Wealth Engine

Category: Model Releases

  • Tencent Hy4: 770B Open Weights, 82% Cheaper Output Than Kimi K3

    Tencent Hy4 preview is a 770-billion-parameter open-weight model released under Apache 2.0 on August 28, 2026, with 49B active parameters and a 1,048,576-token context window. Tencent Cloud prices it at $0.834 per million input tokens and $2.501 per million output — roughly one-eighth the output cost of GPT-5.6 Sol and 82% below Kimi K3. Weights are on Hugging Face.

    Tencent open-sourced its largest model to date on Friday, and the interesting number is not the parameter count. It is the price tag attached to weights anyone can download, modify and resell.

    That combination — frontier-adjacent scores, permissive licensing, and output tokens at $2.501 per million — is the part that moves money.

    What is Tencent Hy4 preview?

    Tencent Hy4 preview is a mixture-of-experts language model with 770B total parameters, of which 49B activate per token. Tencent released it on August 28, 2026, published the weights on Hugging Face under the Apache 2.0 license, and shipped it simultaneously into its own consumer and developer products.

    According to Tencent’s announcement, the model is live in WorkBuddy, CodeBuddy, Yuanbao and ima, with API access through Tencent Cloud TokenHub and OpenRouter.

    WorkBuddy and CodeBuddy are free for two weeks from launch. That is a customer-acquisition subsidy, not a pricing model.

    The architecture, briefly

    The Hugging Face model card lists 78 layers — the first a dense FFN, the remaining 77 MoE — with 256 routed experts plus one shared expert per MoE layer, and top-8 routing per token.

    Vocabulary size is 120,832. Attention is what Tencent calls “Gated Sparse Attention with IndexCache,” reusing sparse indices across layers to keep the million-token window affordable to serve.

    There is also a built-in multi-token-prediction layer of 10B parameters (0.7B activated) for speculative decoding. Tencent says system-level optimization lifted end-to-end throughput 31.8% against its own baseline.

    How much does Tencent Hy4 cost?

    Tencent Cloud lists $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million on cache hits. In renminbi terms Tencent quotes ¥6 input and ¥18 output per million tokens.

    Those are the headline numbers, and they are aggressive against every comparable model.

    Per KuCoin’s summary of the launch materials, that input price is 25% below GLM-5.3 and 70% below Kimi K3; on output it is 36% below GLM-5.3 and 82% below Kimi K3.

    Model Input / 1M Output / 1M Context Weights
    Tencent Hy4 preview $0.834 $2.501 1,048,576 Apache 2.0
    GPT-5.6 Sol (base tier) $4.00 $20.00 n/d Closed
    GPT-5.6 Sol (long context) $8.00 $30.00 n/d Closed

    On output — the expensive half of any agentic workload, where the model writes code, calls tools and re-reads its own work — Hy4 runs at roughly one-eighth of GPT-5.6 Sol’s base rate.

    For a coding agent burning 50 million output tokens a month, that is the difference between about $1,000 and about $125. The comparison only holds if the cheaper model finishes the job in a similar number of tokens, which is exactly the assumption worth testing.

    How does Hy4 score on benchmarks?

    Hy4 posts strong software-engineering numbers and slightly trails the closed frontier on reasoning. Tencent reports 92.3 on GPQA Diamond, 65.7 on SWE-Bench Pro, 82.9 on SWE-Bench Multilingual, 62.9 on SkillsBench v1.1, and 64.3 on Deep-SWE.

    The Deep-SWE result is the standout: 64.3 against 28.0 for the previous generation. That is not incremental.

    Against closed models the gap is real but narrow. Hy4 scores 92.3 on GPQA Diamond versus GPT-5.6 Sol’s 94.6, and 85.4 on Terminal-Bench versus Sol’s 88.8.

    • GPQA Diamond: Hy4 92.3 — GPT-5.6 Sol 94.6
    • Terminal-Bench: Hy4 85.4 — GPT-5.6 Sol 88.8
    • SWE-Bench Multilingual: Hy4 82.9
    • SWE-Bench Pro: Hy4 65.7
    • Deep-SWE: Hy4 64.3, up from 28.0

    Every one of those figures is vendor-reported. None has been independently reproduced at the time of writing, and the model has been public for about a day.

    Is Hy4 better than GLM-5.3 and Kimi K3?

    Marginally, on Tencent’s own evidence. Tencent ran a blind evaluation with 163 internal experts across 203 real engineering tasks. Hy4 preview averaged 2.99 out of 4.00, against 2.94 for Kimi K3 and 2.92 for GLM-5.3.

    A 0.05-point spread on a four-point scale is not a capability gap. It is a tie with a favorable rounding.

    The win-rate breakdown is more honest about how close this is. Against GLM-5.3, Hy4 won 46.8% of comparisons, drew 12.8% and lost 40.4%. Against Kimi K3: 51.2% wins, 7.9% draws, 40.9% losses.

    So Hy4 loses roughly two of every five head-to-head comparisons against models that were already open. And per KuCoin’s read of the same materials, Hy4 “did not lead comprehensively in public benchmarks” and lags GLM-5.3 in code and cybersecurity tests.

    The differentiator here is price and license, not raw capability. That is worth saying plainly, because the launch framing does not. If you are choosing between cheap open coders, our GLM-5.3 vs DeepSeek V4 Pro comparison and the GLM-5.3-Flash vs Qwen3.8-Flash-Next breakdown cover the alternatives.

    Should you self-host Hy4 or use the API?

    Self-hosting is viable at a scale that would have required a cluster a year ago. Tencent documents deployment on a single eight-GPU node using the FP8 quantized variant, with prebuilt containers for vLLM and SGLang at tensor-parallel size 8.

    The full BF16 checkpoint and an FP8 build are both published, along with the AngelSlim toolkit for further compression.

    The API case is stronger for anyone under roughly 100 million tokens a month. One eight-GPU node of current-generation accelerators plus the engineer who babysits it will not come in under $2.501 per million output tokens at low volume.

    The self-host case is stronger for three groups: teams with data-residency constraints, teams already running GPU capacity at low utilization, and teams that want to distill Hy4 into something smaller. Apache 2.0 permits all three without a negotiation. Our earlier analysis of Qwen3.8-Max open weights versus API walks through that math in detail.

    Who wins and loses financially?

    The clearest loser is anyone selling mid-tier closed inference. If a 770B open model at $2.501 output holds up in production, the price umbrella over $20-per-million output tiers gets thinner.

    Kimi K3 is the most directly exposed. An 82% output-price gap against a model that wins only 51.2% of blind comparisons is a hard position to hold.

    Tencent wins on distribution, not on API margin. At ¥6 per million input, the API is a loss leader that routes developers toward Tencent Cloud, WorkBuddy and CodeBuddy — the same playbook that made cheap Chinese inference a strategic instrument rather than a business line. We covered the last turn of that cycle when DeepSeek raised prices by up to 1,100%.

    The GPU vendors win either way. Open weights that need eight accelerators per node create hardware demand that a closed API never surfaces on anyone else’s balance sheet.

    What are the catches?

    Three, and Tencent names one of them itself.

    The company acknowledges in its release notes that the model spends “longer than necessary reasoning” and over-verifies its own work. On a consumption-priced endpoint, verbosity is a bill. A model that is 82% cheaper per token but writes twice as many tokens is 64% cheaper, not 82%.

    Second, serving is thin. OpenRouter shows a single provider — Tencent Cloud — with 3.16-second P50 latency, 38 tokens per second, and 98.65% availability over three days. There is no failover route.

    Third, the context window is a spec, not a guarantee. The model accepts 1,048,576 input tokens but caps completions at 64,000, and nothing in the release claims uniform recall across the full window. For a like-for-like look at long-context pricing, see our cheapest 1M-context model comparison.

    And it is called “preview” for a reason.

    Frequently asked questions

    Is Tencent Hy4 preview free to use?

    The weights are free under Apache 2.0. The hosted API is not — it costs $0.834 per million input tokens and $2.501 per million output. WorkBuddy and CodeBuddy are free for two weeks from the August 28 launch.

    Can I use Hy4 commercially?

    Yes. Apache 2.0 permits commercial deployment, modification, distillation and redistribution without a separate license negotiation or revenue threshold.

    What hardware do I need to run Hy4?

    Tencent documents a single eight-GPU node using the FP8 quantized build, served through vLLM or SGLang at tensor-parallel size 8. Minimum memory figures are not published.

    How big is the context window?

    1,048,576 tokens of input, with completions capped at 64,000 tokens.

    Is Hy4 better than GPT-5.6 Sol?

    Not on published benchmarks. Hy4 scores 92.3 on GPQA Diamond against Sol’s 94.6, and 85.4 on Terminal-Bench against 88.8. It is cheaper by roughly 8x on output.

    Where can I download the weights?

    Hugging Face at tencent/Hy4-preview, with code and deployment instructions on GitHub.

    Have the benchmarks been independently verified?

    No. All published scores are vendor-reported as of August 29, 2026.

    The bottom line

    Hy4 preview is not a capability breakthrough. It wins its own blind evaluation by 0.05 points and loses 40% of head-to-head comparisons against models that were already open-weight.

    It is a pricing event. Tencent shipped near-parity performance under Apache 2.0 at 82% below Kimi K3’s output rate and roughly one-eighth of GPT-5.6 Sol’s, and put the weights on Hugging Face the same day.

    Tencent’s own README calls it “another step change in capability — the largest generation-over-generation gain we’ve measured.” The blind-evaluation table does not support that framing. The invoice does.

    For anyone running high-volume agentic workloads, Hy4 is worth a benchmark run this week — with token-consumption logging turned on, because the verbosity Tencent admits to is where the savings go to die. For anyone selling inference above $20 per million output tokens, the floor moved again.

    Sources

  • Model Hardware Standard: Anthropic Cuts Lab Setup to 8 Hours

    Anthropic released the Model Hardware Standard on August 27, 2026 — a research preview that lets Claude and rival models drive lab robots, pipettes and factory arms through a single spec. Early testers cut integration from weeks to hours. Carnegie Mellon stood up a serial dilution workflow in 8 hours. QuEra took a laser recovery routine from 58% success to 99.3%. No pricing, no revenue, no open-source date.

    Anthropic has spent two years selling tokens that move text. This one moves matter.

    The company published the Model Hardware Standard, or MHS, as a research preview on August 27. It is a specification, not a product — closer to a plug shape than to a machine. And that is exactly the point.

    What is the Model Hardware Standard?

    The Model Hardware Standard is a shared specification that tells an AI agent what a physical device can do and, more importantly, what it must never do. Vendors ship a driver. The agent reads and writes through simple primitives. Anthropic is running it as an invitation-only research preview.

    Per Anthropic’s own announcement, MHS works with any device that exposes a programmable interface. It is model-agnostic by design — Claude is not required.

    That last detail matters more than the demos. Anthropic is not shipping a robot. It is trying to own the socket every robot plugs into.

    How MHS actually works

    A vendor writes one standardized driver. That driver publishes device discovery in a common format and exposes controls, sensor values and safety limits through a shared memory dictionary.

    The agent then reaches the hardware through one of three paths: the Model Context Protocol, a command line interface, or generated code files. Same device, three levels of abstraction.

    Safety limits live in the driver, not in the prompt. Anthropic gives the example of blocking excess laser power at the device layer — so a confused model cannot talk its way past a hardware ceiling.

    Where MCP ends and the Model Hardware Standard begins

    MCP, which Anthropic debuted in 2024, connects models to software: databases, ticket systems, file stores. If you have followed our coverage of how Agent Skills and MCP split the token bill, the architecture will look familiar.

    MHS extends the same logic to things with motors. Anthropic technical staff member Alek Kemeny put it bluntly to TNW: “What MCP did for software, MHS will do for the hardware world.”

    Kemeny has described MCP elsewhere as “kind of like the USB for AI to software connection.” MHS is the industrial-grade version of that pitch.

    What did the early tests actually prove?

    Six organizations ran MHS against real equipment before launch, and the reported results are specific rather than vague. The headline claim is time: Anthropic says MHS “reduces this integration work to hours or minutes,” against a baseline Genentech described as weeks or months of manual work.

    The most concrete number came from quantum computing. QuEra used MHS to rebuild a laser stabilization routine, moving from 58% success at 150 seconds to 99.3% success at 6 seconds — a 25x speedup on a task that was already mostly failing.

    Organization What was automated Reported result
    QuEra Computing Laser stabilization recovery 58% to 99.3% success; 150s to 6s
    Carnegie Mellon Serial dilution workflow 3x faster; 8-hour setup vs. weeks
    Tetsuwan Scientific qPCR liquid handling 9,143 dispenses across 300 transfer types
    Genentech BCA protein assay tuning Converged at ~140 µL/s (water), 10 µL/s (BSA)
    University of Washington Multi-instrument bench Six instruments connected in under a week
    HHMI Janelia Co-development partner Reference implementation

    Anthropic also says it tested six failure conditions on purpose: missing plate, rotated plate, reader busy, disconnected camera, unreachable device, emergency stop. That is a short list for anything touching a factory floor.

    The number that should give buyers pause

    Every figure above comes from partners Anthropic selected and published. None of it is independently benchmarked, and there is no public failure rate across the full preview cohort.

    A 99.3% success rate on a laser is excellent in a lab. On a production line running 20,000 cycles a shift, it is 140 faults.

    Who is backing the Model Hardware Standard?

    Anthropic named ten hardware vendors and six research institutions at launch. The vendor list is the commercially interesting half, because those are the companies that would have to ship MHS drivers in firmware for the standard to matter.

    Vendors listed by Anthropic as supporting or planning support:

    • Amazon Web Services (Strands Robots library)
    • Universal Robots
    • Doosan Robotics
    • Danaher
    • QIAGEN
    • Tecan
    • Automata
    • MBF Bioscience
    • Hugging Face (LeRobot)
    • Raspberry Pi

    Research users include Genentech, Carnegie Mellon, the University of Washington’s Baker and Pinglay labs, HHMI Janelia, QuEra and Tetsuwan Scientific.

    Jonah Cool, Anthropic’s head of partnerships and deployment of science, told Fortune that lab equipment “suffers from proprietary solutions that are very brittle,” and that the goal is to “avoid vendor lock-in for scientists.”

    Read that again from a vendor’s chair. Anthropic is asking Danaher, QIAGEN and Tecan to help dismantle the integration moat that protects their service revenue.

    How much does the Model Hardware Standard cost?

    Nothing, for now — and that is the strategy. MHS is free during the research preview, gated by an invitation waitlist at modelhardwarestandard.com. Anthropic says it will open-source the framework after the preview, but has published no date, no license and no commercial terms.

    Standards are loss leaders. The money is downstream, in the tokens burned by agents that run instruments around the clock.

    An overnight experiment is a 12-hour inference session. Multiply that by a few thousand labs and the economics start to look like a metered utility rather than a chat subscription.

    Who wins and who loses financially?

    The winners are frontier labs with agent products and the robotics vendors with thin software teams. The losers are instrument makers whose margins depend on proprietary integration, and the systems integrators paid by the week to wire benches together.

    Winners

    Anthropic first. The company was reported at a $2 trillion valuation earlier this month, and a hardware standard extends its distribution into a market where it currently sells nothing.

    Robot arm vendors win cheaply. Universal Robots and Doosan get an agent interface without building an AI stack — the same trade that made Unitree’s IPO pop 629% a bet on hardware plus somebody else’s brains.

    Cloud providers win the runtime. AWS shipped Strands Robots support on day one for a reason.

    Losers

    Integration consultancies are the clearest casualty. If a Carnegie Mellon bench goes from several weeks to 8 hours, that is billable work evaporating.

    Proprietary lab software is next. MarketsandMarkets valued lab automation at $6.60 billion in 2026, growing to $8.62 billion by 2031 at a 6.6% CAGR — a slow market where vendors defend share through lock-in, not growth.

    A commoditized driver layer is precisely the thing that breaks that defense.

    Is the Model Hardware Standard safe enough to run a factory?

    Not yet, and Anthropic says so. The company acknowledged that large language models “still lack physical intuition,” and states that safety evaluations are being built during the preview rather than before it. Human approval workflows exist for high-risk actions, but the physical safety roadmap is unfinished.

    The Register, which covered the launch on August 28, raised the obvious dual-use question: a universal spec for driving instruments does not care what the instrument is for.

    The January 2027 regulatory deadline

    EU Machinery Regulation 2023/1230 takes effect on January 20, 2027. It is the first EU rule to cover AI-based safety functions and self-evolving machine behavior.

    TNW notes the awkward implication: an MHS file that constrains how a machine may operate could itself qualify as a regulated safety component. That would put liability on whoever wrote the driver.

    Anthropic has not said who that is. Five months out from the deadline, this is the unpriced risk in the whole announcement.

    How does this fit Anthropic’s broader agent push?

    MHS is the physical endpoint of a strategy that has been visible all year in software. Anthropic has been widening what an agent can touch, from computer-use agents driving desktops to skills that compress tool definitions.

    The competitive timing is not subtle either. Fortune reported that Hugging Face shipped a robotic duck the same day, and that Nvidia is pursuing a $13 billion acquisition of the company.

    Physical AI is where the capital is rotating. German humanoid maker NEURA Robotics raised up to $1.4 billion in Series C funding this year, per TNW.

    Frequently asked questions

    Is the Model Hardware Standard open source?

    Not yet. Anthropic says it intends to open-source the framework after the research preview, but has published no date or license. Drivers built during the preview are being made available for reuse.

    Does MHS only work with Claude?

    No. Anthropic describes MHS as model-agnostic, meaning OpenAI models and open-weight models can drive MHS devices. Whether rival labs adopt a spec authored by a competitor is a separate question.

    How is MHS different from MCP?

    MCP connects models to software. MHS connects them to physical devices, and adds device-level safety limits, sensor state and discovery. MCP is one of three ways to reach an MHS device, alongside a CLI and generated code.

    Can I use it today?

    Only by invitation. Access runs through a waitlist at modelhardwarestandard.com, and Anthropic has described early access as a “handful” of labs and manufacturers in biotech, robotics and quantum computing.

    What hardware is supported?

    Anything with a programmable interface, in principle. In practice, ten named vendors — including Universal Robots, Danaher, QIAGEN, Tecan and Raspberry Pi — are supporting or planning support. Older instruments without a programmable interface are out of scope.

    What is the biggest risk?

    Regulation and liability. EU Machinery Regulation 2023/1230 applies from January 20, 2027, and MHS constraint files may count as regulated safety components — with no clarity yet on who carries responsibility when an agent-driven machine injures someone.

    The bottom line

    The Model Hardware Standard is the most strategically aggressive thing Anthropic has shipped this year, and it contains no product.

    The engineering claims are credible and unusually specific. A 58% to 99.3% jump on QuEra’s laser routine and an 8-hour Carnegie Mellon integration are not marketing numbers. They are the kind of figures a skeptical buyer can go test.

    But every one of them came from a partner Anthropic chose. There is no pricing, no open-source date, no independent benchmark and no answer on who is liable when a driver written by a language model moves a robot arm into a person.

    The verdict: treat MHS as a distribution land-grab, not a revenue event. If ten vendors becomes fifty by January, Anthropic will own the plug shape for physical AI and collect inference rent on every machine that uses it. If the EU deadline arrives with the liability question still open, the same vendors will quietly wait it out.

    Watch the driver count, not the demos.

    Sources

  • 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

  • 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

  • Claude Watermark Backlash: $100 Users Quit Over Invisible Marks

    Anthropic’s Claude watermark is now live worldwide. Every Claude model launched on or after August 2, 2026 embeds an invisible SynthID-Text mark in generated text — across the API, Claude Code and Claude Cowork. Some $100-a-month Max subscribers are canceling in protest. Anthropic says it has seen no statistically significant uptick, while posting an $11.5 billion quarter.

    The feature was announced on August 11. The technical detail landed on Friday, August 15, in an Anthropic blog post covered by TechCrunch. The cancellation screenshots started over the weekend.

    That sequence matters. Anthropic is meeting investors ahead of a possible fall IPO. It just reported its first quarter of positive adjusted operating income.

    And it chose this moment to stamp a machine-readable signature on everything its models write.

    What is the Claude watermark?

    The Claude watermark is an invisible statistical signature embedded in text the model generates. It nudges Claude toward one of several equally valid word choices, following a secret key. Readers see nothing. A detector holding the key can spot the pattern. Image and file outputs get separate provenance metadata instead.

    How the token steering works

    Anthropic uses the SynthID-Text approach that Google DeepMind published in 2024. The system biases “low-stakes” token choices — preferring “overcast” over “grey” in a sentence about cold weather.

    Where only one answer is correct, the watermark stays off. Gizmodo reported that in a prompt like “Paris is the capital of…” there is no room to steer, so nothing is embedded.

    Anthropic told TechCrunch: “Watermarking does not impact the quality of Claude’s output. To a reader, a watermarked response is indistinguishable from an unwatermarked one.” The company describes the effect on speed and token cost as “negligible.”

    What actually gets marked

    Per Anthropic’s own support documentation, embedded watermarks “apply to all generated text.” Files get signed provenance metadata following the C2PA open standard — the documentation names .svg, .png and .jpg.

    Code is the exception. Functional code demands specific tokens, so there is almost nothing to steer. TechCrunch reported that marking shows up mainly in optional elements such as comments.

    Why did Anthropic ship the Claude watermark now?

    Because a legal deadline landed. Article 50 of the EU AI Act applied from August 2, 2026, and requires providers of generative systems to mark synthetic text, audio, image and video in a machine-readable format. Anthropic’s own docs confirm models launched on or after that date “support marking at launch.”

    Legacy models get a grace period. Under the AI Omnibus agreement, systems already on the market before August 2026 have until December 2, 2026 to meet the machine-readable marking requirement.

    The part that annoyed users is the geography. Search Engine Land reported that “Anthropic said it will enable the watermarking system everywhere Claude is available, not just in Europe.”

    That is a choice, not an obligation. A US freelancer with no EU exposure now carries an EU compliance artifact in their drafts — and Anthropic gets to walk into IPO meetings as the lab that shipped transparency early. Read that alongside our coverage of Anthropic’s $2 trillion IPO positioning and the timing looks less accidental.

    Who is affected by the Claude watermark?

    Effectively everyone who touches Claude. Anthropic states the marking applies “everywhere you use Claude, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag,” and that it also applies when supported models are accessed through AWS, Google Cloud or Microsoft Foundry. There is no documented consumer-versus-enterprise carve-out.

    SurfaceMarked?What it means in practice
    Claude appsYes — textChat output carries the signature
    Claude Platform (API)Yes — textYour product’s output is marked too
    Claude CodePartialMostly comments, not functional code
    Claude CoworkYes — text and filesC2PA metadata on generated images
    AWS / Google Cloud / Microsoft FoundryYesCloud resale does not strip the mark
    Models released before Aug 2, 2026Rolling outRetrofit due by Dec 2, 2026

    Are Claude users really canceling their subscriptions?

    Some are, loudly. Multiple subscribers posted cancellation screenshots on X citing the watermark, including math influencer John Ennis, who called it a “ridiculous watermark idea.” The tier repeatedly named in coverage is Claude Max at $100 per month.

    Anthropic’s counter is blunt. The company told Business Insider it had not yet seen a statistically significant uptick in cancellations since the announcement.

    Both things can be true. A few hundred visible cancellations is a trend on X and a rounding error on the income statement.

    The scale gap is the story. Fortune reported Anthropic’s Q2 2026 revenue at over $11.5 billion, against $787 million in Q2 2025 — at least 14-fold growth — with positive adjusted operating income. Q1 2026 was $4.73 billion.

    Consumer subscriptions are not what produced those numbers. The API is. And API customers building products in Europe want the compliance box ticked far more than they want unmarked prose.

    Can you remove the Claude watermark?

    Partly, and inconsistently. Anthropic says light editing “probably won’t remove the watermark completely,” while “a complete rewrite where every word is replaced will.” Detectability scales with text length, so short outputs may carry no usable signal at all.

    • Full rewrites strip it — every token has to change.
    • Light edits usually do not strip it.
    • Short passages may never accumulate enough signal to detect.
    • Code is barely marked, so engineering workflows are largely unaffected.
    • Passing Claude text through a second model is the obvious laundering route — and nothing in the design prevents it.

    There is a nastier edge case. TechCrunch noted the mark can appear even when Claude is only editing text a human wrote — which is exactly the scenario freelancers and consultants are worried about.

    Who wins and who loses financially?

    Anthropic wins on regulatory positioning ahead of an IPO. Detection and provenance vendors win a mandated market. The losers are the people whose pricing depends on clients believing a human wrote the words — and, potentially, Anthropic’s own consumer tier.

    The winners

    Anthropic first. Fortune reports the company filed confidentially and is working with Morgan Stanley, Goldman Sachs and JPMorgan Chase, with an autumn listing that would beat OpenAI and DeepSeek to market. “Already compliant with Article 50” is a clean line in a prospectus.

    Then the provenance stack. C2PA tooling, detection APIs and audit vendors all get a regulatory tailwind — Anthropic has said it plans to release a watermark detection API of its own.

    Platforms benefit too. Axios reported LinkedIn testing an AI detection button, Substack integrating detection technology and Snap limiting promotion of AI-generated video.

    The losers

    Freelance writers, agencies and consultants who bill human rates. A watermark that survives light editing turns “I used AI to tidy this up” into an evidentiary problem with a client.

    Competitors get a gift. Axios reports OpenAI’s watermarking efforts “primarily focus on images and audio rather than text.” For a writer choosing a $100 subscription this week, that asymmetry is the whole decision — and it lands in a market where the price war has already turned.

    There is also an open-weights escape hatch. Self-hosted models carry no vendor watermark at all, which quietly strengthens the case we made in our breakdown of open weights versus API.

    Is the Claude watermark actually reliable?

    No — and Anthropic says so itself. The company cautions that a detected mark “doesn’t prove Claude created the original ideas or wrote the original text,” and that “the absence of a detectable watermark doesn’t mean content wasn’t AI-generated.” Both directions fail.

    Read that carefully. A positive result does not prove authorship. A negative result does not prove human writing.

    What is left is a signal that Claude touched a document at some point. That is a compliance artifact, not evidence.

    The skeptical read: universities, publishers and HR departments will treat it as evidence anyway, because a machine-readable flag is easier to act on than a judgment call. Anthropic has built a tool it explicitly says cannot answer the question everyone will use it to answer.

    Claude watermark FAQ

    Can I turn the Claude watermark off?

    Anthropic’s published documentation lists no opt-out, and none of the major coverage from August 11–16 identified one. Assume it is on.

    Does it apply outside the European Union?

    Yes. Anthropic is enabling it worldwide, not only in the EU, even though Article 50 is a European obligation.

    Does it slow Claude down or cost more tokens?

    Anthropic describes the impact on speed and token cost as negligible. No independent benchmark of that claim has been published yet.

    Will it flag code I wrote with Claude Code?

    Barely. Functional code leaves almost no room for token steering, so marking concentrates in comments and other optional text.

    Which models are affected?

    Claude models launched on or after August 2, 2026 support marking at launch. Earlier models are being retrofitted during the AI Act transition period, which runs to December 2, 2026.

    Can a client or university prove I used Claude?

    Not conclusively. Anthropic states the mark shows Claude processed the text, not that Claude authored it. Expect that nuance to be ignored in practice.

    Do OpenAI and Google watermark text the same way?

    Not equivalently. Axios reports OpenAI’s watermarking focuses on images and audio rather than text. The underlying SynthID-Text method originated at Google DeepMind in 2024.

    The bottom line

    The Claude watermark is a compliance product shipped as a trust product, and the gap between those two things is where the backlash lives.

    Anthropic is not going to lose an $11.5 billion quarter over cancelled $100 subscriptions. The API business it actually runs on will benefit from being first through the Article 50 door.

    But it has handed competitors a specific, nameable reason to switch — and handed self-hosted open weights a genuine selling point that has nothing to do with price or benchmarks.

    Verdict: right call for the IPO, expensive call for the writers. If your work product depends on nobody being able to run a detector over it, this was the week your vendor choice changed.

    Sources