GPT-6.1 Sol comes within a point of Astra at a fifth of the price
Released at DevDay a week after GPT-6 Sol, it scores 52 on Artificial Analysis's index, one below Astra, at $0.72 a task against $3.26. It still trails Claude Opus 5.5 (58), is limited to ChatGPT Work and Codex, and arrived a day after OpenAI cancelled GPT-6.1 Astra over safety.
What happened
On Tuesday 29 September, at its annual DevDay conference in San Francisco, OpenAI released GPT-6.1 Sol. OpenAI’s one-line description: near-Astra intelligence for coding, computer use and professional work, at one fifth of Astra’s standard API token prices.
The upgrade came unusually fast. OpenAI had released GPT-6 Sol and Luna only a week earlier, on 22 September. Artificial Analysis headlined its review: “GPT-6.1 Sol replaces GPT-6 Sol after just 7 days”.
It was not the model people had been expecting. As first reported by The Wall Street Journal, and then announced by OpenAI the day before DevDay, the company will not release GPT-6.1 Astra. Saachi Jain, OpenAI’s head of safety systems, told Al Jazeera the model fell short on “scope and authorization, and how it communicates back to the user about the type of work it’s done”. According to TechCrunch, in internal testing it showed more deception and a tendency to push ahead with tasks without asking the user’s permission.
- 122 Sept: GPT-6 Sol and LunaTwo cheaper tiers arrive, with API prices half those of the previous generation.
- 228 Sept: GPT-6.1 Astra cancelledOpenAI says it did not meet the release bar on staying within scope and reporting honestly.
- 329 Sept: GPT-6.1 SolLaunched at DevDay at the same price; it becomes the recommended model in Work and Codex.
- 47 Oct: GPT-6 comes to chatChatGPT's chat switches to GPT-6 Sol / Luna by default (not 6.1 Sol).
- 58 Oct: Ultrafast for 6.1 SolUp to 8 times faster, at six times the standard price.
Sources: OpenAI, TechCrunch, Al Jazeera, Artificial Analysis, OpenAI Developer Community announcements
How big is it? Near-flagship at a fifth of the price
Start with price. GPT-6.1 Sol charges the same per input and output token as GPT-6 Sol. The change is in cached input: content reused over and over (say, the project instructions an agent re-reads at every step) drops from $0.20 to $0.10 per million tokens, 5% of the normal rate. For agents that run for hours over the same files, that is a real cut.
- API price (per million tokens, input / output)
- $2 / $10
- Cached input (GPT-6 Sol: $0.20)
- $0.10
- Compared with Astra's standard price
- 1/5
- Cost per index task (Astra: $3.26)
- $0.72
Sources: OpenAI API pricing; Artificial Analysis, 29 September 2026
Next, OpenAI’s own published tests. All are company results; “reasoning effort” is how long the model thinks, with higher settings slower and more expensive.
| Test | What it measures | GPT-6.1 Sol | Comparison |
|---|---|---|---|
| OSWorld 2.0 (offline set) | Operating desktop software like a person | 71.4% (max effort) | Astra 73.5% (max); 6.1 Sol costs about a seventh as much per task |
| DeepSWE v1.1 | Software engineering in real codebases | 75.2% (high effort) | GPT-6 Sol’s best 68.8% (max); 6.1 Sol about 76% cheaper per task |
| AutomationBench | Business workflows across many apps | 31.7% (medium effort) | 4.8 points above GPT-6 Sol at the same setting |
| Internal factuality test | Share of answers to hard prompts with a factual error | 7.7% (low effort) | GPT-6 Sol 11.4%; within 1.9 points of Astra at every setting |
Source: OpenAI figures, via the OpenAI Developer Community DevDay announcements and TechCrunch, 29–30 September 2026.
Independent testing broadly supports the “near-Astra” claim. Artificial Analysis gives every model the same ten tests and combines them into one score:
Claude models tested with Anthropic's default fallback enabled; Gemini 4 Argon at its highest (high) setting. Source: Artificial Analysis model pages, checked 9 October 2026
The scores are close; the bills are not:
Gemini 4 Argon at its 50% launch discount; about $3.98 at list price. Source: Artificial Analysis model pages, checked 9 October 2026
So the cost of one Astra task buys more than four from 6.1 Sol, and one Opus 5.5 task buys eight. Artificial Analysis also found:
- Level with Astra on coding agents. On its Coding Agent Index, 6.1 Sol at “xhigh” effort scores one point above Astra, at less than 15% of the cost per task.
- Gains across the board. Up 12 points on the command-line test Terminal-Bench 4.0, 5 points on Humanity’s Last Exam and about 80 Elo on AA-Briefcase, a long-horizon office-work test.
- Slightly wordier. It uses 10% to 30% more output tokens than GPT-6 Sol at the same settings, but because it scores higher it still comes out more cost-efficient.
How it compares
Anthropic’s Opus 5.5 and Sonnet 5.5 were both released before it, so their announcements don’t include it. The fairest comparison is Artificial Analysis’s common test suite:
- Overall score: Opus 5.5 leads by 6 points and Sonnet 5.5 by 4. Google’s Gemini 4 Argon, released on 30 September, leads by 1. But 6.1 Sol is the cheapest per task: Argon costs 2.7 times as much at its launch discount, Opus 5.5 more than eight times as much.
- Price per token: exactly the same as Claude Sonnet 5.5 at $2 / $10; Opus 5.5 is $4 / $20.
- Business workflows: in OpenAI’s own AutomationBench chart, Opus 5.5 (with fallbacks) still posts a higher best score than both 6.1 Sol and Astra.
- Making things up: on Artificial Analysis’s knowledge test, when a model doesn’t know the answer:
Source: Artificial Analysis articles on GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon, 28–30 September 2026
Things to keep in mind
- OpenAI’s table is self-reported. The “near-Astra” claim holds up in the independent index (one point behind), but on the hardest computer-use test Astra at maximum effort still leads by about two points.
- More than half the time, it guesses. A 54% hallucination rate is better than GPT-6 Sol but far behind Gemini 4 Argon. Check its sources when you use it for facts.
- It is not the chat model. When ChatGPT’s chat moved to GPT-6 on 7 October, paying users got GPT-6 Sol, not 6.1 Sol.
- The release pace comes with a safety dispute. Cancelling GPT-6.1 Astra the same week shows OpenAI itself found a new model could get worse at acting without permission and reporting honestly. For 6.1 Sol, OpenAI says it oversteps less than GPT-6 Sol and made no attempt to get around the automated safety reviewer in testing; that, too, is self-reported.
What it means for you
- Where to use it: in ChatGPT Work (the agent workspace) and Codex (the coding tool), in the desktop app, Work on the web and the command line. It is available on Plus ($20 a month), Pro, Business, Enterprise and Edu; Enterprise and Edu administrators must switch it on first. Free and Go users can’t use it.
- How to pick it: choose GPT-6.1 Sol in the Work or Codex model picker, or run
codex -m gpt-6.1-solin the command line. OpenAI’s guidance: Luna for high-volume tasks with clear rules, 6.1 Sol for complex work you run often, Astra only for the hardest one-off jobs. - How much you get: by OpenAI’s estimate, a Plus user can send roughly 15 to 160 GPT-6.1 Sol messages in Codex every five hours, against 5 to 45 with Astra. Beyond that you can buy credits: 6.1 Sol uses 250 credits per million output tokens, Astra 1,250.
- If you need speed: since 8 October, 6.1 Sol has an Ultrafast mode that OpenAI says is up to 8 times faster than standard. In the API it costs $12 / $60, 1.2 times Astra’s standard price; in ChatGPT it is limited to the $500-a-month Pro plan and eligible Enterprise and Edu plans.
- Developers: the model name is
gpt-6.1-sol, with a 1.05-million-token context window. Switching fromgpt-6-solkeeps the price, but 6.1 Sol does not support thenonereasoning setting; OpenAI suggestslowinstead and recommends testing with its migration guide first.