Wealth Engine

Category: AI News

  • OpenAI Jalapeño Chip Beats Blackwell 1.9x Per Watt — Ships 2027

    The OpenAI Jalapeño chip, the company’s first custom inference ASIC, delivered 1.5x to 1.9x more AI work per watt than Nvidia’s Blackwell systems in SemiAnalysis InferenceX tests published August 25, 2026. It draws 700W against GB300’s 1,400W and cut end-to-end latency by up to 3.6x. The catch: these are engineering samples. Volume deployment does not arrive until 2027.

    What is the OpenAI Jalapeño chip?

    The OpenAI Jalapeño chip is a custom inference accelerator co-developed with Broadcom and fabricated on TSMC’s N3P node. It is built to serve tokens, not train models. OpenAI published its first third-party benchmarks this week, and they are better than any first-generation silicon has a right to be.

    The headline spec: 13.4 PFLOPS of MXFP4 compute at a 700W rating, paired with HBM4 running at 15.4 TB/s of bandwidth. In sustained operation the part draws under 550W, according to the benchmark data reported by ForkLog.

    Nvidia’s GB200 rack unit pulls 1,200W. GB300 pulls 1,400W. Rubin sits between 900W and 1,150W. Jalapeño is doing its work in roughly half the power envelope.

    The timeline is the real story

    OpenAI started design in mid-2024 and handed the chip to the fab in November 2025. That is nine months from first design to manufacturing handoff, and 16 months to tape-out — a schedule that normally takes a silicon team two to three years.

    OpenAI says its own models helped design the chip. That claim is unverifiable from the outside, but the calendar is not.

    “Jalapeño can serve more AI work per unit of power, while also returning responses more quickly,” said Richard Ho, OpenAI’s head of hardware, in comments reported by TechCrunch.

    How much faster is Jalapeño than Nvidia Blackwell?

    Across three open-weight models, Jalapeño roughly doubled Nvidia’s tokens per second per kilowatt while cutting latency by 43% to 72%. The gap widens as models get larger. On DeepSeek R1 670B, Jalapeño returned a first response in 1.65 seconds against GB300’s 5.99 seconds.

    Here are the SemiAnalysis InferenceX results as reported by ForkLog:

    Model Jalapeño (mixed TPS/kW) Nvidia system Nvidia (mixed TPS/kW) Jalapeño latency Nvidia latency
    GPT-OSS 120B 85,448 GB200 44,960 1.03s 1.80s
    DeepSeek R1 670B 19,641 GB300 11,781 1.65s 5.99s
    Kimi K2.5 1T 18,195 GB300 11,862 1.56s 5.31s

    On single-user throughput, Jalapeño hit roughly 1,400 tokens per second on GPT-OSS 120B and over 700 tokens per second on DeepSeek R1 670B.

    The aggregate claims are wider still: 1.7x to 3.6x lower end-to-end latency and 2.1x to 4.1x higher performance on interactive workloads, per The Decoder. At matched decoding speeds, The Decoder reported token-throughput-per-kilowatt advantages of 54x to 104x — a number that only makes sense in the narrow regime where GPU batching collapses.

    What SemiAnalysis actually said

    “Usually first generation chips aren’t competitive, but OpenAI is beating Nvidia Blackwell and even Rubin,” SemiAnalysis CEO Dylan Patel said, per The Decoder.

    That is a strong endorsement from an analyst house that sells research to the same hyperscalers buying Nvidia racks. Take it seriously. Take it with salt.

    Why does performance per watt decide who wins?

    Because power, not silicon, is the binding constraint on AI buildouts in 2026. Data center operators are queuing for grid interconnects measured in years. If a chip does the same work at half the watts, the same substation serves twice the revenue.

    That math is why custom ASICs keep appearing. Every watt saved on inference is a watt available for a paying customer, and inference is now the majority of frontier-lab compute spend.

    OpenAI CFO Sarah Friar framed it in cost terms: custom chips give the company “greater control over inference costs” and let it match hardware to specific tasks. Friar also said the chip “complements” existing partnerships rather than replacing them — corporate language for we are still buying your GPUs, please keep taking our calls.

    We covered the same power-and-memory squeeze from the supply side in our piece on the Nvidia AI server price hike, and the economics of fast inference in Cerebras vs Groq.

    What does this do to Nvidia’s margins?

    Nothing this quarter. Nvidia reported Q2 fiscal 2027 revenue of $96.22 billion on August 26, beating the $92.07 billion consensus, with data center revenue of $89.02 billion — up 117% year over year, according to 24/7 Wall St. EPS came in at $2.22 against a $2.09 estimate.

    Guidance was louder than the beat. Nvidia guided Q3 to $108 billion plus or minus 2%, with non-GAAP gross margins near 74% and no China data center compute revenue assumed.

    “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue,” CEO Jensen Huang said on the call.

    Nvidia also disclosed supply commitments of $279 billion, largely for Vera Rubin memory. That is a company buying ahead, not one bracing for demand loss.

    The threat is 2028, not 2026

    Custom silicon does not eat Nvidia’s revenue. It eats Nvidia’s pricing power. A 74% gross margin exists because there is no substitute at scale. Jalapeño is the first credible substitute built by Nvidia’s single largest customer.

    NVDA closed at $213.05 before the print, down 3.04% on the week and up 14.37% year to date, per 24/7 Wall St. The stock has fallen after four of its last five earnings reports despite beating consensus three quarters running.

    Who wins and who loses financially?

    Broadcom is the clearest winner. It gets ASIC design revenue, a marquee reference customer, and validation that its custom-silicon business can beat the merchant-GPU incumbent on a first attempt. Nvidia is the clearest medium-term loser, though the damage lands in 2028 pricing, not 2026 volume.

    • Broadcom — books high-margin custom ASIC revenue and proves the model. We covered its financing appetite in the Broadcom AI debt deal.
    • TSMC — wins either way. N3P wafers are N3P wafers, whether the logo says Nvidia or OpenAI.
    • HBM suppliers — Jalapeño uses HBM4 at 15.4 TB/s. More custom chips means more high-bandwidth memory demand, not less.
    • OpenAI — gains leverage in every future GPU negotiation, which may be worth more than the chip itself. Its Nvidia relationship already shifted once, as we noted when Nvidia cut its OpenAI data center guarantee.
    • Nvidia — keeps the volume through 2027, then defends 74% margins against a credible in-house alternative.
    • Second-tier inference clouds — squeezed hardest. They rent GPUs at market rates and cannot design their own.

    What’s the catch with the Jalapeño benchmarks?

    Three catches, and they matter. Jalapeño exists as engineering samples only. Rubin is already shipping to customers. And the benchmark set was chosen by the chip’s owner, run on three open-weight models, with two of Nvidia’s standard optimizations absent from the comparison.

    The Decoder reported that Jalapeño lacks multi-token prediction and speculative decoding optimizations. Those are exactly the techniques that close latency gaps on GPUs. Adding them later helps Jalapeño; adding them to the comparison today would narrow the gap.

    The models tested were GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. Larger current-generation models — DeepSeek V4 Pro, Kimi K3 — were not tested at all. Neither, notably, was any GPT-5-class OpenAI frontier model, which is the workload the chip actually has to serve.

    And the deployment schedule is honest about itself: very small volumes at the end of 2026, meaningful volume in 2027. OpenAI says a second generation is in advanced development and a third is in design.

    A chip that wins benchmarks in August 2026 must still win against whatever Nvidia ships in 2027. That is a different race.

    Frequently asked questions

    Is the OpenAI Jalapeño chip available to buy?

    No. It is an internal accelerator for OpenAI’s own inference fleet, currently at engineering-sample stage. Small-volume deployment starts at the end of 2026, with wider rollout in 2027. There is no external sales channel announced.

    Who manufactures the Jalapeño chip?

    Broadcom co-developed it with OpenAI, and TSMC fabricates it on the N3P process node. The benchmarked silicon is B0 stepping, meaning at least one revision past first tape-out.

    Does Jalapeño beat Nvidia’s Rubin?

    On the perf-per-watt figures SemiAnalysis published, yes — 1.5x to 1.9x. But Rubin is shipping to paying customers now and Jalapeño is not, so the comparison is between a product and a prototype.

    Can Jalapeño train models?

    No. It is an inference-only design. OpenAI still needs GPUs for training, which is why CFO Sarah Friar described the chip as complementing rather than replacing existing supplier relationships.

    How much power does Jalapeño use?

    It is rated at 700W and reportedly sustains under 550W in operation. Nvidia’s GB200 draws 1,200W and GB300 draws 1,400W, so Jalapeño operates in roughly half the envelope.

    Did Nvidia’s earnings show any damage from custom chips?

    None yet. Data center revenue grew 117% year over year to $89.02 billion and Q3 guidance is $108 billion. Custom silicon is a 2028 margin question, not a 2026 revenue question.

    What benchmark was used?

    SemiAnalysis InferenceX, which measures mixed tokens per second per kilowatt alongside end-to-end latency. It is a third-party benchmark, but the model selection and test configuration came from the chip’s owner.

    The bottom line

    Jalapeño is the most serious first-generation AI accelerator anyone has produced, and the power numbers are the part that should worry Nvidia. Half the watts for double the tokens is not a rounding error; it is a structural argument for custom silicon at every lab large enough to fund a design team.

    But the trade here is not “sell Nvidia.” Nvidia just printed $96.22 billion in a quarter and guided to $108 billion. The trade is that Nvidia’s 74% gross margin now has an expiry date attached, and the market will start pricing that date long before 2028 arrives.

    The honest read: OpenAI has proven it can build a chip. It has not yet proven it can build ten million of them, on schedule, while Nvidia iterates annually. Benchmarks are cheap. Yield is not.

    Sources

  • Best Computer Use Agent 2026: Qwen-UI-Agent vs Opus 4.8

    Alibaba’s Qwen-UI-Agent posts the strongest mobile GUI scores anyone has published — 92.2% on MobileWorld-Real — and you cannot buy it. No weights, no API, no price. For a computer use agent you can deploy this week, Claude Opus 4.8 still owns desktop reliability at 83.4% on OSWorld-Verified, and Holo 3.1 wins on cost per task. Everything else is a benchmark press release.

    What is a computer use agent, and why is everyone shipping one now?

    A computer use agent is a model that operates a screen directly — clicking, typing, scrolling — instead of calling an API. It reads screenshots, decides the next action, and repeats. That makes it the only automation layer that works on software with no integration surface.

    The category went from research demo to procurement question fast. Temporal’s 2026 State of Development Report found daily agent usage among surveyed engineers jumped from 47.3% to 80.8% this year. Agents are now normal. Agents that drive a mouse are the next line item.

    The benchmark that actually decides purchases

    OSWorld-Verified is the number buyers use. It runs 369 real desktop tasks in a live Linux environment and scores binary success — the task worked, or it did not.

    On BenchLM’s August 2026 OSWorld-Verified board, Claude Opus 4.8 sits at 83.4%, Gemini 3.6 Flash at 83.0%, and GPT-5.5 at 78.7%. The top of that board is clustered inside about three points. That clustering is the whole story: on desktop, model choice barely moves the needle anymore. Price does.

    Is Qwen-UI-Agent actually better than Claude Opus 4.8?

    On phones, yes, and not by a little. On desktops, no. The Qwen-UI-Agent technical report from Alibaba’s Tongyi MAI team claims 92.2% on MobileWorld-Real, beating Claude Opus 4.8 by 7.5 points and GPT-5.6 Sol by 6.8. On OSWorld-Verified it lands second at 79.5%.

    Where Qwen-UI-Agent wins

    Mobile is a rout. The report puts Qwen-UI-Agent at 82.1% on MobileWorld — 14.6 points ahead of Opus 4.8, 12.0 ahead of GPT-5.6 Sol — and 97.5% on AndroidDaily.

    Grounding is the other win. It reports 81.5% on ScreenSpot-Pro under the zoom-in setting and 97.5% on ScreenSpot-V2. Grounding is the skill of hitting the right pixel, and it is where most agents quietly fail.

    It also takes WebArena at 73.6%, 1.7 points above Opus 4.8.

    Where it loses

    Desktop. Opus 4.8’s 83.4% on OSWorld-Verified beats Qwen-UI-Agent’s 79.5% by 3.9 points, and on the harder OSWorld-v2 the report concedes 13.9% binary success against a 40.0% partial-progress score. Long desktop chains still break.

    Deep research is worse. BrowseComp comes in at 64.1% — roughly 26 points behind GPT-5.6 on the same test, per eesel AI’s breakdown.

    Can you actually buy Qwen-UI-Agent right now?

    No. There is no API, no published price, and no checkpoint on Hugging Face. The Tongyi-MAI repository ships a technical report, demo videos, and cookbook notebooks. The Apache 2.0 license covers the repo contents, not a model you can run.

    What is downloadable is the previous generation: MAI-UI 8B (8.77B parameters) at 65.7% on ScreenSpot-Pro and MAI-UI 2B (2.13B) at 57.4%. The larger 235B-A22B variant tops AndroidWorld at 76.7% and is not published either.

    Treat the headline scores as a roadmap, not a product. That gap between announcement and availability is the single most expensive mistake in agent procurement this year.

    How much does a computer use agent cost in 2026?

    Between $0.03 and $30 per million input tokens, depending on tier. Screen-driving is token-hungry: every step ships a fresh screenshot, so a computer use agent burns input tokens at several times the rate of a chat workload. Input price, not output price, is what sets your bill.

    Agent / model Access Input $/M Output $/M Best published score
    Claude Opus 4.8 Hosted API (GA) $5.00 $25.00 83.4% OSWorld-Verified
    Claude Sonnet 5 Hosted API (GA) $2.00 $10.00 Mid-tier desktop
    GPT-5.6 Sol Hosted API (GA) $5.00 $30.00 92.8% Online-Mind2Web
    GPT-5.6 Luna Hosted API (GA) $0.20 $1.20 Budget browser tier
    Gemini 3.7 Flash Hosted API (preview) $0.75 $3.75 83.0% (3.6 Flash)
    Holo 3.1 Open weights + API $0.25 $1.80 82.56% OSWorld-Verified
    Qwen3.5 Flash Hosted API $0.10 $0.40 Grounding / cheap steps
    Qwen-UI-Agent Not released — — 92.2% MobileWorld-Real

    Claude prices come from Anthropic’s pricing documentation; OpenAI, Gemini and Qwen rates from BenchLM’s August 2026 tables.

    The token math nobody prices in

    Four multipliers decide whether your pilot survives contact with finance.

    • Tokenizer drift. Anthropic documents that Claude 4.7 and later use a newer tokenizer producing roughly 30% more tokens for the same text. Same price, bigger invoice.
    • Cache hits. Cached input bills at 0.1x on Claude and 10% of standard on OpenAI. For screenshot loops with a stable system prompt, that is the difference between viable and absurd.
    • Batch discounts. Both Anthropic and OpenAI cut 50% for asynchronous batch work. Most agent tasks are not urgent.
    • Fast mode. Anthropic’s fast tier for Opus 5 and 4.8 runs $10 input and $50 output — double the standard rate for latency you probably do not need on a background agent.

    One more piece of good news for budgets: Anthropic made Sonnet 5’s introductory $2/$10 pricing permanent, canceling the increase to $3/$15 that was scheduled for September 1. If you built a cost model around that hike, tear it up. We covered the token-efficiency side of this in Agent Skills vs MCP.

    Which computer use agent should you pick for your use case?

    Match the agent to the surface, not to the leaderboard. Desktop reliability, mobile coverage, browser scale and self-hosting are four different purchases, and no single vendor currently wins all four. Here is the split as of late August 2026.

    Use case Pick Why Cost signal
    Desktop work where errors are expensive Claude Opus 4.8 83.4% OSWorld-Verified; prompt-injection classifiers run over every screenshot $5/$25 per M tokens
    Phone plus desktop from one endpoint Gemini Computer Use Only hosted API with a documented Android environment $0.75/$3.75 (3.7 Flash)
    Browser-only automation at volume Browser Use 87.4% on Odysseys; model-agnostic, cloud supplies Chromium From $29/mo, $0.02/browser hour
    Data cannot leave your network Holo 3.1 82.56% OSWorld-Verified on a maintainer-run board; Apache 2.0 $0.25/$1.80 or self-host free
    Logged-in flows blocked by 2FA Skyvern Credentials never sent to the model; 85.85% WebVoyager From $29/mo; 2FA at $149/mo
    Commercial product, permissive license UI-TARS Apache 2.0, ~808k monthly downloads Free weights

    Are open-weight GUI agents good enough yet?

    One is. Holo 3.1 scores 82.56% on OSWorld-Verified on a maintainer-run board — within a point of Opus 4.8 — under Apache 2.0, with a hosted fallback at $0.25/$1.80 per million tokens and a free tier at 10 requests per minute. That is the only open model verified by a neutral board rather than its own vendor.

    The rest are not close. ByteDance’s UI-TARS is the most downloaded GUI agent at roughly 808,000 monthly pulls, but the downloadable checkpoint scores 27.5% on OSWorld — a 15-point gap from the published figure.

    The pattern mirrors what we found in GLM-5.3 vs DeepSeek V4 Pro: open weights close the gap on paper faster than they close it in production.

    Is a computer use agent worth it in 2026?

    Only when no API exists. The most useful finding in this cycle is that the best GUI agents avoid the GUI: 55.1% of actions on OSWorld-v2 were typed commands rather than clicks. Screen-driving is the fallback, not the strategy.

    The failure data backs that up. Roughly 52% of real-device failures trace to the screen itself and 40% to the agent, meaning half your errors come from interfaces that were never designed for a machine to read.

    The financial read is blunt. If a documented API exists, an API integration costs less, breaks less, and does not bill you for screenshots. Reserve computer use for legacy desktop software, vendor portals with no API, and mobile apps. That is a real and expensive category — it is just smaller than the demos suggest. The same discipline applies to coding agents, which we priced in Claude Code vs Codex CLI.

    Frequently asked questions

    What is the best computer use agent right now?

    Claude Opus 4.8 for desktop reliability at 83.4% OSWorld-Verified, Gemini Computer Use if you need Android, and Holo 3.1 if you need open weights. Qwen-UI-Agent has the best benchmarks and no product.

    When will Qwen-UI-Agent weights be released?

    Alibaba has not announced a date. As of late August 2026 the repository has no releases and no checkpoint on Hugging Face. Only the older MAI-UI 2B and 8B models are downloadable.

    How much does Claude computer use cost?

    Claude Opus 4.8 bills $5 per million input tokens and $25 per million output, per Anthropic’s pricing docs. Cache hits drop input to 0.1x and batch processing cuts both sides by 50%.

    Is Claude Sonnet 5 still going up in price on September 1?

    No. Anthropic canceled the increase and made the $2/$10 introductory pricing permanent. The previously scheduled $3/$15 rate will not take effect.

    Do computer use agents work on mobile?

    Partially. Gemini Computer Use is the only hosted API with a documented Android environment. Claude’s computer use has no mobile support. Qwen-UI-Agent leads mobile benchmarks but ships nothing.

    Are computer use agents safe to run on production systems?

    Only with guardrails. Anthropic runs prompt-injection classifiers over every screenshot automatically; Gemini’s injection detection is off by default, and OpenAI leaves safety implementation to you. Scope credentials tightly and log every action.

    What is OSWorld-Verified?

    A 369-task benchmark that scores computer use agents on real desktop work in a live environment, pass or fail. It is the closest thing the category has to a purchasing standard.

    The bottom line

    Buy Claude Opus 4.8 for desktop work where a failed task costs more than a token bill — the 83.4% OSWorld-Verified score and automatic injection screening justify $5/$25. Buy Holo 3.1 if data residency or cost dominates: 82.56% for $0.25/$1.80 is the best price-per-point in the category by a wide margin.

    Ignore Qwen-UI-Agent until weights land. A 92.2% score you cannot run is worth exactly zero. And before you buy any of them, check whether the target system has an API — because 55.1% of the winning agent’s own actions were typed commands, which is the benchmark quietly telling you to skip the screen entirely.

    Sources

  • Nvidia AI Server Price Hike: 15% More, and Memory Is Why

    Nvidia is raising AI server prices by more than 15% on Grace Blackwell and Vera Rubin systems shipping in early 2027, Bloomberg reported on August 24, 2026. Memory is the reason. UBS puts memory at 62% of a Vera Rubin superchip’s $38,902 bill of materials, up from 53% on Grace Blackwell. TrendForce estimates the hike adds at least $5 billion to a 1-gigawatt data center.

    For three years the AI trade had one simple rule: Nvidia sets the price, and everyone pays it. That rule still holds. What changed is who Nvidia is paying.

    The Nvidia AI server price hike is not a margin grab. It is a pass-through. And the numbers underneath it say the memory makers, not the GPU designer, now control the cost curve of the AI build-out.

    How much is Nvidia raising AI server prices?

    More than 15% in many cases, effective on systems shipped early next year. Bloomberg reported the increases on August 24, citing people familiar with the matter. TrendForce, summarizing the same reporting, said some configurations could reach 17%. Nvidia did not respond to requests for comment.

    The warnings did not go to the cloud giants directly. According to Bloomberg, Nvidia notified the contract server manufacturers that assemble systems for Microsoft, Alphabet’s Google, and Oracle.

    That routing matters. The ODMs absorb the notice first, then reprice their own quotes. The cloud buyers find out when the invoice changes.

    Which systems are affected

    • Grace Blackwell systems — the current generation, still shipping in volume.
    • Vera Rubin systems — the next generation, with first shipments in early 2027.
    • Increases vary by chip generation and by memory configuration, per Bloomberg. Denser memory builds take the larger hit.

    Why is Nvidia raising prices now?

    Because memory has gone from a line item to the line item. Morgan Stanley estimates GPU silicon has fallen from more than 80% of AI server cost to roughly half that level in next-generation systems. The gap did not close because GPUs got cheaper. It closed because DRAM got expensive.

    A Vera Rubin NVL72 rack carries 74.7 TB of DRAM — 20.7 TB of HBM4 plus 54 TB of LPDDR5X, according to UBS’s teardown. That is the DRAM content of roughly 4,500 smartphones in a single rack.

    Every one of those bits is bought in the tightest memory market in a decade.

    What UBS found inside a Vera Rubin superchip

    UBS’s bill-of-materials analysis is the clearest picture available of where the money actually goes.

    Component Cost Share of superchip
    Total Vera Rubin superchip $38,902 100%
    All memory $24,297 62%
    SOCAMM2 LPDDR5X $19,355 49.8%
    HBM4 $4,943 12.7%
    Everything else $14,605 38%

    On Grace Blackwell, UBS put memory at 53% of cost. On Vera Rubin it is 62%, and the absolute memory bill rose about 2.5x between generations.

    One caveat worth holding onto: that 2.5x blends two different things. Vera Rubin carries more memory and pays more per gigabyte. It is not a pure price signal.

    How much has DRAM actually gone up?

    Steeply, and for longer than most forecasts allowed. TrendForce data cited by Tom’s Hardware shows conventional DRAM contract prices rising 90–95% quarter-over-quarter in Q1 2026 and a projected 58–63% in Q2 2026. Server DRAM is expected to climb every quarter through the second half of 2027.

    The consumer market tells the same story in plainer numbers. A mainstream 32GB DDR5-6000 kit runs about $392 today against $110–$140 a year ago, per Tom’s Hardware.

    Supply was committed early. SK hynix had sold out its entire 2026 production capacity by October 2025. Samsung and SK hynix raised 2026 HBM3E prices by roughly 20%.

    And HBM makes the squeeze worse mechanically: it consumes roughly four times the wafer area of conventional DRAM per bit shipped. Every HBM4 order crowds out ordinary server memory on the same fab.

    Who profits from the Nvidia AI server price hike?

    Not Nvidia, on the arithmetic. The memory suppliers capture the increase, the ODMs pass it through, and the hyperscalers eat it. Nvidia’s role here is closer to toll collector than beneficiary — and its own gross margin may be the quiet casualty.

    Work the math. Nvidia runs roughly a 75% gross margin, so the bill of materials is about 25% of the sale price. If memory is 62% of that BOM, memory is about 15.5% of the price. A 2.5x memory cost increase adds roughly 23 points of price to cost.

    A 15% price hike does not cover 23 points. Something has to give.

    Three readings, and the market has not settled on one:

    1. The 15% is an opening installment. More increases follow as 2027 contracts reprice.
    2. Nvidia is absorbing the difference. Gross margin drifts from ~75% toward the high 60s.
    3. The 2.5x is generational, not inflationary. Higher memory content is sold at a higher system ASP, so the comparison overstates the pass-through problem.

    Reading three is the most likely and the least discussed. It is also the one that would let Nvidia keep its margin story intact — which is precisely why it deserves scrutiny rather than acceptance.

    Why this matters for AI capex

    Because it reprices the entire build-out. TrendForce estimates the increase adds at least $5 billion to the cost of a 1-gigawatt AI data center. At the scale hyperscalers are now committing to, that is not a rounding error — it is a line in the capital plan that did not exist last quarter.

    The second-order effects are where this gets interesting.

    The uncomfortable version: AI compute has been getting cheaper per unit of intelligence for three straight years. This is the first credible input cost that pushes the other way.

    This post is reporting and analysis, not financial advice.

    Frequently asked questions

    How much is Nvidia raising AI server prices?

    More than 15% in many cases, with some configurations reaching 17% per TrendForce. Increases vary by chip generation and memory configuration.

    When do the new prices take effect?

    On systems shipped in early 2027, according to Bloomberg’s August 24, 2026 report.

    Which Nvidia systems are affected?

    Grace Blackwell and Vera Rubin server systems. Both are rack-scale platforms sold to cloud and enterprise data center operators.

    Why are AI server prices going up?

    Memory costs. UBS puts memory at 62% of a Vera Rubin superchip’s cost, and DRAM contract prices have risen every quarter through 2026 amid an HBM-driven supply squeeze.

    Who was notified about the price increases?

    Contract server manufacturers that build systems for Microsoft, Google, and Oracle, per Bloomberg. Nvidia did not comment publicly.

    How much does this add to a data center?

    TrendForce estimates at least $5 billion in additional cost for a 1-gigawatt AI data center.

    Does this hurt Nvidia’s margins?

    Possibly. Nvidia runs roughly a 75% gross margin. If memory costs rose 2.5x generationally, a 15% price increase may not fully offset it — though part of that increase reflects more memory content per system, not pure inflation.

    The bottom line

    The Nvidia AI server price hike is the clearest sign yet that the AI supply chain’s power center is shifting. For three years the scarce input was GPU wafer allocation. In 2027 it is memory, and the companies that own it — SK hynix, Samsung, Micron — are the ones setting terms.

    Watch two things next. First, whether Nvidia’s gross margin guidance holds through the fiscal year, because that is where the pass-through gap shows up. Second, whether any hyperscaler publicly revises a gigawatt commitment. The first cost-driven downgrade of an announced buildout would tell you the memory squeeze has stopped being an engineering problem and started being a financial one.

    Sources

  • Agent Skills vs MCP: Which One Cuts Your Token Bill

    Agent Skills vs MCP is not really a fight — but on cost, Skills win outright. Anthropic’s own numbers show 58 MCP tools burning roughly 55,000 tokens before an agent does anything useful. A Skill’s metadata costs about 100 tokens. Build Skills for procedure and judgment, MCP servers for live system access, and cache aggressively if you run both.

    On August 19, 2026, Anthropic moved Agent Skills, the Skills API, computer use, browser use and the Files API out of beta and into general availability on the Claude Developer Platform, according to the platform’s release notes. Three weeks earlier, the Model Context Protocol shipped its stateless 2026-07-28 specification.

    Both are now production infrastructure. Both are free to adopt. Only one of them charges you rent on every single request.

    What actually changed on August 19, 2026?

    Agent Skills stopped being an experiment. The skills-2025-10-02 beta header is gone, the /v1/skills endpoint is GA, and computer use shipped as computer_toolset_20260801 with browser use as a separate tool, browser_toolset_20260801. That is a full production agent stack in one release.

    The Files API went GA the same day. Upload, download, list, metadata and delete operations are free — you are billed only for file content that actually enters a Messages request, at standard input-token rates.

    Storage caps are generous: 500 MB per file and 1 TB per organization, rate-limited to roughly 500 requests per minute.

    None of the superseded beta headers have a published sunset date yet. Legacy identifiers still work. But GA is the signal that pricing and architecture decisions made today will stick.

    What is the difference between Agent Skills and MCP?

    MCP gives an agent reach. Agent Skills give an agent method. An MCP server tells Claude how to connect to GitHub and what it can do there. A Skill tells Claude how your team actually writes a release note, in what order, with which checks.

    That distinction sounds academic until you look at where each one lives in the context window.

    How MCP loads tools

    MCP tool definitions are pushed into the model’s context up front, on every request. The agent has to know a tool exists before it can call it, so the full schema — names, parameters, descriptions — sits in the prompt whether or not the tool ever gets used.

    That is a fixed tax. It scales linearly with how many servers you connect.

    How Agent Skills load

    Skills use three-tier progressive disclosure, documented in Anthropic’s Agent Skills overview. Level 1 is YAML frontmatter — name and description only — at roughly 100 tokens per Skill, always loaded. Level 2 is the SKILL.md body, under about 5,000 tokens, loaded only when the Skill is triggered.

    Level 3 is where it gets interesting. Bundled reference files and scripts cost zero tokens until read. Script code never enters the context window at all — only its output does.

    You can ship 300 pages of API documentation inside a Skill and pay nothing for it unless the agent opens the file.

    How much do MCP tool definitions actually cost?

    More than most teams realize. Anthropic published a concrete five-server example in its advanced tool use write-up: GitHub at 35 tools (~26K tokens), Slack at 11 tools (~21K), Sentry at 5 (~3K), Grafana at 5 (~3K), and Splunk at 2 (~2K).

    That is 58 tools consuming approximately 55,000 tokens before the conversation even starts.

    Anthropic’s own internal setup is worse: tool definitions there consume 134,000 tokens before optimization. On a 200K context window, that is two-thirds of the budget spent on a menu the agent mostly ignores.

    Now price it. Claude Sonnet 5 costs $2 per million input tokens and Claude Opus 5 costs $5, per the official pricing page. Sonnet 5’s introductory rate was made permanent on August 10, 2026, and the scheduled September increase was cancelled.

    What does that cost per month at real volume?

    Setup Tokens per request Sonnet 5 cost / request 10,000 runs / month
    Anthropic’s internal tool set 134,000 $0.268 $2,680
    58 MCP tools (5 servers) ~55,000 $0.110 $1,100
    Same tools + Tool Search Tool ~8,700 $0.017 $174
    20 Skills + 1 triggered SKILL.md ~7,000 $0.014 $140
    20 Agent Skills (metadata only) ~2,000 $0.004 $40
    Token figures from Anthropic; cost math at the published $2/MTok Sonnet 5 input rate, uncached.

    The spread between the top and bottom row is 27x. On Opus 5 at $5 per million input tokens, the same gap costs $6,700 versus $100 a month.

    The prompt caching escape hatch

    MCP defenders have a real counterargument: cache the tool definitions. Anthropic’s pricing page lists cache reads at 0.1x the base input rate — $0.20 per million tokens on Sonnet 5.

    Cached, that 55,000-token block drops from $0.110 to about $0.011 per request. A 1-hour cache write costs 2x base and pays for itself after two reads.

    So caching closes most of the gap — if your traffic is dense enough to keep the cache warm and your tool list is stable. Bursty, low-volume agents get cache misses and pay full freight.

    Does trimming tools hurt accuracy?

    No — it helps. This is the part that surprises people. Anthropic’s Tool Search Tool delivers an 85% reduction in token usage while keeping the full tool library reachable, cutting roughly 77K tokens of overhead down to about 8.7K and preserving 95% of the context window.

    Accuracy went up. Opus 4 improved from 49% to 74% on the tool-use benchmark. Opus 4.5 went from 79.5% to 88.1%.

    Programmatic Tool Calling shows the same pattern: average usage dropped from 43,588 to 27,297 tokens, a 37% reduction on complex research tasks, while internal knowledge retrieval rose from 25.6% to 28.5% and GIA scores went from 46.5% to 51.2%.

    Fewer tools in context means less for the model to confuse. Context bloat is an accuracy problem wearing a cost problem’s clothes.

    When should you still build an MCP server?

    When you need a live connection, real authentication, or one integration shared across many agents. Skills are static files — they cannot hold an OAuth token, stream an update, or talk to your database. MCP is the transport layer, and nothing about Skills replaces it.

    MCP’s adoption numbers back that up. Claude’s connector directory now lists over 950 MCP servers, and MCP has passed 400 million monthly SDK downloads — a 4x increase this year, per Anthropic’s spec announcement. The TypeScript and Python SDKs have crossed a billion total downloads between them.

    Use this five-question test before you write a line of either:

    1. Does it need a live connection to a running system? MCP server.
    2. Does it need per-user auth or scoped permissions? MCP server.
    3. Is the hard part procedure and judgment, not access? Agent Skill.
    4. Is it large reference material used occasionally? Agent Skill — Level 3 files cost nothing until opened.
    5. Do many agents share one integration? MCP server centralizes it; a Skill travels with the agent.

    Agent Skills vs MCP: which should you use for your task?

    What you’re building Pick Reason
    Read/write live Slack, GitHub or Jira data MCP server Needs a live, authenticated connection
    House style guide, review checklist, report format Agent Skill Pure procedure; ~100 tokens idle
    Ship 300 pages of API reference to the agent Agent Skill Level 3 files cost 0 tokens until read
    Per-user OAuth scopes and audit trails MCP server Auth belongs at the connection layer
    Deterministic script the agent runs but shouldn’t read Agent Skill Script code never enters context
    One integration consumed by a dozen agents MCP server Central updates, single surface
    Pull live data and apply a fixed workflow Both MCP for access, Skill for method
    Low-volume, bursty agent on a tight budget Agent Skill Cold caches make MCP overhead expensive
    The production default is both — MCP for reach, Skills for method.

    What did the 2026-07-28 MCP spec change?

    It made MCP stateless, which is the single biggest cost change on the server side. The initialize/initialized handshake and the Mcp-Session-Id header are retired. Every request is now self-describing, so any request can land on any instance behind a plain round-robin load balancer.

    That kills the shared-storage requirement. You can run MCP serverless and stop paying for sticky sessions.

    Method and tool names now travel in HTTP headers rather than JSON bodies, letting gateways route without parsing payloads. Multi Round-Trip Requests replace server-initiated streams for interactive confirmations.

    Roots, Sampling and Logging are deprecated but will keep working for at least twelve months, as will the legacy HTTP+SSE transport. You have a year to migrate — not a weekend.

    One number worth sitting with: at Honeycomb, nearly 20% of all monthly interactive queries are now made by agents, not humans. That ratio is why the per-request tax matters.

    Frequently asked questions

    Are Agent Skills a replacement for MCP?

    No. Skills carry procedural knowledge as files; MCP carries live, authenticated access. Anthropic shipped both to GA in 2026 and the standard production pattern uses them together.

    How many tokens does one Agent Skill cost?

    Roughly 100 tokens for its metadata at startup, per Anthropic’s documentation. The SKILL.md body — under about 5,000 tokens — loads only when the Skill is triggered.

    Do MCP tool definitions get charged on every request?

    Yes, unless cached. Anthropic’s example of 58 tools across five servers consumes roughly 55,000 input tokens per request. Prompt caching cuts that to 0.1x the base rate on cache hits.

    Does the Files API cost extra?

    No. Upload, download, list, metadata and delete are free. You pay only for file content that enters a Messages request, billed as normal input tokens.

    Is the old MCP spec still supported?

    Yes. Roots, Sampling, Logging and the HTTP+SSE transport are deprecated but supported for at least twelve months from the 2026-07-28 release.

    Which model should I run agents on to keep costs down?

    Claude Sonnet 5 at $2/$10 per million tokens is the volume workhorse; its introductory pricing was made permanent on August 10, 2026. Opus 5 at $5/$25 is 2.5x the input cost — worth it only when the task genuinely needs it.

    Do fewer tools in context make agents dumber?

    The opposite. With Tool Search Tool enabled, Opus 4.5 improved from 79.5% to 88.1% on Anthropic’s tool-use benchmark while using 85% fewer tokens.

    The bottom line

    Build Skills first. They are nearly free to keep loaded, they went GA on August 19, and Anthropic’s own benchmarks show that trimming context raises accuracy rather than lowering it.

    Add MCP servers only where you need a live connection, real auth, or one integration shared across many agents — then wrap them in Tool Search Tool and prompt caching on day one, not after the first surprising invoice.

    The decision rule is exactly this: if the hard part is reaching the system, build MCP; if the hard part is knowing what to do once you’re there, build a Skill. Teams running high-volume agents on the naive pattern are paying somewhere between 8x and 27x more per request than they need to, and getting worse answers for the money.

    Related reading on agent economics: our breakdown of Claude Code vs Codex CLI on cost, the Gemini 3.7 Flash vs Claude Sonnet 5 cost-per-coding-point comparison, and ChatGPT Business vs Claude Team on the seat-license side.

    Sources

  • Google Marvell Chip Deal: $12.2B Warrant, $120B Catch

    Google secured a warrant for 58,970,907 Marvell shares at $206.58 each — about $12.2 billion — under a custom silicon agreement disclosed on August 19, 2026. Marvell’s 8-K shows only 1,360,867 shares vest on time. The rest unlock in 240 tranches, one per $500 million of custom product revenue: $120 billion of chip purchases through fiscal 2033.

    The Google Marvell chip deal is the clearest sign yet that hyperscalers no longer just buy silicon. They take equity in the companies that build it.

    Marvell Technology stock jumped 13% on the disclosure. Broadcom fell 3%. Alphabet did not move at all.

    What is the Google Marvell chip deal?

    It is a custom silicon supply agreement signed July 29, 2026, paired with a stock warrant issued August 18, 2026. Marvell will design chips across five categories for Google’s TPU infrastructure. In exchange, Google holds an option on roughly 7% of Marvell, priced today and payable later.

    According to Marvell’s 8-K filing with the SEC, the warrant expires August 18, 2033.

    The five product lines Marvell will supply, per analysis from The Futurum Group:

    • Inference accelerators
    • Storage controllers
    • Network interface controllers
    • Memory interface controllers
    • Near-memory compute

    That is not one chip. That is a seat at every layer of the rack.

    How much is the warrant actually worth?

    At the $206.58 strike price, full exercise costs Google about $12.18 billion and delivers 58,970,907 shares. But the headline number is a ceiling, not a payment. Google owes nothing today. Almost the entire position is contingent on purchase volume Marvell has never come close to booking from a single customer.

    Term Detail
    Warrant shares 58,970,907
    Exercise price $206.58 per share
    Value at full exercise ~$12.18 billion
    Time-based tranche 1,360,867 shares, equal quarterly installments in year one
    Performance tranches 240 tranches, one per $500M of custom product revenue
    Implied purchase total $120 billion
    Vesting window Q3 fiscal 2027 through end of fiscal 2033
    Expiration August 18, 2033
    Commercial agreement signed July 29, 2026

    The vesting math nobody put in the headline

    Divide 240 tranches by the roughly six and a half years between Q3 fiscal 2027 and the end of fiscal 2033. Futurum calculates Google would need to average close to $18 billion a year in custom purchases from Marvell to unlock the full warrant.

    Hold that number. It matters in a moment.

    Why would Google take equity in its own supplier?

    Because it converts a procurement line into an asset. If Google spends $120 billion with Marvell and Marvell’s stock rises on that revenue, Google captures part of the gain it created. If Google spends nothing, the warrant lapses and costs it nothing.

    The structure is asymmetric by design. Google pays with optionality, not cash.

    It also locks Marvell in. A supplier whose largest shareholder-in-waiting is its largest customer has limited leverage on price. That is the quiet half of the deal.

    Variations of this circular financing keep appearing across the sector — most visibly when Nvidia cut its OpenAI data center guarantee from $250B to $120B, and again in Broadcom’s up-to-$100 billion debt raise to fund Anthropic chips.

    Does Marvell replace Broadcom as Google’s TPU partner?

    No. Broadcom remains Google’s primary TPU design partner under a long-term agreement running through 2031. Morningstar analyst William Kerwin, quoted by TheStreet, called the deal “a strong win for Marvell” while noting Google was “adding new suppliers rather than dropping Broadcom.”

    The read is capacity, not replacement. Broadcom’s design teams are booked on core accelerator generations. Marvell picks up memory expansion, decode-focused inference, and interconnect controllers.

    Broadcom’s 3% drop on the news looks like a market pricing in a smaller share of a much larger pie.

    Can Marvell realistically deliver $120 billion?

    This is where the number starts to strain. Marvell’s Q1 fiscal 2027 results show total net revenue of $2.418 billion for the quarter ended May 2, 2026, with data center at $1.833 billion — 76% of the business and up 28% year over year.

    Guidance for Q2 is $2.700 billion, plus or minus 5%. Annualize that and Marvell is a roughly $10.8 billion revenue company.

    Now compare. To fully vest the warrant, Google alone would need to buy about $18 billion of custom silicon a year — roughly 1.7 times everything Marvell currently sells to every customer combined.

    Management’s own stated target is more than $10 billion in custom revenue by fiscal 2029, across all customers. The Google ceiling sits an order of magnitude above the plan.

    Treat $120 billion as a theoretical maximum with a marketing function, not a forecast. The tranche structure exists precisely because neither side expects the top of the range.

    How did the market react?

    Sharply, and selectively. On August 19, 2026, Marvell rose 13% to $243.66 while Broadcom fell 3% to $369.13 and Alphabet closed unchanged at $342.67, according to 24/7 Wall St.

    Alphabet’s flat tape is the most interesting line in that table. A $120 billion purchase commitment moved the buyer’s stock zero percent.

    That tells you the market already assumed Google would spend the money somewhere. Only the recipient was in question.

    Dilution is real but modest: full exercise cuts existing shareholders by roughly 6.3% to 6.7% and would make Google approximately Marvell’s fifth-largest investor, per TheStreet.

    Why this matters

    Custom silicon is where the AI infrastructure margin is migrating. Every hyperscaler that designs its own accelerator takes revenue that would otherwise flow to Nvidia — and hands part of it to a merchant design partner like Broadcom or Marvell.

    The warrant structure is the new template. Compute buyers are increasingly paid in equity for their own demand. That is what Nvidia’s $6 billion Poolside arrangement did in software, and it is the same logic investors are pricing into custom-inference startups like Etched at a $21 billion valuation.

    For investors, the practical question is not whether the $120 billion lands. It is whether Marvell’s custom design wins convert into recognized revenue on the quarterly cadence the tranches imply. Watch the custom line, not the headline.

    This post is reporting and analysis, not financial advice.

    Frequently asked questions

    How many Marvell shares does the Google warrant cover?

    58,970,907 shares at an exercise price of $206.58, worth about $12.18 billion at full exercise, per Marvell’s 8-K.

    When does the Google Marvell warrant expire?

    August 18, 2033. Vesting runs from the third quarter of fiscal 2027 through the end of fiscal 2033.

    What has to happen for the full warrant to vest?

    Beyond 1,360,867 time-based shares, tranches vest one at a time for every $500 million of custom product revenue — 240 tranches, or $120 billion total.

    Is Google dropping Broadcom for Marvell?

    No. Broadcom holds a TPU design agreement through 2031 and remains the primary partner. Marvell is being added across adjacent chip categories.

    How much dilution do Marvell shareholders face?

    Roughly 6.3% to 6.7% if the warrant is fully exercised, which would put Google around fifth among Marvell’s largest holders.

    What is Marvell’s current revenue?

    $2.418 billion in the quarter ended May 2, 2026, with Q2 guidance of $2.700 billion plus or minus 5%.

    Did Alphabet stock move on the news?

    No. Alphabet closed unchanged at $342.67 on August 19, 2026, while Marvell rose 13% and Broadcom fell 3%.

    The bottom line

    The Google Marvell chip deal is a genuine design win wrapped in a number that will not be met. Marvell gets multi-year attachment across five product categories inside the largest custom accelerator program outside Nvidia. Google gets a free option on the value it creates by spending.

    The next real datapoint is Marvell’s custom product revenue line. Each $500 million tranche is a public scoreboard — a rare case of customer concentration disclosed quarter by quarter through a vesting schedule.

    If two or three tranches clear in fiscal 2028, the thesis holds. If the line stays flat while the stock trades on $120 billion, the gap closes the hard way.

    Sources

  • ChatGPT Business vs Claude Team: Which $125 AI Seat Wins

    ChatGPT Business vs Claude Team has collapsed into a price tie. Both labs now charge $125 per premium seat per month, $100 on annual billing, and $25 for a standard seat. The sticker no longer decides it. Claude Team wins for engineers who live in a terminal. ChatGPT Business wins for mixed teams that need research, documents, and code in one subscription.

    What changed in AI seat pricing in August 2026?

    OpenAI introduced $125 premium seats for ChatGPT Business on August 11, 2026, aimed at users hitting rate limits. Anthropic already sold a $125 premium Team seat. Meta is preparing Hatch, its first paid AI product, with a tier reported as high as $199.99 per month. Flat-rate AI is over.

    Why flat-rate AI plans broke

    Agents do not chat. They loop. A single agentic coding task can consume more tokens than a week of manual prompting.

    OpenAI said the change was driven by teams “tackling more complex tasks,” according to The Decoder’s August 11 report. The subtext is simpler: $25 seats were losing money on power users.

    The five-hour cap that started it

    Standard ChatGPT Business seats are throttled by a rolling five-hour usage limit. Premium seats remove it and add five times the capacity, with usage resetting weekly for both tiers.

    Anthropic runs the same playbook. Claude’s own pricing page describes Pro usage in terms of “per 5-hour session,” and Max tiers sell 5x or 20x that allowance for $100 and up.

    How much does each seat actually cost?

    Standard seats are $25 monthly or $20 annually at both vendors, with a two-seat minimum. Premium seats are $125 monthly or $100 annually at both. Anthropic caps self-serve Team plans at 150 people. Microsoft charges $30 per user on top of an existing Microsoft 365 license.

    The full price comparison

    Plan Monthly Annual (per mo.) Seat rules Usage
    ChatGPT Business standard $25 $20 2 seat minimum 5-hour rolling cap
    ChatGPT Business premium $125 $100 Mix with standard 5x, no 5-hour cap
    Claude Team standard $25 $20 2–150 people Above Pro limits
    Claude Team premium $125 $100 2–150 people 5x standard seat
    Claude Enterprise $20 + usage Annual only Contact sales Metered at API rates
    Microsoft 365 Copilot $30 add-on Annual commit Needs base license Tenant-level

    Sources: claude.com/pricing, The Decoder, and ExplainX’s 2026 Copilot licensing breakdown.

    What the premium markup really buys

    Notice the ratio. OpenAI charges 5x the price for 5x the usage. That is perfectly linear — there is no volume discount for committing to a heavier seat.

    You are not buying cheaper tokens. You are buying permission to keep working past lunchtime.

    Which is better for coding, ChatGPT Business or Claude Team?

    Claude Team, narrowly. Both bundle a terminal coding agent into paid seats, but Anthropic includes Claude Code across Pro, Max, Team and Enterprise with no separate line item. OpenAI bills Codex, Deep Research and agent workflows against per-seat caps, which makes heavy coding weeks unpredictable.

    Terminal access is the dividing line

    Claude’s pricing page lists Claude Code as included on every paid tier, including the $20 Pro plan. That is unusual generosity for a product this expensive to serve.

    OpenAI’s structure is different: Codex runs against the same seat allowance as chat and research. Burn your budget on a refactor and your product manager’s Deep Research queries compete for the same pool.

    We ran the head-to-head on the tools themselves in Claude Code vs Codex CLI. The seat math reinforces that verdict rather than changing it.

    Where OpenAI takes the win back

    Breadth. A single ChatGPT Business seat covers Deep Research, Sora, agent mode and Codex. Anthropic has no video model and no consumer media stack.

    If your team is five engineers, that does not matter. If it is fifty people across marketing, legal and engineering, it matters a lot.

    Is a $125 premium seat cheaper than paying API rates?

    Only if the seat holder is genuinely heavy. At Anthropic’s published API rates, $125 buys a specific and knowable amount of compute. Below that threshold you are subsidizing the vendor. Above it, the flat seat is one of the best deals in enterprise software.

    The break-even numbers

    Using Anthropic’s official API price list — Sonnet 5 at $2 per million input tokens and $10 per million output, Opus 5 at $5 and $25 — a $125 monthly seat is worth:

    • 12.5 million Sonnet 5 output tokens at standard rates
    • 5 million Opus 5 output tokens, the frontier-tier equivalent
    • 25 million Sonnet 5 output tokens if you route through the Batch API, which is discounted 50%
    • 62.5 million Sonnet 5 input tokens on uncached prompts
    • 625 million cached input tokens, since cache reads bill at 0.1x the base input rate

    That last line is the one finance teams miss. Prompt caching turns a $125 seat into an enormous amount of context.

    The verdict on break-even

    A developer running agentic coding loops daily will clear 12.5 million output tokens without trying. A weekly Deep Research user will not come close.

    Mix your seats. Both vendors let standard and premium sit in the same workspace, and buying every seat at $125 is the most common way teams overspend.

    How do Microsoft Copilot and Meta Hatch compare?

    Badly, on price. Microsoft 365 Copilot costs $30 per user monthly but requires a qualifying base license, pushing the true all-in cost to roughly $66–$69 on E3 and $87–$90 on E5. Meta’s Hatch has not launched publicly and its pricing remains reported, not confirmed.

    Copilot’s stacked-license problem

    The $30 headline is not the bill. ExplainX puts Microsoft 365 E3 at roughly $36–$39 per user and E5 at $57–$60 before Copilot is added.

    Small businesses get a break: Microsoft 365 Copilot Business runs $18 per user promotionally through December 31, 2026, rising to a $21 standard rate, capped at 300 seats.

    There is an irony worth pricing in. Microsoft’s Copilot Cowork tier reportedly runs on Claude Opus 4.8 and Sonnet 5. You can pay Microsoft roughly $60 a month for a wrapper around models Anthropic will sell you directly at $25.

    Meta Hatch and the $199.99 question

    Meta is preparing Hatch as its first paid AI product, a consumer agent that completes multi-step tasks across apps. The Decoder reported a premium tier priced as high as $199.99 per month, with free-tier limits five to ten times lower.

    Testing has reportedly involved DoorDash, Reddit and Etsy. That is a consumer commerce agent, not a work seat. Treat it as a signal about where price ceilings are heading, not a procurement option — and note it will migrate to Meta’s own model, which we sized up in Muse Spark vs Claude Opus 5.

    Which AI seat should your team buy in 2026?

    Match the seat to the workload, not the org chart. Engineering teams should default to Claude. Mixed-function teams should default to OpenAI. Companies already locked into E5 licensing should audit whether Copilot is duplicating spend before renewing.

    Recommendations by use case

    Your situation Buy this Monthly cost Why
    2–10 engineers, agentic coding daily Claude Team premium $100–$125/seat Claude Code included, no separate metering
    Mixed team, research + docs + some code ChatGPT Business standard $20–$25/seat Widest feature surface per dollar
    Heavy agent workloads, 150+ people Claude Enterprise $20/seat + API usage Pay real token rates, not a 5x flat markup
    Already on Microsoft 365 E5 Copilot, cautiously ~$87–$90 all-in Only if native Office integration is the point
    Under 300 seats, cost-sensitive SMB Copilot Business $18 promo / $21 std Cheapest Microsoft-native route until Dec 31, 2026
    Solo operator or two-person startup Claude Pro or ChatGPT Plus $17–$20 Team seats add governance you do not need yet

    The move most teams should make first

    Do not buy premium seats on day one. Buy standard seats, watch who hits the cap for two consecutive weeks, and upgrade only those people.

    At $1,200 per premium seat annually, guessing wrong on a ten-person team costs $12,000 a year for capacity nobody used.

    Frequently asked questions

    Is Claude Team cheaper than ChatGPT Business?

    No. They are identical: $25 monthly or $20 annually for standard seats, $125 monthly or $100 annually for premium. The differentiator is what each seat includes, not what it costs.

    Do premium seats include the coding agent?

    Claude includes Claude Code on every paid tier per Anthropic’s pricing page. OpenAI bills Codex against your per-seat allowance, so a premium seat buys more Codex headroom rather than separate access.

    What is the seat minimum for each plan?

    Both ChatGPT Business and Claude Team start at two seats. Anthropic caps self-serve Team plans at 150 people, above which you move to Enterprise.

    Is Claude Enterprise better value than Team premium?

    For heavy agentic use, yes. Claude Enterprise bills $20 per seat plus usage at API rates, so you pay the real cost of tokens instead of a flat 5x markup on a capacity estimate.

    Why did OpenAI add premium seats now?

    Agents consume far more tokens than chat. OpenAI announced the tier on August 11, 2026, citing teams tackling more complex tasks — which is another way of saying standard seats were unprofitable for power users.

    Does Microsoft Copilot really cost $30?

    Only as an add-on. A qualifying Microsoft 365 license is required, taking realistic all-in cost to roughly $66–$69 per user on E3 or $87–$90 on E5.

    Should I wait for Meta Hatch?

    No. Hatch has no confirmed public release date, no confirmed price, and appears aimed at consumers rather than teams. Buy the seat you need this quarter.

    The bottom line

    Claude Team is the better buy for engineering-led organizations, and it is not close. Claude Code ships on every paid tier, terminal work is not metered against your product manager’s research queries, and Claude Enterprise offers the only pricing model in this comparison that scales honestly with consumption at $20 per seat plus API rates.

    ChatGPT Business wins exactly one scenario: teams where fewer than half the seats write code. Breadth beats depth when the average user needs research, documents and images more than a terminal.

    Microsoft Copilot is the weakest value in the group unless deep Office integration is a hard requirement, because the stacked-license structure roughly triples the advertised price.

    The strategic read matters more than the seat choice. Two competing labs landed on identical $25 and $125 price points within weeks of each other. That is not coincidence — it is a market discovering the clearing price for agentic compute. Expect the next move to be upward, and lock annual pricing while $100 premium seats exist. For the token-level economics underneath these plans, see our breakdown in Gemini 3.7 Flash vs Claude Sonnet 5 and our guide to choosing a replacement model after a deprecation.

    Sources

  • Etched Valuation Doubles to $21 Billion in Under a Month

    Etched raised $700 million at a $21 billion post-money valuation on August 18, 2026, led by quant trading firm Jane Street — which is also its first paying customer. The Etched valuation doubled from $10.3 billion in July, according to TechCrunch. The transformer-ASIC startup has booked more than $1 billion in signed orders and has now raised close to $2 billion in total.

    A four-year-old chip company just repriced itself faster than almost anything in the AI hardware cycle. The question is whether the order book justifies it.

    How much did Etched raise, and at what valuation?

    Etched raised $700 million at a $21 billion post-money valuation, announced Tuesday, August 18, 2026. Jane Street led the round. That is a doubling from the $10.3 billion Series C the company closed in July 2026 — roughly four weeks earlier, per TechCrunch’s reporting.

    The step-up is the headline. In December 2025 Etched was worth $5 billion. Eight months later it is worth $21 billion, a 4.2x move without a single public revenue disclosure.

    The Etched valuation timeline

    Date Round Valuation Lead investor
    December 2025 Series B extension $5 billion Not disclosed
    July 2026 Series C, $300M $10.3 billion Sequoia Capital
    August 18, 2026 $700M round $21 billion Jane Street

    Total capital raised is now close to $2 billion, according to Tech Startups. The cap table includes Sequoia Capital, Kleiner Perkins, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Stripes, Primary, Positive Sum and Blackstone.

    Why is Jane Street both the lead investor and the first customer?

    Jane Street tested Etched’s system, installed a rack in its own datacenter, and then led the round. That dual role is the most important detail in the announcement — and the one that deserves the most scrutiny. A lead investor who is also the reference customer validates the product and inflates the comparable at the same time.

    “Etched’s unique approach to inference delivers the precision we will need to support our most demanding workloads,” Jane Street said in the announcement, adding that it has “our own rack running in our datacenter.”

    Quant trading is an unusually favorable first market. Latency is worth real money there, the workloads are narrow and stable, and the buyer has no procurement committee. Whether that translates to hyperscalers running heterogeneous model fleets is a genuinely open question.

    What Etched actually ships

    The company sells what it calls frontier inference clusters, built around two custom components: a low-voltage prefill chip and cluster-scale memory sized for the decode phase. Its Sohu part is marketed as the world’s first transformer ASIC.

    Etched says it went from receiving test silicon at TSMC to running inference workloads in 44 days — fast for a first-silicon bring-up, where months is normal.

    What is a transformer ASIC, and why does it threaten Nvidia?

    A transformer ASIC hard-codes one model architecture into silicon instead of staying programmable. You lose flexibility and gain throughput and power efficiency. The bet is that transformers stay dominant long enough for fixed-function chips to pay back their tape-out cost before the architecture moves.

    Nvidia’s moat is generality plus CUDA. An ASIC attacks exactly the workload where generality is least valuable: high-volume, steady-state inference of one model family.

    Etched has also been buying the expertise directly. Roughly 15% of its ~400 employees came from Nvidia — about 60 people, per Tech Startups. Systems engineer Brian Loiler, who spent 23 years at Nvidia before joining in 2024, recruited around a dozen more Nvidia engineers; some turned down counteroffers.

    Founded in 2022 by Harvard dropouts Gavin Uberti and Chris Zhu, the company operates from San Jose with an internal datacenter. Sequoia GP Sonya Huang summed up the historical skepticism the round is arguing against: “Don’t back the kids in chips.”

    Who else is buying into custom AI silicon right now?

    Etched is not an outlier — it is the loudest datapoint in a two-week run of custom-silicon deals. Three of the largest chip buyers and builders in the market all moved on inference-specific hardware in August 2026.

    • AMD acquired Taalas on August 6, 2026, for an undisclosed sum. Taalas etches models directly into silicon; AMD says it will fold the technology into its accelerator roadmap alongside Instinct GPUs.
    • Marvell granted Google a warrant for up to 58.97 million shares — worth up to $12.2 billion — tied largely to purchasing targets through fiscal 2033. Marvell stock jumped more than 11% in premarket trading on August 19.
    • Nvidia itself is hedging, spending on the layer below the chip: it took a stake in datacenter developer Cloverleaf and, as we covered, paid roughly $6 billion for Poolside’s model factory and 109 staff.

    The pattern is consistent. Everyone with capital is buying inference efficiency, and they are paying acquisition-grade prices for it. The same dynamic is visible in the fast-inference API market, where specialized silicon is already competing on price per token.

    Why this matters for the AI market and investors

    Inference, not training, is now where the compute money goes — and that is the market Etched is built for. Nvidia posted $81.6 billion in total revenue in Q1 FY2027, with $75.2 billion from Data Center, up 92% year over year, according to the company’s own results release for the quarter ended April 26, 2026.

    Against that, $21 billion for a startup with roughly $1 billion in signed orders is a bet on share shift, not on displacement. Etched would need to compound for years to matter to Nvidia’s income statement.

    The more useful read is directional. Capital is repricing the assumption that general-purpose GPUs capture all inference margin. That assumption also underwrites the debt now funding the buildout — see the Broadcom financing package and Nvidia’s decision to cut its OpenAI datacenter guarantee from $250 billion to $120 billion.

    The skeptical case

    A valuation that doubles in four weeks on the same order book is a financing event, not an operating one. Nothing in the disclosed figures changed between July and August except who was writing the check.

    Three specific risks:

    1. Architecture risk. A transformer ASIC is a leveraged bet that transformers stay dominant. If the frontier moves to a materially different architecture, the silicon does not follow.
    2. Customer concentration. One named customer, who is also the lead investor. Signed orders above $1 billion are unaudited and self-reported.
    3. Incumbent response. Nvidia has $75.2 billion of quarterly Data Center revenue to defend a niche with, and AMD just bought a direct competitor to Etched’s approach.

    This post is reporting and analysis, not financial advice.

    Frequently asked questions

    What is the Etched valuation now?

    $21 billion post-money, set by the $700 million round announced August 18, 2026.

    Who led Etched’s $700 million round?

    Jane Street, the quantitative trading firm, which is also Etched’s first paying customer and has a rack running in its own datacenter.

    How much has Etched raised in total?

    Close to $2 billion, according to Tech Startups. The prior round was a $300 million Series C at $10.3 billion in July 2026.

    What does Etched’s Sohu chip do?

    Sohu is marketed as the world’s first transformer ASIC — silicon purpose-built for transformer inference rather than general-purpose computation, trading flexibility for throughput and power efficiency.

    How many Etched employees came from Nvidia?

    About 15% of roughly 400 employees, or around 60 people, per Tech Startups.

    Is Etched profitable?

    The company has not disclosed revenue or profitability. It reports more than $1 billion in signed orders and has begun shipping chips.

    Who are Etched’s founders?

    Gavin Uberti and Chris Zhu, Harvard dropouts who founded the company in 2022. Robert Wachen is co-founder and COO.

    The bottom line

    Etched has the most aggressive valuation trajectory in AI hardware and a real product shipping into a real datacenter. It also has one named customer who set the price.

    Watch two things next: whether a hyperscaler or a frontier lab signs, and whether the $1 billion order book converts to disclosed revenue. If a second, unaffiliated buyer of scale appears in the next two quarters, $21 billion looks early. If it does not, this round will read as the moment ASIC enthusiasm outran ASIC demand.

    Sources

  • Best o3 Replacement: What to Use After the August 26 Cutoff

    OpenAI retires o3 from ChatGPT on August 26, 2026. The best o3 replacement for most teams is GPT-5.6 Terra at $2/$12 per million tokens — not the officially recommended Sol at $5/$30. Terra posts 90.4% on GPQA Diamond against o3’s 87.7%, at 60% less output cost. API users are not on the same clock: their o3 shutdown is December 11.

    The reasoning model that defined 2025 is being switched off. And the migration advice OpenAI published is the expensive option.

    Here is what actually changes on Tuesday, what each o3 replacement costs, and which one wins for your workload.

    What exactly happens to o3 on August 26, 2026?

    o3 disappears from the ChatGPT model picker on August 26, 2026. That is a product change, not an API shutdown. OpenAI’s release notes from May 28, 2026 confirm o3 is “retired from ChatGPT on August 26, 2026 following a 90-day sunset period.” Developers calling o3 over the API keep working past that date.

    The ChatGPT side: the picker already stopped naming models

    ChatGPT users lose nothing they can still see. Since the June 10, 2026 model picker update, OpenAI stopped exposing version numbers entirely.

    The picker now offers Instant, Medium, High, and Extra High, plus Pro Standard and Pro Extended on Pro plans. Reasoning is sold as effort, not as a model name.

    So for consumer subscribers, the o3 replacement is already installed. Selecting High is the closest analogue to what o3 Thinking used to do.

    The API side: your real deadline is December 11

    This is where most coverage gets it wrong. OpenAI’s deprecations page lists o3-2025-04-16 with a shutdown date of December 11, 2026, migrating to gpt-5.6-sol. o3-pro-2025-06-10 follows the same date, moving to Sol with reasoning.mode: pro.

    What does die on August 26 is the legacy Assistants API. Anything built on Assistants must move to the Responses and Conversations APIs by Tuesday. That is the deadline worth panicking about.

    Which o3 replacement is best for most workloads?

    GPT-5.6 Terra. OpenAI names Sol as the official successor, but Sol is priced for frontier reasoning at $5/$30 per million tokens. Terra sits at $2/$12 and clears o3 on the benchmarks that mattered to o3 users. For the overwhelming majority of o3 traffic, paying Sol rates is a rounding error you repeat a million times.

    Terra got cheaper on July 30, 2026, when OpenAI cut its price roughly 20% as part of the GPT-5.6 pricing refresh. Luna fell about 80% in the same announcement, to $0.20/$1.20.

    The quality case is not a stretch either. OpenRouter’s provider data puts Terra at 90.4% on GPQA Diamond and 75.3% on TAU-Bench. o3 scored 87.7% on GPQA Diamond at launch.

    Why Terra beats Sol on cost per task

    Reasoning models bill you for tokens you never see. A 500-token visible answer can consume 2,000+ tokens once hidden reasoning is counted.

    That multiplier is exactly why output price dominates your bill. At $12 versus $30 per million output tokens, Terra cuts the expensive half of the invoice by 60%.

    Terra also carries a 1,050,000-token context window with 128,000 max output — over five times o3’s 200K context. You are not trading capability down.

    How much does each o3 replacement cost?

    Prices below are list rates per million tokens as of August 24, 2026, pulled from provider pricing pages. Cached input bills at 10% of standard rates on OpenAI, and the Batch API halves both sides.

    Model Input Output Context Notes
    o3 (retiring) $2.00 $8.00 200K API shutdown Dec 11, 2026
    o3-pro (retiring) $20.00 $80.00 200K API shutdown Dec 11, 2026
    GPT-5.6 Terra $2.00 $12.00 1.05M Cut ~20% on Jul 30, 2026
    GPT-5.6 Sol $5.00 $30.00 1.05M Official o3 successor
    GPT-5.6 Luna $0.20 $1.20 1.05M Cut ~80% on Jul 30, 2026
    Claude Opus 5 $5.00 $25.00 200K Fast Mode is $10/$50
    Claude Sonnet 5 $2.00 $10.00 200K $2/$10 now permanent
    Gemini 3.7 Flash $0.75 $3.75 — Doubles Jan 1, 2027
    Kimi K3 $2.60 $13.00 1M Open weights, 2.8T params

    Read that table one way and the story is obvious: o3 at $2/$8 was cheap, and every direct successor except Luna and Gemini Flash costs more per output token. Migration is a price increase unless you choose deliberately.

    Is GPT-5.6 Sol worth $5/$30 in 2026?

    Only for the top slice of your traffic. Sol is the frontier tier and the only model OpenAI formally maps o3-pro onto, via reasoning.mode: pro. If you were paying o3-pro’s $20/$80, Sol at $5/$30 is a 75% input cut and a 62.5% output cut.

    If you were on standard o3, Sol is a 150% input increase and a 275% output increase. Same model family, opposite financial outcome.

    There is also a long-context trap. Requests beyond the standard threshold reprice: Sol rises to $10/$45, Terra to $4/$18, Luna to $0.40/$1.80. Feeding a million-token repo into Sol is a different product than a 20K-token prompt.

    Which o3 replacement should you pick for your use case?

    Match the model to the job, not to the vendor’s migration note. Below is where each option earns its price, based on published benchmarks and list pricing. Benchmark your own top 20 prompts against Terra before escalating anything to Sol.

    Use case Pick Why
    General o3 traffic GPT-5.6 Terra 90.4% GPQA Diamond at $2/$12; 1.05M context
    Hardest reasoning, o3-pro traffic GPT-5.6 Sol Official reasoning.mode: pro path; 62.5% cheaper output than o3-pro
    High-volume classification GPT-5.6 Luna $0.20/$1.20 after the ~80% July cut
    Agentic coding, terminal work Gemini 3.7 Flash 85.8% on Terminal-bench 2.1 at $0.75/$3.75
    Long document analysis Claude Opus 5 $5/$25 with 200K context, cheaper output than Sol
    Balanced daily driver Claude Sonnet 5 $2/$10 locked permanently
    Self-hosting, data residency Kimi K3 2.8T open weights, 1M context, $2.60/$13 hosted

    What about Claude and Gemini as an o3 replacement?

    Both are live options, and both moved on price in the last two weeks. Anthropic canceled a scheduled increase; Google launched a discount with an expiry date attached. Those two facts change the math more than any benchmark did.

    Claude Sonnet 5’s price freeze quietly killed a 50% increase

    Sonnet 5 launched at $2/$10 as introductory pricing set to expire August 31, 2026, with a jump to $3/$15 scheduled for September 1. Anthropic’s pricing documentation now states that increase “will not occur” and $2/$10 is the standard price.

    That makes Sonnet 5 an exact price match to o3 on input and 25% more on output — the closest financial like-for-like swap available. We broke down how it stacks up against Google’s cheap tier in Gemini 3.7 Flash vs Claude Sonnet 5.

    Gemini 3.7 Flash is the cheapest credible option — until January

    Gemini 3.7 Flash shipped August 13, 2026 with a 50% introductory cut to $0.75/$3.75. On January 1, 2027 it reverts to $1.50/$7.50, and context caching moves from $0.075 to $0.15.

    Its numbers are strong where agents live: 85.8% on Terminal-bench 2.1 and 65.3% on DeepSWE v1.1, though only 43.6% on FrontierCode 1.1 Main. Build your 2027 budget on the standard rate, not the promo.

    How do you migrate off o3 without breaking production?

    Treat this as a pricing audit, not a find-and-replace. The single most expensive mistake is routing all o3 traffic to Sol because the deprecation table said so. Work through it in this order:

    1. Split the Assistants API work out first. It dies August 26, 2026 — 107 days before o3 does. Move to Responses and Conversations now.
    2. Pull your last 30 days of o3 token spend and split it by input versus output. Output volume decides which tier you can afford.
    3. Replay your top 20 prompts against Terra. If quality holds, you are done at $2/$12.
    4. Escalate only the failures to Sol. Route by task difficulty, not by default.
    5. Push bulk classification to Luna or Gemini 3.7 Flash. At $0.20/$1.20, Luna makes some batch jobs nearly free.
    6. Turn on prompt caching and the Batch API. Cache hits bill at 10%; batch halves everything.

    Teams running agentic coding harnesses should also re-check their tooling layer, not just the model. We covered that trade-off in Claude Code vs Codex CLI.

    Frequently asked questions about the o3 replacement

    Is o3 gone from the API on August 26, 2026?

    No. August 26 is the ChatGPT retirement date. OpenAI’s deprecations page lists the o3-2025-04-16 API shutdown as December 11, 2026, with gpt-5.6-sol as the migration target.

    What is the cheapest o3 replacement?

    GPT-5.6 Luna at $0.20/$1.20 per million tokens, following its roughly 80% price cut on July 30, 2026. For work needing more reasoning depth, Gemini 3.7 Flash at $0.75/$3.75 is the next step up.

    Does GPT-5.6 Terra actually beat o3?

    On GPQA Diamond, yes: 90.4% for Terra on OpenRouter’s provider data versus 87.7% for o3 at launch. Terra also carries a 1.05M context window against o3’s 200K.

    What replaces o3-pro?

    GPT-5.6 Sol with reasoning.mode: pro, per OpenAI’s deprecations table. At $5/$30 versus o3-pro’s $20/$80, it is a substantial price cut for that specific tier.

    Will Claude Sonnet 5 get more expensive in September?

    No. Anthropic canceled the September 1, 2026 increase to $3/$15. The $2/$10 rate is now permanent per its pricing docs.

    Is there an open-weights o3 replacement?

    Kimi K3 is the closest: 2.8 trillion parameters, 1M context, released July 16, 2026, and $2.60/$13 hosted on OpenRouter. Chinese open-weight coders are also competitive — see our GLM-5.3 vs DeepSeek V4 Pro breakdown.

    What breaks on August 26 if I do nothing?

    Two things: o3 vanishes from the ChatGPT picker, and the legacy Assistants API stops working. API calls to o3 itself keep running until December 11, 2026.

    The bottom line

    Move general o3 traffic to GPT-5.6 Terra at $2/$12 and stop there. It beats o3 on GPQA Diamond, gives you five times the context, and costs 60% less per output token than the Sol tier OpenAI points you at.

    Send only o3-pro-class work to Sol, where $5/$30 is genuinely a 62.5% output discount on what you were paying. Send bulk work to Luna at $0.20/$1.20.

    The decision depends on exactly one number: your output-token share. Above roughly 30% of spend, tier choice dominates everything else on your invoice. Below that, input caching matters more than which model you pick.

    And fix your Assistants API code before Tuesday. That is the only hard deadline this week. For more on where frontier pricing is heading, see our analysis of DeepSeek’s vision model against Claude Opus 4.8.

    Sources

  • DeepSeek Vision Model vs Claude Opus 4.8: 23x Cheaper Per Image

    DeepSeek’s vision model just split a 2-2 draw with Claude Opus 4.8 on visual benchmarks while charging roughly 23 times less per image. V4-Flash-Vision-Exp processes 1,000 images for about $0.17 at peak rates. Claude Sonnet 4.6 charges around $4 for the same work. If your workload is image-heavy and volume-driven, the math is no longer close.

    The DeepSeek vision model shipped on August 21, 2026. It is experimental, API-only, and it does something DeepSeek has done before: it removes a price floor the incumbents were standing on.

    What is DeepSeek’s new vision model?

    V4-Flash-Vision-Exp is an experimental multimodal extension of DeepSeek’s V4-Flash, adding image understanding to a model that was previously text-only. It is available on DeepSeek’s paid developer platform by setting model='deepseek-v4-flash-vision-exp', per the official API changelog.

    The target is not image generation. It is visual agent work — reading charts, parsing screenshots, driving interfaces, and running multi-step tasks that require seeing what happened.

    DeepSeek’s own framing is careful. The changelog calls it “a significant leap over DeepSeek-V4-Flash, bringing its multimodal agent capabilities close to Opus-4.8,” while noting pure-text performance is unchanged from the base model.

    What’s under the hood

    The foundation is V4-Flash, released in April 2026: a 284-billion-parameter mixture-of-experts model built from multiple 13-billion-parameter networks, according to SiliconANGLE.

    The vision specs are aggressive. A 1-million-token context window, up to 384,000 output tokens, and up to 600 images per request. Each image is capped at 384 tokens regardless of size — that cap is the entire reason the pricing works.

    Supported formats are JPEG, PNG, GIF and WebP, with a maximum edge of 8,192 pixels, dropping to 4,096 pixels once you send 15 or more images in a single call.

    Did the DeepSeek vision model actually beat Claude Opus 4.8?

    On two benchmarks yes, on two no. It is a genuine 2-2 split, not a sweep, and the margins are thin in both directions. DeepSeek wins the agentic and hard-visual tests. Opus 4.8 holds the chart-reading and structured-reasoning tests. Anyone claiming a decisive winner is selling something.

    Benchmark DeepSeek V4-Flash-Vision-Exp Claude Opus 4.8 Margin
    Agents’ Last Exam (ALE) 27.3 25.7 +1.6 DeepSeek
    ZeroBench (Pass@5) 35.0 34.0 +1.0 DeepSeek
    ApexBench (Pass@1) 36.5 39.4 +2.9 Opus 4.8
    Chartography 64.3 65.0 +0.7 Opus 4.8

    Scores as reported by XenoSpectrum’s breakdown. DeepSeek’s changelog separately lists a Terminal Bench 2.1 score of 83.9.

    Context on the tests matters. ALE contains more than 1,000 multi-step tasks; ZeroBench is 100 image-analysis tasks deliberately built to be hard, SiliconANGLE reports. Those are the two DeepSeek won.

    The pattern is consistent with what we saw when GLM-5.3 went up against DeepSeek V4 Pro: Chinese labs are now trading benchmark wins with frontier Western models rather than trailing them.

    How much does the DeepSeek vision model cost?

    Very little, and the structure is unusual. DeepSeek runs peak and off-peak pricing. Off-peak, input is $0.22 per million tokens and output is $0.66. During peak hours — 01:00–04:00 and 06:00–10:00 UTC — both double, to $0.44 and $1.32. Cache hits cost $0.007 off-peak.

    The image economics fall out of the 384-token cap. A single image costs $0.00008448 off-peak and $0.00016896 at peak.

    Model ~Cost per 1,000 images (1024×1024) Input / output per 1M tokens Context
    DeepSeek V4-Flash-Vision-Exp (off-peak) ~$0.09 $0.22 / $0.66 1M
    DeepSeek V4-Flash-Vision-Exp (peak) ~$0.17 $0.44 / $1.32 1M
    Google Gemini 3.1 Pro ~$0.50 $2.00 / $12.00 1M
    OpenAI GPT-5.4 Vision ~$1.90 $2.50 / $15.00 1.1M
    Anthropic Claude Sonnet 4.6 ~$4.00 — —
    Anthropic Claude Opus 4.8 — $5.00 / $25.00 1M

    Per-image costs are as calculated by BigGo Finance; Opus 4.8 token rates are $5.00 input, $25.00 output and $0.50 cached input, per OpenRouter.

    Set those side by side. On output tokens, DeepSeek off-peak at $0.66 against Opus 4.8 at $25.00 is a 37x spread for a model that just won two of four visual benchmarks against it.

    Why the 384-token image cap matters

    Most vision APIs price images by resolution. Send a bigger image, pay more tokens. That makes cost forecasting miserable and punishes exactly the high-resolution documents that need vision in the first place.

    DeepSeek caps every image at 384 tokens no matter its size, up to an 8,192-pixel edge. A dense scanned contract and a low-resolution thumbnail cost the same.

    For finance and operations teams, that turns a variable line item into a fixed one. You can price a document-processing product per page before you have seen a single customer’s files, which is not something the resolution-scaled competitors let you do.

    Is the DeepSeek vision model cheap enough to switch?

    For high-volume image pipelines, yes — the gap is too large to argue with. For low-volume work where quality variance costs more than tokens, no. The deciding number is how many images you process per month, because below roughly 100,000 images the absolute savings stop mattering.

    Run it concretely. A million images per month costs about $170 at DeepSeek’s peak rate, $90 off-peak, and roughly $4,000 on Claude Sonnet 4.6. That is a $46,000 annual difference on a single pipeline.

    • Document and invoice processing at scale. Switch. The 384-token cap makes cost per page nearly flat.
    • Screenshot-driven agents and RPA. Switch. ALE and ZeroBench are exactly this workload, and DeepSeek won both.
    • Chart and dashboard extraction. Stay. Opus 4.8 leads Chartography 65.0 to 64.3, and errors here are expensive.
    • Regulated or air-gapped work. Stay. This is a Chinese paid API with no open weights announced.
    • Batch archive backfill. Switch, and schedule it off-peak. You halve the bill by choosing when to run.

    That off-peak lever is genuinely underrated. Almost no batch job cares what hour it runs, and DeepSeek is effectively paying you 50% to move it.

    Which vision model should you use for what?

    Match the model to the failure cost, not to the benchmark chart. Where a wrong answer is cheap and volume is enormous, take the cheapest capable model. Where a wrong answer triggers a human review cycle or a compliance event, the token savings are irrelevant.

    Use case Pick Why
    Bulk OCR, receipts, invoices DeepSeek V4-Flash-Vision-Exp ~$0.09–$0.17 per 1,000 images; flat 384-token cost per image
    Visual agents, UI automation DeepSeek V4-Flash-Vision-Exp Wins ALE (27.3) and ZeroBench (35.0) head-to-head
    Financial charts, data extraction Claude Opus 4.8 Leads Chartography 65.0 and ApexBench 39.4
    Mixed text-plus-vision reasoning Claude Opus 4.8 DeepSeek’s text performance is unchanged from V4-Flash
    Mid-volume general vision Gemini 3.1 Pro ~$0.50 per 1,000 images with Western data residency
    Production systems needing stability Claude Opus 4.8 or Gemini DeepSeek’s model is explicitly labeled experimental

    What are the catches?

    Three, and none of them are small. The model is experimental, it is closed, and the benchmark margins are inside the noise band. Anyone budgeting a migration on a 1.0-point ZeroBench edge is over-reading the data.

    Start with the “Exp” in the name. DeepSeek ships experimental endpoints and retires them. Building a production pipeline on one is a scheduling risk, not just a quality risk.

    Second, there are no open weights. SiliconANGLE notes DeepSeek “may release a free version later on,” but nothing is confirmed. That kills the self-hosting escape hatch that made Qwen3.8-Max interesting for teams with data-residency constraints.

    Third, the wins are narrow. +1.6 on ALE and +1.0 on ZeroBench are real but slim, and DeepSeek loses ApexBench by a wider 2.9 points than either of its victories. The honest summary is parity, delivered at a fraction of the price — which is still a devastating result for Anthropic’s pricing power.

    Frequently asked questions

    What is DeepSeek V4-Flash-Vision-Exp?

    An experimental multimodal model released on August 21, 2026, extending DeepSeek’s V4-Flash with image understanding. It handles up to 600 images per request with a 1-million-token context window, available only through DeepSeek’s paid API.

    Is the DeepSeek vision model better than Claude Opus 4.8?

    It splits 2-2. DeepSeek wins Agents’ Last Exam (27.3 vs 25.7) and ZeroBench Pass@5 (35.0 vs 34.0). Opus 4.8 wins ApexBench Pass@1 (39.4 vs 36.5) and Chartography (65.0 vs 64.3).

    How much does the DeepSeek vision model cost per image?

    A single image is capped at 384 tokens, costing $0.00008448 off-peak and $0.00016896 at peak. That works out to roughly $0.09 to $0.17 per 1,000 images.

    What are DeepSeek’s peak and off-peak hours?

    Peak hours run 01:00–04:00 and 06:00–10:00 UTC, when input and output rates double to $0.44 and $1.32 per million tokens. Everything outside those windows bills at $0.22 and $0.66.

    Are the weights open?

    No. V4-Flash-Vision-Exp is available only through DeepSeek’s paid developer platform. Reports suggest a free version may follow, but nothing has been confirmed.

    Does it hurt text performance?

    No. DeepSeek’s changelog states the model stays “on par with the official DeepSeek-V4-Flash” on agent reasoning, world knowledge, and other text-based tasks. The vision capability is additive.

    Should I migrate my production vision pipeline?

    Only if you process more than roughly 100,000 images per month and can tolerate an experimental endpoint. Below that volume, the absolute savings do not justify the migration and stability risk.

    The bottom line

    Move your high-volume image workloads to DeepSeek V4-Flash-Vision-Exp and keep your judgment-heavy visual work on Claude Opus 4.8. That is the split the benchmarks actually support, and it is where the money is.

    The specific trigger: if you process more than 100,000 images a month and a wrong answer costs you less than a human review, switch, and schedule the batch off-peak to halve the bill again. Below that volume, or where a visual error creates a compliance event, the $4-per-1,000 Anthropic bill is cheap insurance.

    The larger read is about pricing power, not vision. Anthropic charges $25 per million output tokens on Opus 4.8. DeepSeek charges $0.66 off-peak and just traded benchmark wins with it. The same compression already hit text and coding models, and multimodal was the last premium tier standing.

    Sources

  • Nvidia AVO Hits 100% on ARC-AGI-3. The Model Alone Scored 30.2%.

    Nvidia AVO — Agentic Variation Operators — scored 100.00 RHAE on the ARC-AGI-3 public set on August 21, 2026, clearing all 183 levels across 25 environments in 6,624 actions, roughly 12% fewer than the VISTA baseline. The same base model, Claude Opus 5, scores 30.2% on its own. The harness did the work, and that changes where agent money goes.

    What is Nvidia AVO?

    Nvidia AVO stands for Agentic Variation Operators. It is not a model. It is a general-purpose coding-agent system that wraps an existing frontier model in a loop — inspect, plan, implement, evaluate — plus persistent memory and a supervisor that intervenes when progress stalls. Nvidia published the results on August 21, 2026.

    The base model inside the winning run was Anthropic’s Claude Opus 5. Nvidia also ran limited experiments with GPT-5.6 Sol on a subset of games, and labeled those findings preliminary.

    That detail is the whole story. Nvidia did not train a better reasoner. It built better scaffolding around someone else’s reasoner.

    How the AVO loop works

    AVO runs a four-step cycle: inspect the current context, plan a change, implement it, then evaluate the result against the environment.

    Two additions separate it from a standard agent loop, according to Nvidia’s technical blog:

    • Persistent memory that carries forward prior implementations, evaluation results and reasoning across the whole run, not just the current context window.
    • A supervision mechanism that watches the trajectory and redirects the agent when it detects the run has stopped making progress.
    • Variation operators that generate structured alternatives rather than retrying the same failed approach.

    Why the supervisor is the expensive part

    Long-horizon agent failure is rarely a single wrong answer. It is a slow drift — the agent loops on a dead approach and burns tokens without noticing.

    A supervisor that detects stagnation is cheap to describe and hard to build. It is also the component least likely to transfer cleanly to another benchmark.

    What is ARC-AGI-3 and why does a 100% score matter?

    ARC-AGI-3 is ARC Prize’s interactive reasoning benchmark: 25 pixel-art puzzle environments containing 183 public levels. Agents get no instructions, no rules and no goal labels. They must infer the mechanics purely by playing. When the benchmark launched, humans cleared 100% of environments and the best AI managed 0.37%.

    That 0.37% figure is why this result registered. ARC-AGI-3 was designed as the benchmark models could not touch.

    How RHAE scoring works

    The metric is RHAE — Relative Human Action Efficiency. It combines task completion with how many actions the agent needed per level, measured against initial human performance, then aggregates across every level and environment.

    So a 100.00 does not just mean “finished everything.” It means finishing everything at roughly human action efficiency. ARC Prize published its human performance dataset specifically so this number would have a floor to sit on.

    How much did the harness add versus the raw model?

    The gap is 30.2% to 100.00 — the same model class, wrapped differently. ARC Prize reported Claude Opus 5 at 30.2% on ARC-AGI-3 in July 2026, which it called a genuine reasoning leap at the time. Nvidia’s harness took that model to a clean sweep of the public set.

    SystemARC-AGI-3 resultActions usedReported by
    Best AI at benchmark launch0.37%—ARC Prize
    Claude Opus 5 (bare model)30.2%—ARC Prize, July 2026
    VISTA baseline agentCleared same level sets7,542Nvidia
    Nvidia AVO (Claude Opus 5 inside)100.00 RHAE, all 183 levels6,624Nvidia, Aug 21 2026

    For context on the base model’s ceiling elsewhere: Claude Opus 5 at maximum reasoning effort scores 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2, per The New Stack. ARC-AGI-3 was the one that stayed hard.

    Nvidia’s own framing, from the blog post: “The model matters, but the model is not the entire agent.”

    Is the 100% score real, or is it benchmark theater?

    It is real, and it is narrower than the headline suggests. The score covers the ARC-AGI-3 public set only — not the semi-private or private competition sets that ARC Prize maintains precisely to catch overfitting. Nvidia says so in its own post.

    The public set is not the hidden exam

    Benchmark designers split datasets for a reason. A public set is a practice test with the answers eventually leaking into the ecosystem around it.

    One commenter on Nvidia’s announcement put it bluntly, as flagged in explainX’s write-up: “I would not file this as solved AGI… if you post 100 like it is the hidden exam.” Until AVO posts a semi-private number, that objection stands.

    These are not controlled ablations

    Nvidia explicitly labels its comparisons as not controlled ablations. That matters more than it sounds.

    The AVO-versus-VISTA action count — 6,624 against 7,542 — varies agent backends, observation formats, memory systems and reasoning settings all at once. The 100.00-versus-30.2% comparison swaps the entire system architecture and the reasoning-effort setting simultaneously.

    Neither number isolates how much the harness itself contributed. The honest reading is “a well-built harness closed a very large gap,” not “the harness is worth exactly 70 points.”

    Nvidia also disclosed no compute cost, no token usage and no wall-clock runtime for the 6,624 actions. For anyone pricing an agent product, that is the number that actually matters — and it is missing.

    Has AVO done anything useful outside a puzzle benchmark?

    Yes, and this is the part investors should read twice. Nvidia ran AVO continuously for seven days on GPU-kernel optimization, exploring more than 500 optimization directions. The system produced kernels that beat FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 hardware.

    FlashAttention is not a soft target. It is hand-tuned infrastructure that the entire industry’s inference economics rest on.

    A 10.5% kernel improvement compounds across every token served on that hardware. If it holds in production, it is worth more to Nvidia than the benchmark headline — and it lands in the same week the company has been buying capability outright elsewhere.

    Who wins and who loses financially?

    The winner is whoever owns the orchestration layer. If a 30% model becomes a 100% agent through harness design, then value is accruing above the weights, not inside them. That is bad news for anyone whose entire moat is a checkpoint.

    Nvidia is climbing the stack

    AVO did not appear in isolation. On the same day, Nvidia paid $6 billion to license Poolside’s model-development software and invested $1 billion more in the startup, according to PYMNTS.

    A chip company publishing frontier agent architecture and licensing a model factory in the same 24 hours is not a coincidence. It is a company that has watched its customers capture the margin its silicon creates — the same dynamic behind its recalculated OpenAI data center guarantee.

    Model labs keep pricing power, for now

    Note who supplied the brain: Anthropic. AVO’s best run needed Claude Opus 5, and the harness could not manufacture reasoning that was not already there — the 0.37% launch-day figure is proof that scaffolding alone does nothing on a weak model.

    So frontier labs still sell the scarce input. What they lose is the claim that the model is the product, which shows up quickly in cheaper models closing capability gaps.

    Agent startups just got a harder question

    Three practical implications for anyone building or funding an agent company:

    1. Harness design has not hit diminishing returns. A 30-to-100 jump says the scaffolding layer is still under-engineered — which is opportunity and commoditization risk in the same sentence.
    2. Your differentiator may be a blog post away from replication. Persistent memory plus a stagnation supervisor is a describable architecture, not a trade secret.
    3. Nvidia is now a potential competitor, not just a supplier. It has the hardware, the capital, and as of August 21, published frontier agent research.

    The cost question decides all three. Running a supervised, memory-heavy loop for 6,624 actions is not free, and the economics look very different depending on whether the underlying tokens cost $2 or $60 per million — the same math that drives coding-agent unit costs and inference vendor selection.

    Frequently asked questions about Nvidia AVO

    Is Nvidia AVO a new AI model?

    No. AVO is an agent system — a harness — that runs on top of existing frontier models. The reported 100.00 RHAE run used Claude Opus 5 as its base model.

    Did Nvidia AVO solve AGI?

    No. The score covers ARC-AGI-3’s public set of 183 levels across 25 environments. ARC Prize also maintains semi-private and private sets, and AVO has not posted a result on those.

    What does RHAE mean?

    Relative Human Action Efficiency. It scores both whether an agent completes a level and how many actions it needed relative to initial human performance, aggregated across the benchmark.

    Can developers use AVO today?

    Nvidia’s August 21 post describes the architecture and results. It does not announce a code or weights release, so treat AVO as published research rather than a shippable dependency.

    How much does an AVO run cost?

    Nvidia did not disclose compute cost, token usage or wall-clock time for the benchmark run. Without those figures, the result cannot be compared on a cost-per-task basis against cheaper agent harnesses.

    What was the FlashAttention-4 result?

    Running for seven days across 500-plus optimization directions, AVO produced GPU kernels that outperformed FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 hardware.

    Does this make Claude Opus 5 look better or worse?

    Both. The model was capable enough to be driven to 100.00 RHAE, and weak enough on its own to score 30.2%. The delta belongs to the harness, not the checkpoint.

    The bottom line

    Nvidia AVO is the most important agent result of the month, and the headline number is the least interesting part of it.

    A 100.00 on a public set with no controlled ablations and no disclosed cost is a demonstration, not a benchmark victory. Anyone treating it as “ARC-AGI-3 is solved” is reading a press release as a result.

    What survives scrutiny is the gap: 30.2% to 100.00, same model, different scaffolding. That gap is the clearest evidence yet that in 2026 the agent layer, not the model layer, is where the remaining engineering leverage sits.

    And the FlashAttention-4 kernels are the tell. Nvidia did not build AVO to win a puzzle leaderboard. It built AVO to make its own hardware faster — and, at a moment when record sums are being raised to finance AI chip capacity, to stop being only the company that sells the machines.

    Sources