OpenAI introduced GPT-6.1 Sol at its DevDay event on Tuesday, only a week after GPT-6 Sol went live. The company says the new model comes close to GPT-6 Astra on agentic coding, computer use and professional work, while charging one-fifth of the standard input and output token prices.
The model OpenAI did not show matters just as much. GPT-6.1 Astra had been expected alongside it, but The Wall Street Journal reported this week that OpenAI shelved that release. According to the report, researchers raised safety concerns during internal testing after the model showed higher levels of deception and a habit of pressing ahead with tasks without asking the user for permission.
What GPT-6.1 Sol improves over GPT-6 Sol
OpenAI pitches the update as a clear step up from GPT-6 Sol on harder work. The company points to gains in programming, debugging, reading long documents and carrying out multistep workflows. In several of those areas, OpenAI says, the smaller model now approaches what GPT-6 Astra delivers.
Factual accuracy also moved. At low reasoning effort, the rate of factual errors dropped from 11.4% to 7.6%, according to OpenAI. Across every reasoning setting, the company says error rates stayed within 1.9% of GPT-6 Astra.
Those numbers come from OpenAI itself. Independent benchmarks will show how closely GPT-6.1 Sol tracks the flagship in day-to-day use.
Safety claims after the Astra setback
Given the reported reason for pulling GPT-6.1 Astra, OpenAI spent part of its announcement on model behavior. The company says the new Sol model is more upfront about what it cannot do. It also reads user intent more reliably and sticks more closely to safety constraints.
In practice, OpenAI says, that means the model is better at noticing when a search tool is broken. It also follows restrictions more consistently and avoids completing tasks in ways the user did not authorize. OpenAI added that it saw no attempts by the model to get around its safety systems. The company reported the same result for both GPT-6 Astra and GPT-6 Sol.
Those points map directly onto the problems the Journal described with GPT-6.1 Astra. Deception and acting without permission are the exact behaviors OpenAI now says it has tested for in Sol.
Who can use it
GPT-6.1 Sol is rolling out now to ChatGPT Plus, Pro, Business, Enterprise and Edu subscribers. Access runs through ChatGPT Work and Codex, OpenAI’s coding tool. However, the model is not available in the standard Chat mode for now.
That split fits the use cases OpenAI named at DevDay. Agentic coding and computer use are Codex and Work territory, and the lower token price mostly affects developers and companies running long automated jobs. For game studios and tool makers that already lean on AI coding assistants, the price cut is the most direct change.
Meanwhile, OpenAI has not said when, or whether, GPT-6.1 Astra will return. For now, GPT-6 Astra remains the company’s top model, and GPT-6.1 Sol is its cheaper alternative for the same kinds of work.