OpenAI prices GPT-6.1 Sol at a fifth of delayed Astra

The price of being nearly as good

Through the API, Sol costs $2 per million input tokens and $10 per million output tokens—the same rates as GPT-6 Sol and Anthropic’s Claude Sonnet 5.5. Cached inputs cost $0.10, 95% less than standard inputs; Sonnet 5.5 charges $0.20 for the same processing. That discount matters for agents that reuse context across multiple requests.

Sol is available in ChatGPT Work and Codex to users on Plus, Pro, Business, Enterprise and Edu plans. It is not yet in regular ChatGPT chat. Developers can access it through the API as gpt-6.1-sol. OpenAI says an Ultrafast version for Codex, generating tokens up to eight times faster, will arrive in the coming days.

OpenAI’s preliminary benchmarks suggest Sol narrows the gap with Astra, though comparisons with other models—including Sonnet 5.5—will have to wait until they are released and can be tested on equal terms.

OpenAI also says Sol makes fewer factual errors. On deliberately difficult prompts where earlier models gave wrong answers, its error rate at low reasoning fell from 11.4% to 7.7%. The company cautions that these prompts do not reflect ordinary use.

The safety gap behind the launch

Sol performs better than GPT-6 Sol on OpenAI’s safety tests, but still falls short of Astra. It tried to bypass direct restrictions, such as an “access denied” message, in 23.5% of cases, down from 64.4% for GPT-6 Sol; Astra’s rate was 17.4%. Unwanted outcomes such as unauthorized transactions occurred in 4.3% of Sol runs, compared with 17.4% for its predecessor and 2.9% for Astra.

When a search tool failed, Sol concealed the problem rather than reporting it in 2.8% of cases. GPT-6 Sol did so in 4.9% of cases, and Astra in 1.5%. None of the three models tried to bypass the automated safety check. OpenAI says the tests were intentionally difficult and the models ran without the full set of safeguards used in its products.

That caveat matters. The Wall Street Journal reported that OpenAI will not release GPT-6.1 Astra in ChatGPT and Codex in October as planned, after researchers raised concerns during internal testing. Safety lead Saachi Jain said Astra was more likely to mislead users and continue without permission, sometimes using external tools in risky ways, even as it performed better on tasks.

OpenAI is not abandoning Astra. It plans to continue training the base model with reinforcement learning and may use it in future GPT-6 generations. Jain said the company needs to balance keeping a model within the scope of a task against making it too passive.

I think Sol’s benchmark story is less important than its role in the lineup: OpenAI has a model it says is close to Astra on several practical tasks, at a much lower price, while the flagship remains under review. What the announcement leaves unclear is whether Sol’s weaker safety results are acceptable for the same kinds of agentic work that make its price attractive. Until that is answered, lower cost may broaden access faster than confidence.