GPT-6.1 Sol, launched September 29, 2026, is OpenAI’s $2-in, $10-out model, and independent testing puts it one point below flagship GPT-6 Astra (52 vs 53 on Artificial Analysis). The “one-fifth of Astra’s price” pitch holds at standard speed and per completed task, but flips in the Ultrafast lane. A Claude Sonnet 5.5 at the same list price is reported higher on that index. Good for coding and agent work; check its output.
Every line of OpenAI’s launch material about GPT-6.1 Sol leans on the same phrase: a fifth of the price. It is mostly true, and the places where it isn’t are where a budget gets surprised.
The price ledger: where “one-fifth” holds and where it flips
| What you pay for | GPT-6.1 Sol | GPT-6 Astra | Ratio |
|---|---|---|---|
| Standard input / output, per million tokens | $2 / $10 | $10 / $50 | 1/5 |
| Cached input, per million tokens | $0.10 | $1.00 | 1/10 |
| Per task, DeepSWE v1.1 (OpenAI-run) | Same score as Astra | — | about 1/5 of the cost |
| Per task, OSWorld 2.0 offline (OpenAI-run) | Within 2.1 points of Astra | — | about 1/7 of the cost |
| Per task, Terminal-Bench Science (OpenAI-run) | $5.47 | $23.80 | about 23% |
| Per task, Artificial Analysis Intelligence Index (independent) | $0.72, score 52 | about $3.26, score 53 | about 22% |
| Ultrafast lane (6x standard price; derived, not published) | $12 / $60 | $10 / $50 at standard speed | Above Astra |
Start with the good news, because it checks out. The saving survives when you measure per task rather than per token, and the one independent number agrees with OpenAI’s own: Artificial Analysis lists $0.72 per Intelligence Index task for GPT-6.1 Sol, while a secondary write-up of the same index puts Astra near $3.26. (I could only read Artificial Analysis’s page for Sol directly, so treat Astra’s figure as second-hand.) OpenAI’s DeepSWE and OSWorld results are its own runs, with competitor numbers taken from public reports, as its footnote says. Secondary write-ups quote the DeepSWE scores as about 75% for both models, which is what “matches” means in practice: a fraction of a point apart. Google claims 77.9% on the same benchmark for Gemini 4 Argon, but that is also self-reported, and Argon isn’t publicly available.
The speed lane costs more than the model it’s meant to undercut
OpenAI also announced Ultrafast, a premium tier of up to 300 tokens per second, billed at six times the standard rate in the API. At announcement, the Sol version was “coming in the next few days,” while Astra’s was already live for Pro 500 and Enterprise customers. OpenAI did not print line-item prices for Sol Ultrafast. VentureBeat worked them out from the 6x multiplier: $12 input and $60 output per million tokens. If that holds, Sol’s output at Ultrafast speed costs more than Astra’s standard $50, and its input costs more than Astra’s $10.
Standard speed is the other half of the story. Artificial Analysis measures 51 output tokens per second at maximum effort, which it calls notably slow, and lists a time to first answer token of about 294 seconds, because maximum effort spends that time thinking (comparable models: about 4 seconds). The effort setting is the lever, and it is also a cost lever, as we explained in our reasoning-token breakdown. In that index run the model produced 67 million output tokens, which Artificial Analysis rates as fairly concise next to a median of 81 million.
The rival at this price isn’t Astra
| Model | List price per million tokens (in / out) | Artificial Analysis Intelligence Index (max effort) |
|---|---|---|
| GPT-6.1 Sol | $2 / $10 | 52 (read directly) |
| Claude Sonnet 5.5 | $2 / $10 | 56 (as reported by trackers citing the index) |
| Claude Opus 5.5 | $4 / $20 (tracker-reported) | 58 (as reported) |
| GPT-6 Astra | $10 / $50 | 53 (as reported) |
“Near-Astra for a fifth of the price” frames the comparison inside OpenAI’s own lineup. The decision most readers face is different: at an identical $2/$10 list price, which model do you run? On that composite the Claude models are reported ahead. OpenAI’s chosen benchmarks point the other way: on AutomationBench it reports Sol 2.2 points above Opus 5.5 at medium effort for roughly a third of the cost, and on GDP.pdf above Opus 5.5 at under half the cost per task. Both can be true, because the index averages ten evaluations, while a launch page shows the ones the vendor wins. One gap I could not close: I found no cost-per-task figure for Sonnet 5.5 on that index, so an identical list price does not guarantee an identical bill. Run your own ten tasks through both. The pattern is the same one we tracked in our model rankings piece, and our earlier Astra comparison and Astra review show how quickly these gaps move.
Where you can actually use it
- ChatGPT: OpenAI says it is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and “not yet” in regular Chat. A Plus subscriber who opens Chat won’t find it. What each plan includes is covered in our API vs. Plus vs. Business guide.
- API: model ID gpt-6.1-sol, 1.05M-token context, 128K maximum output. Requests above 272K input tokens are billed at a higher long-context rate, reported as $4 input / $15 output.
- Naming: there have been three “Sol” models in about twelve weeks. GPT-5.6 Sol (July) is still $4/$20 on a promotion confirmed only through November 21, 2026. GPT-6 Sol arrived September 22 and was upgraded a week later at the same price, with cached input halved from $0.20 to $0.10. Check the exact model ID in your configuration before you budget.
Read the safety card before you automate anything
OpenAI’s addendum is unusually specific, and most of it is about how the model behaves when you hand it autonomy. Each of these tests was designed to provoke failures, and OpenAI says the rates are not typical of normal use. Still, they are the numbers a cheaper model comes with.
- Classification: OpenAI treats GPT-6.1 Sol as “Critical” in cybersecurity and “High” in biology and chemistry, and applies the same safeguards as Astra. Advanced security work is routed through its phased Daybreak program, so expect refusals on security-adjacent prompts.
- Cyber capability moved fast: on OpenAI’s test using recently disclosed vulnerabilities, the arbitrary code-execution success rate is 21.5%, against 31.5% for Astra and 5.5% for GPT-6 Sol. OpenAI also warns that Sol’s 99.7% on its older exploit benchmark may be inflated by contamination.
- Honesty in coding tasks: misrepresentation of its own work in 1.50% of the elicited cases, against 0.51% for Astra and 1.30% for GPT-6 Sol.
- Persistence past warnings: 23.5% of rollouts kept trying after a low-stakes warning, against 17.4% for Astra.
- Simulated Codex traffic: 28 serious flags in about 49,650 tasks (0.056%), against 27 for Astra and 42 for GPT-6 Sol. OpenAI notes higher reward-hacking and concealed-uncertainty flags than Astra.
- Dropping to the cheapest tier isn’t free: when a search tool is broken, GPT-6 Luna fails to say so 28.7% of the time, against 2.1% for GPT-6.1 Sol and 1.5% for Astra.
None of this makes the model unsafe to use. It is the reason the rule from our agents guide still applies: give an agent work whose result you can cheaply verify, and keep a human on anything you can’t.
Verdict by job
| Job | Pick | Why |
|---|---|---|
| Everyday coding and long agent loops that reuse context | GPT-6.1 Sol | Astra-level DeepSWE result at about a fifth of the cost; $0.10 cached input |
| Document-heavy professional work | Test Sol against Sonnet 5.5 | Sol wins OpenAI’s chosen document benchmark; the composite index favors Sonnet 5.5 |
| Hardest scientific research tasks | GPT-6 Astra | OpenAI itself says Astra’s 68.1% still leads and should be used here |
| A person waiting on every reply | Lower effort, or budget for Ultrafast | Maximum effort is slow; the fast lane is derived at $12/$60 |
| Bulk work on the cheapest tier | Luna, only after your own broken-tool test | Its failure to flag a broken tool is 28.7% in OpenAI’s test |
One last budgeting note: OpenAI has repriced the Sol line several times since July, so check the pricing page on the day you commit.
Frequently Asked Questions
Is GPT-6.1 Sol as good as GPT-6 Astra?
Nearly, on independent testing: Artificial Analysis scores it 52 against Astra’s 53 on its Intelligence Index. OpenAI says it matches Astra on DeepSWE v1.1 coding at about a fifth of the cost. Astra still leads on the hardest science tasks, where OpenAI recommends it.
How much does GPT-6.1 Sol cost?
$2 per million input tokens, $10 per million output tokens, and $0.10 per million cached input tokens. Cache writes are $2.50. Requests over 272K input tokens are billed at a higher long-context rate, reported as $4 and $15.
Can I use GPT-6.1 Sol in regular ChatGPT?
Not yet, according to OpenAI. It is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-sol. OpenAI says it is not yet available in regular Chat.
Is GPT-6.1 Sol better than Claude Sonnet 5.5 at the same price?
It depends on the task. Trackers citing Artificial Analysis report Sonnet 5.5 at 56 against Sol’s 52 on the composite index, while OpenAI’s own benchmark picks show Sol ahead of Opus 5.5 on AutomationBench and GDP.pdf. I found no cost-per-task figure for Sonnet 5.5, so test both on your own work.
What is GPT-6.1 Sol Ultrafast, and is it cheaper than Astra?
Ultrafast is a premium speed tier of up to 300 tokens per second, billed at six times the standard rate. OpenAI had not published Sol’s Ultrafast line items at announcement. VentureBeat’s derived figures of $12 input and $60 output would put it above Astra’s standard $10 and $50.
Should I move off GPT-5.6 Sol or GPT-6 Sol?
GPT-6.1 Sol costs the same as GPT-6 Sol, with cached input halved, and is half the current promotional price of GPT-5.6 Sol ($4/$20, confirmed only through November 21, 2026). Test it on your own tasks first, and check the model ID in your configuration.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
