GPT-6 Sol
GPT-6 Sol is the middle tier of the GPT-6 family, released by OpenAI on Sep 22, 2026, together with Luna. It costs US$ 2 per million input tokens and US$ 10 per million output, half of GPT-5.6 Sol's promotional price, and matches Claude Sonnet 5.5's rate card to the cent up to 272 thousand tokens. A week later OpenAI released GPT-6.1 Sol at the same price with higher scores; 6 Sol still makes sense for integrations that depend on none effort or on tools through Chat Completions.
GPT-6 Sol in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- OpenAI
- API identifier
gpt-6-sol- Release
- Sep 22, 2026, together with GPT-6 Luna
- Input
- US$ 2 per million tokens
- Output
- US$ 10 per million tokens
- Cache read
- US$ 0.20 per million (10% of input)
- Cache write
- US$ 2.50 per million (1.25x input)
- Above 272 thousand input tokens
- US$ 4 input, US$ 0.40 cache and US$ 15 output, for the whole request
- Batch and Flex
- US$ 1 input and US$ 5 output
- Fast mode
- US$ 4 and US$ 20; no Ultrafast
- Context window
- 1,050,000 tokens (up to 922 thousand input)
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Apr 20, 2026
- Reasoning
- effort
nonetomax, defaultmedium - Tools on Chat Completions
- only with
reasoning_effortset tonone; with reasoning, through the Responses API - Where it runs
- OpenAI API, ChatGPT Work, Codex and Amazon Bedrock (us-east-1)
- Cyber risk
- High on the Preparedness scale, with the GPT-5.6 safeguard package
Where GPT-6 Sol does well, where it falls behind and who it is for
Where it does well
- Price: US$ 2 and US$ 10 per million tokens, 50% below GPT-5.6 Sol's promotional price (US$ 4 and US$ 20).
- Cheap automation: 33.2% on AutomationBench at xhigh effort, for US$ 0.27 per task, above Claude Opus 5 at max (26.9%) at 9% of the cost, in OpenAI's figures.
- Agents: 56.4% on Agents' Last Exam at max effort, above Opus 5's best score, at a 60% lower cost per task, according to OpenAI.
- Fewer factual errors: about half the mistakes of its predecessor, per OpenAI. At medium effort, 6.9% of answers with errors on hard prompts, against 15.0% for GPT-5.6 Sol.
- Flexible API: accepts
noneeffort and tool calls through Chat Completions in that mode, which 6.1 Sol and Astra do not.
Where it falls behind
- It already has a successor: GPT-6.1 Sol costs the same for input and output, rereads the cache at half the price and beats 6 Sol on all six performance tests OpenAI published with it.
- Science in the terminal: 27.6% on Terminal-Bench Science at max effort, against 57.0% for 6.1 Sol and 68.1% for Astra.
- Against its own predecessor: at max effort, 68.8% on DeepSWE and 64.4% on OSWorld 2.0 offline, below GPT-5.6 Sol's 72.7% and 66.2%, though at much lower cost per task.
- Images: until Sep 25, 2026, an encoding bug degraded image understanding. OpenAI fixed it and recommends rerunning evaluations that use images.
- Health: in the safety report, 53.2 on HealthBench and 30.1 on HealthBench Hard, below GPT-5.6 Sol's 57.0 and 33.1.
Who it makes sense for
- Integrations that call tools through Chat Completions without reasoning, at
noneeffort, and do not want to move to the Responses API yet. - High-volume workflow automation, where cost per task matters more than a few points of score.
- Teams comparing against Claude Sonnet 5.5: the price per token is the same, so a test with your own tasks settles it.
- Anyone still on GPT-5.5 in Codex: OpenAI points to GPT-6 Sol as the replacement, and GPT-5.5 leaves ChatGPT and Codex on Oct 14, 2026.
When to pick another model
- New projects that do not depend on
none: start with GPT-6.1 Sol, same price and higher scores. - Scientific research and long terminal work: 6.1 Sol scores twice as much on Terminal-Bench Science, and Astra more still.
- Requests that often go past 272 thousand tokens: Claude Sonnet 5.5 keeps US$ 2 and US$ 10 up to 1 million.
- High-volume classification, extraction and summaries: GPT-6 Luna costs one twentieth per token.
What OpenAI published at launch
GPT-6 Sol came out on Sep 22, 2026 with Luna, trained with methods similar to Astra's. OpenAI halved the price relative to GPT-5.6 Sol's promotional rate and changed its pitch: instead of claiming the best model, it compared cost per task against Claude models.
In the table, GPT-6 Sol does not lead any row, but it is the cheapest per task in every one. Against Opus 5, OpenAI highlights a similar or higher score at a fraction of the cost. Against its own predecessor, the gain is uneven: on DeepSWE and OSWorld, GPT-5.6 Sol at max effort scores higher, while spending more than twice as much.
| GPT-6 Sol OpenAI | GPT-5.6 Sol OpenAI | Claude Opus 5 Anthropic | GPT-6 Astra OpenAI | |
|---|---|---|---|---|
| AutomationBench automation |
33.2% (US$ 0.27) | 28.8% (US$ 0.67) | 26.9% (US$ 3.05) | 41.4% (US$ 1.73) |
| Agents' Last Exam agent tasks |
56.4% (US$ 2.93) | 53.6% (US$ 5.08) | 55.9% (US$ 7.29) | 59.3% (US$ 6.23) |
| DeepSWE v1.1 software engineering |
68.8% (US$ 2.74) | 72.7% (US$ 6.46) | 73.7% (US$ 11.84) | 74.1% (US$ 4.43) |
| FrontierCode 1.1 Main coding |
49.3% (US$ 2.14) | 47.5% (US$ 5.19) | 53.4% (US$ 4.31) | 53.3% (US$ 4.59) |
| OSWorld 2.0 offline computer use |
64.4% (US$ 3.25) | 66.2% (US$ 7.71) | 70.2% (US$ 24.11) | 73.5% (US$ 9.07) |
| Factual errors hard prompts, lower is better |
4.5% (US$ 0.13) | 8.4% (US$ 0.39) | not measured | 3.9% (US$ 0.48) |
Charts from the GPT-6 Sol and Luna announcement. Each model's best score at any effort level; in brackets, the average cost per task. The dot marks the best in each row. Competitor scores come from public reports, according to OpenAI. Some scores here differ from the Astra announcement, such as GPT-5.6 Sol on AutomationBench (28.8% against 18.1%) and Astra on OSWorld (73.5% against 72.6%).
Same price as Claude Sonnet 5.5, up to 272 thousand tokens
GPT-6 Sol and Claude Sonnet 5.5 share the same rate card to the cent: US$ 2 input, US$ 10 output, US$ 0.20 cache read and US$ 2.50 cache write. With price tied, the choice moves from cost per token to cost per task, which depends on how many tokens each model spends to solve the same problem.
The difference shows up on large requests. Above 272 thousand input tokens, OpenAI doubles input and cache and charges 1.5x output for the whole request. Anthropic charges the same rates up to 1 million tokens. For apps that send whole documents with every request, Sonnet 5.5 is cheaper per token in that range.
| GPT-6 Sol OpenAI | Sonnet 5.5 Anthropic | GPT-6.1 Sol OpenAI | |
|---|---|---|---|
| Input | US$ 2 | US$ 2 | US$ 2 |
| Output | US$ 10 | US$ 10 | US$ 10 |
| Cache read | US$ 0.20 | US$ 0.20 | US$ 0.10 |
| Cache write | US$ 2.50 | US$ 2.50 | US$ 2.50 |
| Input above 272 thousand tokens | US$ 4 | US$ 2 | US$ 4 |
| Output above 272 thousand tokens | US$ 15 | US$ 10 | US$ 15 |
| Context window | 1.05 million | 1 million | 1.05 million |
OpenAI and Anthropic pricing pages, checked on Oct 1, 2026. For Anthropic, 5-minute cache writes.
Astra, 6.1 Sol, Sol or Luna: choosing within OpenAI
GPT-6 Sol was born as the middle tier of the family, below Astra and above Luna. A week later, on Sep 29, OpenAI released GPT-6.1 Sol at the same input and output price, with half-price cache and higher scores. Since then, OpenAI's model guide lists only three models: Astra, 6.1 Sol and Luna.
In practice: for new projects, 6.1 Sol. For simple high-volume tasks, Luna, at one twentieth of the price per token. For the most demanding work, Astra, at five times. GPT-6 Sol remains for teams that need none effort or tools through Chat Completions, which 6.1 Sol does not accept.
| GPT-6 Astra | GPT-6.1 Sol | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|---|
| Input | US$ 10 | US$ 2 | US$ 2 | US$ 0.10 |
| Output | US$ 50 | US$ 10 | US$ 10 | US$ 0.50 |
| Cache read | US$ 1 | US$ 0.10 | US$ 0.20 | US$ 0.01 |
| Cache write | US$ 12.50 | US$ 2.50 | US$ 2.50 | US$ 0.125 |
| Above 272 thousand tokens input and output |
US$ 20 and US$ 75 | US$ 4 and US$ 15 | US$ 4 and US$ 15 | US$ 0.20 and US$ 0.75 |
| Context window | 1.05 million | 1.05 million | 1.05 million | 1.05 million |
| Knowledge cutoff | Apr 30, 2026 | Apr 30, 2026 | Apr 20, 2026 | May 18, 2026 |
| Accepts effort none | no | no | yes | yes |
| Role, per OpenAI | highest intelligence | balance of cost and intelligence | superseded by 6.1 Sol | fastest and most cost-effective |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026.
Examples of using GPT-6 Sol
An assistant checks orders and stock through function calls, without long reasoning, on Chat Completions at none effort. GPT-6 Sol accepts this setup; with 6.1 Sol, you would have to move to the Responses API.
An agent reads emails, updates the CRM and triggers tasks across several systems. On AutomationBench, GPT-6 Sol at xhigh effort scored 33.2% for US$ 0.27 per task, in OpenAI's figures.
With the same price per token, it is worth running the same ten real tasks on both and comparing accuracy and cost per task. Keep requests under 272 thousand tokens so the price comparison stays fair.
Where GPT-6 Sol stands out and where it falls behind, by the numbers
GPT-6 Sol has published results on 26 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 26, 2025 and Dec 21, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
It is not in the bottom half of its generation on any test with 6 or more models compared.
See all 26 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 94.3% | #9 / 73 | #9 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 75 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| Dtbench | 97.3% | #5 / 64 (tied with 2) | #5 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 59.08 | #7 / 64 | #7 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 30.9% | #7 / 76 | #7 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 57.6% | #17 / 74 | #17 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 89.8% | #5 / 64 | #5 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 89.6% | #7 / 43 | #7 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 2,462 | #3 / 57 | #3 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,689 | #6 / 67 | #6 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 95.5% | #12 / 43 (tied with 1) | #12 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 60.7% | #14 / 59 | #15 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 56.0% | #12 / 58 | #13 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 56.0% | #9 / 58 (tied with 1) | #9 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 83.0% | #12 / 64 (tied with 1) | #12 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 90.0% | #5 / 54 | #5 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 14,428 | #2 / 43 | #2 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 54.3% | #22 / 48 | #22 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 49.3% | #9 / 43 | #9 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 26.4% | #11 / 36 | #11 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 58.3% | #8 / 32 | #8 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 36.9% | #7 / 29 | #7 / 31 | 54.4% Gemini 4 Argon |
| Ebr Bench | 53.3% | #6 / 22 | #6 / 27 | 76.2% GPT-6 Astra |
| Spireclimb | 7.5% | #3 / 7 | #3 / 7 | 25.7% GPT-6 Astra |
| Spireclimb Sts1 A20 Heart | 1.7% | #3 / 7 | #3 / 7 | 20.0% GPT-6.1 Sol |
| Spireclimb Sts2 A10 | 13.3% | #3 / 7 | #3 / 7 | 34.7% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare GPT-6 Sol with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Gemini 4 Argon | Google DeepMind | 164.56 | 1,608 | n/a | Sep 30, 2026 |
| Claude Opus 5 | Anthropic | 162.60 | 1,573 | 80.1 | Jul 24, 2026 |
| GPT-6 Sol | OpenAI | 162.55 | n/a | 79.3 | Sep 22, 2026 |
| Claude Fable 5 | Anthropic | 162.13 | 1,542 | 83.0 | Jun 9, 2026 |
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
GPT-6 Sol or Claude Opus 5?
Claude Opus 5 is the model from another lab closest to GPT-6 Sol in capability. Across the 22 tests both took, this is the score:
| Test | GPT-6 Sol | Claude Opus 5 |
|---|---|---|
| Chess puzzles | 56.0% | 43.0% |
| Vending Bench 2 | 14,428 | 11,182 |
| Frontiermath Tier 4 V2 | 90.0% | 73.2% |
| Blueprint Bench 2 | 36.9% | 30.4% |
| Apex Agents | 54.3% | 65.8% |
| ProofBench | 83.0% | 99.0% |
| Ebr Bench | 53.3% | 45.7% |
| Gdp Pdf | 26.4% | 24.0% |
| FrontierCode | 49.3% | 53.4% |
| Lmca | 59.08 | 63.30 |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
Other models from OpenAI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| GPT-6.1 Sol | Sep 29, 2026 | 166.32 | n/a | 81.6 | API access |
| GPT-6 Sol | Sep 22, 2026 | 162.55 | n/a | 79.3 | API access |
| GPT-6 Luna | Sep 22, 2026 | 156.31 | n/a | 72.0 | API access |
| GPT-6 Astra | Sep 3, 2026 | 166.41 | n/a | 82.2 | API access |
| GPT-5.6 Sol | Jul 9, 2026 | 161.73 | 1,527 | 81.1 | API access |
| GPT-5.6 Terra | Jul 9, 2026 | 159.43 | 1,515 | 77.9 | API access |
| GPT-5.6 Luna | Jul 9, 2026 | 156.33 | 1,478 | 73.6 | API access |
| GPT-5.5 Instant | May 5, 2026 | 142.45 | 1,456 | n/a | API access |
| GPT-5.5 Pro | Apr 23, 2026 | 161.97 | n/a | n/a | API access |
Have you used GPT-6 Sol? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about GPT-6 Sol
How much does GPT-6 Sol cost?
US$ 2 per million input tokens and US$ 10 per million output. Cache reads cost US$ 0.20 and cache writes US$ 2.50. Above 272 thousand input tokens, the whole request moves to US$ 4 input and US$ 15 output. Batch and Flex are half price, and Fast mode costs double.
GPT-6 Sol or GPT-6.1 Sol?
For new projects, 6.1 Sol: same input and output price, half-price cache and a better result on all six performance tests OpenAI published, such as 57.0% against 27.6% on Terminal-Bench Science. GPT-6 Sol only has the edge if you need none effort or tools through Chat Completions.
Does GPT-6 Sol cost the same as Claude Sonnet 5.5?
Yes, up to 272 thousand input tokens: US$ 2 input, US$ 10 output, US$ 0.20 cache read and US$ 2.50 cache write on both. Above that, OpenAI doubles input and cache and charges 1.5x output, while Anthropic keeps its rates up to 1 million. The independent scores are at the top of this page.
Is GPT-6 Sol in ChatGPT?
At launch it came to ChatGPT Work and Codex on the Plus, Pro, Business, Enterprise and Edu plans, but not to Chat. In Codex, it is the recommended replacement for GPT-5.5, which leaves on Oct 14, 2026.
Where does GPT-6 Sol rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-6 Sol ranks #8 out of 267 models, with an ECI of 162.55. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is GPT-6 Sol better than Claude Opus 5?
Across the 22 tests both models took and Epoch AI compiled, GPT-6 Sol comes out ahead on 13, behind on 9 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is GPT-6 Sol best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of GPT-6 Sol are: Vending Bench 2 (#2 of 43); ALE-Bench (#3 of 57); Frontiermath Tiers 1 3 V2 (#5 of 64). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- OpenAI, GPT-6 Sol and GPT-6 Luna announcement
- OpenAI, GPT-6 Sol model page (documentation)
- OpenAI, API pricing
- OpenAI, GPT-6 Astra system card (GPT-6 Sol and GPT-6 Luna appendix)
- OpenAI, API changelog (image fix on Sep 25, 2026)
- OpenAI, GPT-6.1 Sol announcement (charts including GPT-6 Sol)
- OpenAI, latest model guide
- OpenAI, models in Codex and ChatGPT Work
- OpenAI, models on Amazon Bedrock (documentation)
- Anthropic, pricing (documentation)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking