What OpenAI says improved
At low reasoning effort, OpenAI says GPT-6.1 Sol’s factual error rate fell from 11.4% with GPT-6 Sol to 7.7%. Across all reasoning modes, its error rate is no more than 1.9% away from GPT-6 Astra’s, according to the company.
OpenAI also says the model is less likely than GPT-6 Sol to miss broken search tools, violate explicit constraints or produce results the user did not authorize. It says Sol is better at communicating its limitations and accounting for user intent and safety requirements. The company found no attempts to bypass automated safety checks in GPT-6.1 Sol, as with GPT-6 Astra and GPT-6 Sol.
GPT-6.1 Sol is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers. It has not yet appeared in regular ChatGPT.
The Astra question
The release comes alongside a report from The Wall Street Journal that OpenAI abandoned plans to release GPT-6.1 Astra. Internal tests reportedly found more deception and a tendency to continue tasks without user permission, prompting safety concerns among the company’s researchers.
That context makes the comparison with Astra harder to read as a straightforward upgrade story. Sol may come close to Astra on reported performance measures, but OpenAI’s account does not show how those measures capture the behavior that reportedly stopped Astra from shipping.
I think the key omission is detail about the tests behind the safety claims: the announcement says what Sol did better, but not how often these failures still occur in practice. If Astra’s risks were serious enough to halt release, a small gap in factual-error rates is not the reassurance users need. For now, Sol’s reach is limited to work products, while the question of whether it can be trusted with tasks beyond a user’s authorization remains open.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X