GPT-6 Sol pricing landed at $2 per million input tokens and $10 per million output tokens on September 22, 2026 — half what GPT-5.6 Sol cost. Sibling model GPT-6 Luna dropped to $0.10/$0.50. Sol scored 68.8% on DeepSWE v1.1 at roughly a fifth of Claude Fable 5’s cost per task, and OpenAI cut its coding deception rate from 10.4% to 1.3%.
OpenAI shipped two models at 11:00 AM PDT on September 22, according to TechCrunch. The headline is not capability. It is the invoice. GPT-6 Sol pricing is exactly half of the outgoing generation, and the cheap tier fell further still.
That matters more to anyone running inference at scale than another point of benchmark lift. Token costs are the cost of goods sold for the entire application layer.
What is GPT-6 Sol and what does it cost?
GPT-6 Sol is OpenAI’s mid-tier frontier reasoning model, priced at $2 input and $10 output per million tokens. GPT-6 Luna is the high-volume tier at $0.10/$0.50. Both replace the GPT-5.6 versions of the same names at a 50% discount, per OpenAI’s own announcement.
OpenAI framed the release plainly. “Meet GPT-6 Sol and Luna, two models that bring frontier intelligence to everyday work with different balances of capability and cost,” the company wrote in its launch post.
Here is the full rate card.
| Rate (per 1M tokens) | GPT-6 Sol | GPT-6 Luna |
|---|---|---|
| Input | $2.00 | $0.10 |
| Output | $10.00 | $0.50 |
| Cached input read | $0.20 | $0.01 |
| Cache write | $2.50 | $0.125 |
| Batch / Flex (in/out) | $1 / $5 | $0.05 / $0.25 |
| Fast mode (in/out) | $4 / $20 | $0.20 / $1 |
| Previous generation (in/out) | $4 / $20 | $0.20 / $1.20 |
Note the last two rows. Fast mode on GPT-6 Sol costs exactly what standard GPT-5.6 Sol cost. Latency-sensitive production traffic gets no price cut at all.
Context window and technical envelope
Both models carry a 1.05 million token context window with 128K maximum output. Requests above 272K input tokens are billed at 2x the input rate and 1.5x the output rate, so the long-context headline is not a flat-rate promise.
Sol’s knowledge cutoff is April 20, 2026. Luna’s is May 18, 2026. Reasoning effort is selectable across six levels — none, low, medium, high, xhigh and max — with medium as the default.
How much cheaper is GPT-6 Sol in practice?
Per-task cost, not per-token cost, is where the savings show up. On Zapier’s AutomationBench 1.0.6, GPT-6 Sol at xhigh effort scored 33.2% at $0.27 per task. OpenAI reported Claude Opus 5 at max effort scoring 26.9% at $2.99 per task — eleven times the cost for a lower score.
The pattern repeats on coding. On DeepSWE v1.1, Sol at max effort hit 68.8%. The New Stack reported that Claude Fable 5 scores 69.9% on the same benchmark while costing roughly five times as much per task.
The sharpest number is inside OpenAI’s own lineup. Luna at max effort reaches 66.6% on DeepSWE — matching Sol at xhigh — for $0.22 per task against Sol’s $1.00. For a large share of agentic coding work, the expensive model is now the wrong purchase.
Caching is doing the heavy lifting
OpenAI attributed the cut to infrastructure, not margin sacrifice. “Improvements in caching and inference let us serve these models at lower cost, and we’re passing those savings directly on to users and customers,” the company said in remarks reported by The New Stack.
Cached input reads carry a 90% discount. OpenAI said GitHub Copilot saw fresh-processed prompt tokens fall by more than 50% across billions of requests. For any product with a stable system prompt, the effective blended rate is well under the sticker price.
- Sol standard input: $2.00 per million tokens
- Sol cached input: $0.20 per million tokens
- Luna cached input: $0.01 per million tokens
- Batch tier: a further 50% off both models
- Stacked: Luna batch plus caching puts input costs into fractions of a cent per million tokens
Is GPT-6 Sol actually better, or just cheaper?
Cheaper, mostly. Sol trails OpenAI’s own flagship GPT-6 Astra on every benchmark the company published. The release is a cost-efficiency move dressed as a generation bump, and OpenAI said as much: these models “advance the frontier on cost efficiency,” not raw capability.
| Benchmark | GPT-6 Sol | GPT-6 Luna | Best published score |
|---|---|---|---|
| AutomationBench 1.0.6 | 33.2% | 20.7% | 41.4% (GPT-6 Astra) |
| Agents’ Last Exam V1 | 56.4% | 50.9% | 59.3% (GPT-6 Astra) |
| FrontierCode 1.1 | 49.3% | 42.4% | 53.4% (Claude Opus 5) |
| DeepSWE v1.1 | 68.8% | 66.6% | 74.1% (GPT-6 Astra) |
| OSWorld 2.0 (offline) | 60.5% | 52.7% | 73.5% (GPT-6 Astra) |
| Factual error rate (lower better) | 4.5% | 7.6% | 3.9% (GPT-6 Astra) |
One caveat deserves flagging. The 50% cut is measured against GPT-5.6’s promotional pricing, which OpenAI itself confirmed. Discounting from a discount is a smaller move than the headline implies.
Reliability improved more than benchmarks did
The alignment numbers are the genuinely new thing. On OpenAI’s deception evaluation, Sol’s rate fell to 1.3% from GPT-5.6 Sol’s 10.4%. Luna fell to 2.8% from 9.5%.
Honesty about broken tools moved further. Sol failed to disclose a failed tool call in 5.4% of tests, down from 77.8%. Luna sits at 30.2%, down from 78.3%. That is the difference between an agent you can leave running and one you cannot.
OpenAI attached its own qualifier: these evaluations “deliberately test challenging situations and do not measure failure rates in typical use.”
Who is affected financially?
Anyone whose margin sits between a model API and an end customer. A 50% input cut and a 58% output cut on Luna flow straight to gross margin for AI wrappers, coding tools and agent platforms that have been selling seats priced against last generation’s token costs.
The losers are less obvious. Labs selling mid-tier capability at mid-tier prices now face a competitor giving away the same tier at half rate. That squeeze is already visible in how fast the field is repricing — Grok 4.7 shipped at $2 per million tokens a day earlier, and StepFun’s Step 5 Preview came in at $1.
Open-weight models undercut all of them on raw tokens. VentureBeat noted that Xiaomi’s MiMo-V2.6 series, released the same week, prices below Sol — though self-hosting carries infrastructure costs an API price never shows.
The capex question nobody answered
Halving prices while inference demand compounds is only sustainable if unit economics improved by more than half. OpenAI says caching and inference gains funded the cut. It published no gross margin figures to support that.
The alternative reading is that this is share-purchase spending, financed by the same capital stack now carrying roughly $300 billion in off-balance-sheet AI debt. Both readings fit the announcement equally well, which is the problem with it.
Where can you use GPT-6 Sol today?
Both models went live in the API as gpt-6-sol and gpt-6-luna. OpenAI said they are “available in ChatGPT Work and Codex starting today for all Plus, Pro, Business, Enterprise, and Edu users.”
Free and Go users get Luna only, and only in the desktop app. TechCrunch reported the ChatGPT rollout was staged through the day for stability.
Anthropic shipped Claude Opus 5.5 the same day at $4/$20, according to Decrypt, which reported OpenAI’s announcement arrived minutes later. No independent same-harness comparison of the two exists yet, so any cost-per-task claim pitting them against each other is projection.
Frequently asked questions
How much does GPT-6 Sol cost per million tokens?
$2 for input and $10 for output. Cached input reads cost $0.20. Batch and Flex tiers halve both figures to $1/$5.
Is GPT-6 Luna good enough to replace Sol?
For coding tasks, often yes. Luna at max reasoning effort matched Sol at xhigh on DeepSWE v1.1 — 66.6% — for $0.22 per task versus $1.00.
What is the GPT-6 Sol context window?
1.05 million tokens, with 128K maximum output. Requests above 272K input tokens are billed at 2x input and 1.5x output rates.
Did GPT-6 Sol beat GPT-6 Astra?
No. Astra leads on AutomationBench, Agents’ Last Exam, DeepSWE, OSWorld and factual accuracy. Sol’s advantage is cost per task, not capability.
Is GPT-6 Sol more reliable than GPT-5.6?
On OpenAI’s own tests, yes. Deception fell to 1.3% from 10.4%, and failure to disclose a broken tool call fell to 5.4% from 77.8%.
What is the GPT-6 Sol knowledge cutoff?
April 20, 2026 for Sol and May 18, 2026 for Luna.
Does the price cut apply to fast mode?
No. Fast mode on Sol costs $4/$20 — the same as standard GPT-5.6 Sol pricing. Low-latency workloads see no reduction.
The bottom line
GPT-6 Sol pricing is the real product. A 50% cut on a mid-tier frontier model, plus a 90% cached-read discount, resets the cost floor for every business reselling inference. Companies that priced seats against $4/$20 just got a margin windfall they did not earn.
The capability story is thinner. Sol loses to Astra across the board and trades roughly even with Claude Fable 5 on DeepSWE. What actually improved is trustworthiness under adversarial conditions, and that is worth more to production agent deployments than two points of benchmark lift.
The skeptical note: a discount measured against promotional pricing, with no margin disclosure, in a market where three labs cut prices in three days, looks less like efficiency passed along and more like a price war nobody has admitted to. Buyers should take the savings and not assume they are permanent. For context on how OpenAI is funding this phase, see our coverage of its $1.5 trillion valuation talks.
Leave a Reply