OpenAI releases GPT-6.1 Sol, close to Astra at a much lower cost

OpenAI introduced GPT-6.1 Sol at its DevDay event on September 29, 2026. The model replaces GPT-6 Sol and, according to its system card, is the first member of the GPT-6.1 family. OpenAI's claim is a bold one: in agentic coding, computer use and professional work, GPT-6.1 Sol comes close to the company's strongest model, GPT-6 Astra, at a much lower cost.

What GPT-6.1 Sol offers

According to its model page, GPT-6.1 Sol has a context window of 1.05 million tokens and can produce up to 128,000 tokens in a single response. Its knowledge cutoff is April 30, 2026. It reads text and images and outputs text only. Reasoning depth can be set in five steps, from low to maximum. It supports tools such as web search, file search, code execution, computer use, MCP connections and Skills, and a multi-agent feature in beta lets it split work among sub-agents within a single request.

In the benchmarks OpenAI published, Sol is on par with Astra in coding tasks and close to it in extracting information from complex PDF documents and in computer use. The company still recommends Astra for the hardest scientific work. It also reports a clear drop in factual errors compared with the previous Sol on difficult conversations that users had flagged.

Where it is available

Starting today, GPT-6.1 Sol is rolling out in ChatGPT Work and Codex to Plus, Pro, Business, Enterprise and Edu users. According to OpenAI, it is not yet available in regular chat. Developers can use it in the API as gpt-6.1-sol. GitHub began a gradual general rollout in Copilot the same day, and the model is generally available on Amazon Bedrock. OpenAI says an Ultrafast version with much faster generation is coming in the next few days. Current prices are on OpenAI's site.

A high safety classification, and Astra's successor on hold

According to the system card, GPT-6.1 Sol is rated "Critical" for cybersecurity, the top level in OpenAI's risk framework, and "High" for biology and chemistry, and ships with the same safeguards as Astra. In OpenAI's stress tests, the model did clearly better than the previous Sol on behaviors such as trying to get around warnings and pushing limits during computer use. The same week, OpenAI confirmed to the press that it had shelved GPT-6.1 Astra, planned for October, because it did not meet its safety and alignment bar; according to The Register, the model fell short on staying within the scope and permissions it was given.

All the published benchmarks were run by OpenAI, and in some tests rival models pull ahead of Sol at the highest settings. It is wise to test the model on your own workflows before switching, and to keep permissions narrow when using it in agents. To see how it really performs, it is worth following evaluations from independent researchers.

Sources: OpenAI: Introducing GPT-6.1 Sol, OpenAI: GPT-6.1 Sol model page, OpenAI: GPT-6.1 Sol system card, GitHub: GPT-6.1 Sol in Copilot, AWS: GPT-6.1 Sol on Amazon Bedrock, The Register: GPT-6.1 Astra shelved.