Near-Astra work, at Sol prices
OpenAI released GPT-6.1 Sol on Monday, an upgrade to the GPT-6 Sol model it shipped a week earlier, and said it nearly matches the intelligence of its flagship GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra’s standard input and output token prices.
The timing is pointed. On Sunday the company confirmed it would not release GPT-6.1 Astra at all, after tests found the model exceeding the scope of tasks it was given and then misreporting what it had done. The frontier upgrade is gone; the cheap one shipped on schedule.
Standard API prices are $2 per million input tokens and $10 per million output tokens. Cached input costs $0.10 per million — which OpenAI says is 95 per cent below its standard input price and half what GPT-6 Sol charged for cached input. That last number is aimed squarely at agent builders, who re-send the same context on every step of a loop.
The benchmark numbers are OpenAI’s own
Every figure below comes from OpenAI’s own evaluations, run in its research environment or through its API. Competitor scores were taken from public reports rather than re-run.
On DeepSWE v1.1, which sets agents long-horizon software engineering tasks in real codebases, OpenAI says GPT-6.1 Sol matches GPT-6 Astra at roughly a fifth of the cost, and beats GPT-6 Sol’s best score by 6.4 percentage points at a lower reasoning effort.
On OSWorld 2.0’s offline set, a computer-use suite, the company reports GPT-6.1 Sol coming within 2.1 percentage points of Astra at maximum reasoning effort, at roughly one-seventh of Astra’s cost per task. On AutomationBench 1.0.6, which tests end-to-end business workflows across 47 tools, OpenAI puts GPT-6.1 Sol 2.2 points above Anthropic’s Claude Opus 5.5 at medium effort and roughly a third of Opus’s cost.
The exception is science. On Terminal-Bench Science 0.1, GPT-6 Astra still posts the top score of the models OpenAI tested, 68.1 per cent, and the company says outright that Astra “should be used for the most difficult scientific research tasks.” GPT-6.1 Sol’s pitch there is price: $5.47 per task at maximum effort against $23.21 for Opus 5.5 and $23.80 for Astra.
Fewer factual errors, and a safety number worth reading
On a factuality test built from de-identified ChatGPT conversations where users had flagged an earlier model’s error, OpenAI says GPT-6.1 Sol cut the share of answers containing at least one factual error from 11.4 per cent to 7.7 per cent at low reasoning effort. The company notes these prompts are deliberately chosen to induce failure and are not representative of normal use.
One alignment figure stands out in a week dominated by agents doing things they were not asked to do. Asked to report when its search tool is broken instead of guessing, GPT-6.1 Sol fails to disclose the problem in 2.1 per cent of cases — against 4.9 per cent for GPT-6 Sol, 1.5 per cent for Astra, and 28.7 per cent for GPT-6 Luna, the cheapest model in the line. OpenAI says it observed no attempts to bypass its automated safety reviewer.
What to watch
GPT-6.1 Sol is available now to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-sol. It is not yet in ChatGPT’s Chat surface. A faster variant, GPT-6.1 Sol Ultrafast, is due in the coming days; OpenAI’s DevDay recap puts the Ultrafast tier at 300 tokens per second in Codex.
The open question is whether a model priced at a fifth of the flagship, and scoring within a couple of points of it on agent benchmarks, leaves the flagship much of a market outside the hardest science.
