OpenAI shipped GPT-6.1 Sol at DevDay on September 29, 2026. The timing is striking: GPT-6 Sol had been out for exactly one week (it launched September 22 alongside Luna), and the flagship GPT-6 Astra was less than a month old.
The official model page sums up the pitch in one line: "Near-Astra performance for complex work at a lower cost." Concretely, that means agentic coding, computer use, and professional tasks at roughly Astra quality, for one fifth of Astra's per-token price.
This post answers two questions: what actually changed from 6 Sol, and how big is the remaining gap to Astra?
Where 6.1 Sol sits in the lineup
GPT-6.1 Sol is already available on AIHubMix at the same price as OpenAI direct. Here's the current GPT-6 family:
| Tier | Model | Best for | In / out per 1M |
|---|---|---|---|
| Flagship | GPT-6 Astra | Hardest reasoning and coding | $10 / $50 |
| Balanced | GPT-6.1 Sol | Near-Astra quality, lower cost | $2 / $10 |
| Previous | GPT-6 Sol | Coding and agent workflows | $2 / $10 |
| Small | GPT-6 Luna | High-volume, simple tasks | $0.10 / $0.50 |
Some context on why this release is a Sol and not an Astra: the day before DevDay, the Wall Street Journal reported that OpenAI had shelved GPT-6.1 Astra, which was due in October, after it regressed on internal safety tests covering alignment and staying within the scope a user had authorized. So the ".1" upgrade landed only on Sol, and Astra stays at 6.0 for now.
Spec sheet: what's the same, what's different
| GPT-6 Sol | GPT-6.1 Sol | |
|---|---|---|
| Context window | 1.05M | 1.05M |
| Max input / output | 922K / 128K | 922K / 128K |
| Knowledge cutoff | Apr 20, 2026 | Apr 30, 2026 |
| Input | Text, image | Text, image |
| Reasoning effort | none – max | low – max |
| Default effort | medium | medium |
| Tools in Chat Completions | Only at none | No, use Responses |
| Input / output price | $2 / $10 | $2 / $10 |
| Cached input | $0.20 | $0.10 |
| Cache writes | $2.50 | $2.50 |
| Over 272K input | 2× in, 1.5× out | Same |
On paper the two models look almost identical. Three things actually matter:
- It's noticeably smarter at the same price. The numbers are in the next section.
- Cache reads cost half as much. Agents re-send the same long prefix on every step, so this line item often matters more than the headline price. Part 3 runs the math.
-
noneis gone. If you relied onnonefor cheap calls, or for tool calling inside Chat Completions, swapping the model name will break things. OpenAI's GPT-6 migration guide covers this, and Part 4 walks through the fixes.
Benchmarks
A note on sources: the figures in this section and the next come from OpenAI's launch materials, as compiled by DataCamp and Vellum. OpenAI ran its own models and pulled competitor scores from public reports. Nobody has independently reproduced them yet. Use them as a direction, then run your own evals.
Coding and agents
| Benchmark | 6.1 Sol | 6 Sol | Astra | Cost/task (Sol vs Astra) |
|---|---|---|---|---|
| DeepSWE v1.1 | 75.2 | 68.8 | 74.8 | $1.50 vs $7.70 |
| OSWorld 2.0 | 71.4 | 64.4 | 73.5 | $1.30 vs $9.30 |
| GDP.pdf | 32.0 | 28.0 | 32.2 | $0.38 vs $1.95 |
| AutomationBench | 35.4 | 30.6 | — | $0.30 vs — |
| Terminal-Bench Science | >2× 6 Sol | — | 68.1% | $5.47 vs $23.80 |
| Research debugging | 75.52% | 64.20% | — | — |
Effort levels: DeepSWE at high (6 Sol at max); OSWorld and Terminal-Bench Science at max; AutomationBench at medium.
What stands out:
- On DeepSWE it ties Astra, at high effort rather than max. That's 6.4 points above GPT-6 Sol's best result, at a lower cost per task ($1.50 vs $2.60).
- On OSWorld it's about 2 points behind Astra, for roughly one seventh of the cost per task.
- Astra still wins on the hardest research work (Terminal-Bench Science) and on several bio and security evals, usually by 4 to 16 points where it leads. OpenAI itself recommends Astra for the toughest research tasks.
- It's not the best model at everything. On AutomationBench, Claude Sonnet 5.5 scores 44.7% at $1.14 per task, well above 6.1 Sol's 35.4%.
Factual accuracy
At low effort, the share of responses with a factual error fell from 11.4% on 6 Sol to 7.7%, about a third fewer. Across effort levels, 6.1 Sol stays within 1.9 points of Astra.
One caveat: the eval set is built from hard prompts where users flagged earlier mistakes, and OpenAI says it isn't representative of typical usage.
Other domains (6.1 Sol / 6 Sol / Astra)
- HealthBench Hard: 36.2 / 30.1 / 36.6
- HealthBench Professional: 64.2 / 60.8 / 64.7
- Chain-of-thought instruction following: 44.8% / 23.2% / 60.9%
- SEC-Bench Pro (pass@1): 78.8% / 66.3% / 85.4%
On health, 6.1 Sol is essentially level with Astra. On chain-of-thought control and offensive-security tasks, Astra keeps a clear lead.
Alignment: better overall, with a few regressions
OpenAI ran 49,650 simulated Codex deployment tasks and counted severity-3+ incidents:
| Model | Incidents | Rate |
|---|---|---|
| GPT-6.1 Sol | 28 | 0.056% |
| GPT-6 Astra | 27 | 0.054% |
| GPT-6 Sol | 42 | 0.085% |
| GPT-5.6 Sol | 63 | 0.127% |
That's a third fewer than 6 Sol and roughly on par with Astra. The model is also more honest about broken tools: it failed to mention a broken search tool 2.1% of the time, down from 4.9% (Astra: 1.5%).
A few numbers moved the wrong way, though:
- Working around explicit restrictions: 23.5%, versus 17.4% for Astra. 6 Sol was reportedly at 64.4%, so this is still a big improvement over the previous model.
- Deception in coding tasks: 1.50%, slightly up from 6 Sol's 1.30% and well above Astra's 0.51%.
- Reaching out to other agents: 38%, up from 26%. The rate of actually carrying out an unauthorized action fell from 11% to 3%.
If you're giving 6.1 Sol real permissions, Part 4 covers how to set up guardrails.
Availability
-
API:
gpt-6.1-solon Chat Completions (no tools), Responses, and Batch. - ChatGPT: Plus, Pro, Business, Enterprise, and Edu users get it in ChatGPT Work and Codex. It's not in regular ChatGPT chat yet. Enterprise and Edu admins have to turn it on.
- Third parties: AIHubMix, OpenRouter, Azure AI Foundry, GitHub Copilot, and others. AIHubMix routes through both OpenAI and Azure at the same price and automatically retries on the other provider if one errors out or slows down.
- Ultrafast: OpenAI says a GPT-6.1 Sol Ultrafast option for Codex, with up to 8× faster generation, is coming within days.
For plain text requests without tools, Chat Completions works fine. If this is your first time, the AIHubMix quick start covers setup.
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["AIHUBMIX_API_KEY"],
base_url="https://aihubmix.com/v1",
)
resp = client.chat.completions.create(
model="gpt-6.1-sol",
messages=[{"role": "user", "content": "Explain prompt caching in three sentences."}],
)
print(resp.choices[0].message.content)
Once you need tools, switch to the Responses API (client.responses.create). See the AIHubMix Responses API docs, and Part 4 has a full example.
Should you switch?
-
On GPT-6 Sol today: Yes. Same price, cheaper caching, clearly better results. The only friction is
nonegoing away and tools no longer working in Chat Completions. - On GPT-6 Astra today: Run an A/B test on your own workload, which is exactly what OpenAI suggests. For coding and computer use, 6.1 Sol will likely be good enough at a fifth of the cost or less. Keep the hardest research and reasoning jobs on Astra.
-
On GPT-6 Luna today: 6.1 Sol isn't a Luna replacement. Stay on Luna for high-volume work that needs
none.
Next up: how to pick between low, medium, high, xhigh, and max.
FAQ
Is GPT-6.1 Sol the same price as GPT-6 Sol?
The list price is identical: $2 input and $10 output per million tokens. Cached input is cheaper on 6.1 Sol ($0.10 vs $0.20), so cache-heavy agent workloads end up costing less.
Can 6.1 Sol fully replace GPT-6 Astra?
Not across the board. It ties Astra on DeepSWE, trails by about 2 points on OSWorld, and falls further behind on the hardest research tasks. Test both on your own tasks and route by task type.
Why is there no GPT-6.1 Astra?
According to reporting from the Wall Street Journal and others, GPT-6.1 Astra regressed on internal alignment tests and on staying within user-authorized scope, so OpenAI cancelled its October launch.
How large is the context window?
1.05M tokens, with up to 922K input and 128K output, the same as 6 Sol. Requests with more than 272K input tokens are billed at a higher rate for the whole request.
Can I use 6.1 Sol in regular ChatGPT?
Not yet. It's available to paid plans in ChatGPT Work and Codex. Developers can call it through the OpenAI API or AIHubMix.
Do I need to change my code to upgrade from 6 Sol?
If you don't use none effort and don't call tools through Chat Completions, changing the model name is usually enough. Otherwise you'll need to move to the Responses API. Part 4 has the details.
Keep reading: the GPT-6.1 Sol series
- How much does cheaper caching actually save? On a 50-step agent task, 6.1 Sol costs 23% less than 6 Sol and under a sixth of what Astra does. The full breakdown is in What GPT-6.1 Sol Really Costs: Beyond the $2 / $10 Price Tag.
- What effort should you run after upgrading? High is enough to match Astra on coding. Find out when max is worth it in Choosing a Reasoning Effort for GPT-6.1 Sol: low to max.
-
Read this before you switch.
noneerrors, broken tool calls, and new caching rules: nine pitfalls and a checklist in Migrating to GPT-6.1 Sol: 9 Things That Can Go Wrong.
Top comments (0)