Think of a good law firm. The senior partner is brilliant and bills accordingly. You’d never ask them to go through two hundred pages of supplier contracts looking for the cancellation clause. You give that to a trusted senior associate, someone almost as sharp and a lot cheaper.
GPT-6.1 Sol is that senior associate. It isn’t OpenAI’s most powerful model, but it is close, and it’s the one you can afford to use for most of the week’s work.
What it is, and where you’ll find it
OpenAI released GPT-6.1 Sol on September 29, 2026, at its DevDay event, one week after GPT-6 Sol. OpenAI describes it as near-Astra intelligence for coding, computer use, and professional work, at a fifth of the price of its flagship GPT-6 Astra.
Where you can use it matters for most readers. In ChatGPT, GPT-6.1 Sol lives in ChatGPT Work and Codex, the parts of the app built for longer tasks, files, and agents. It’s available on the Plus, Pro, Business, Enterprise, and Edu plans. It is not yet in regular Chat, and it isn’t on the free plan. Developers can use it through the API as gpt-6.1-sol, and coding tools such as Cognition’s Devin added it on launch day.
How close to the flagship?
Artificial Analysis, an independent testing firm, ran its benchmarks on launch day. Its Intelligence Index combines ten hard evaluations, from office work and business automation to scientific coding and factual knowledge. GPT-6.1 Sol scored 52. GPT-6 Astra scored 53, and last week’s GPT-6 Sol scored 48. That puts GPT-6.1 Sol 10th of the 222 models on the board.
Now look at the price. Artificial Analysis measured about $0.72 to run one task of its index with GPT-6.1 Sol, against $3.26 for Astra. That’s one point less capability for less than a quarter of the cost.
For fairness, Anthropic still leads this index. Claude Opus 5.5 scored 58 and Claude Sonnet 5.5 scored 56. But those scores cost much more to reach: $5.98 and $7.60 per task.
What it’s good at in a working day
OpenAI’s launch tests focus on the kind of work that fills an office week. Keep in mind that these are OpenAI’s own numbers.
- Reading difficult PDFs. On GDP.pdf, the model answers questions about real documents from finance, law, healthcare, and seven other fields, with dense tables, charts, and fine print. OpenAI’s chart shows 32.0% at high effort for $0.35 per task. Claude Opus 5.5’s best is 28.8% at $0.83, and GPT-6 Astra’s best is 32.2% at $1.91, so Sol nearly matches the flagship for a fifth of the price.
- Multi-step chores. On AutomationBench, an agent completes end-to-end workflows across 47 business tools, such as updating a CRM, filing a ticket, and sending a follow-up. At medium effort, GPT-6.1 Sol scores 31.7% for $0.19 per task, 2.2 points above Claude Opus 5.5 at the same setting. Read the whole chart, though: at max effort Opus 5.5 reaches 42.5% and Astra 41.4%, while Sol tops out at 36.1%. Sol wins on routine chores per dollar; the pricier models still finish more of the hardest ones.
- Using a computer. On OSWorld 2.0, the model clicks and types its way through real desktop apps. OpenAI reports 71.4% at $1.27 per task, against 73.5% for Astra at $9.44.
Fewer made-up facts, but not none
OpenAI tested factual accuracy on deliberately hard questions taken from real conversations where users had flagged an earlier model’s mistake. At low effort, the share of answers containing a factual error fell from 11.4% to 7.7% compared with GPT-6 Sol, roughly a third fewer.
Artificial Analysis’s independent knowledge test gives a more sober picture. When GPT-6.1 Sol doesn’t know an answer, it still makes one up 54% of the time. That’s better than GPT-6 Sol’s 60%, but worse than Claude Sonnet 5.5’s 47%. The practical advice doesn’t change: for an obscure name, date, or figure, ask it to search the web or cite a source.
A word on the Astra that didn’t ship
GPT-6.1 Sol was not the only model OpenAI had planned for DevDay. The day before, the company cancelled GPT-6.1 Astra after internal tests found it didn’t reliably stay within the scope and permissions of its tasks. GPT-6.1 Sol is a separate model with its own safety report. In one of OpenAI’s tests, an agent’s search tool is quietly broken, and the model should say so rather than guess. GPT-6.1 Sol failed to disclose the problem in 2.1% of cases, against 4.9% for GPT-6 Sol and 1.5% for GPT-6 Astra. That’s a real improvement, though not a guarantee for every situation.
The honest catch
It’s behind a paywall and a tab. For a model aimed at everyday work, the biggest limitation is simply access: no free plan and no regular Chat, only ChatGPT Work and Codex on paid plans.
It reads images but doesn’t draw. GPT-6.1 Sol takes text and images as input and writes text back. It doesn’t generate images, audio, or video itself.
Claude is still ahead on the broad index. If you want the highest independent scores regardless of cost, Claude Opus 5.5 and Sonnet 5.5 are ahead. GPT-6.1 Sol’s advantage is doing nearly as well for much less.
Who should use it
| If your day looks like… | Reach for | Why |
|---|---|---|
| Contracts, reports, and PDFs in ChatGPT Work | GPT-6.1 Sol | Near-Astra document work at a fraction of the cost |
| The hardest research and computer-use jobs | GPT-6 Astra | OpenAI’s strongest model, a point or two ahead |
| High-stakes judgment, obscure facts | Claude Opus 5.5 | Highest on the broad independent index |
| Everyday work on a free plan | Claude Sonnet 5.5 or Gemini | Strong models without a subscription |
The everyday verdict
GPT-6.1 Sol won’t impress you by being the smartest model on the board. It impresses you by being almost as smart as OpenAI’s flagship at a fraction of the cost. If you already pay for ChatGPT and spend your day in Work, making it your default is the easy choice. Check the obscure facts, and switch to Astra only when a task truly needs the last couple of points.
