OpenAI on Tuesday introduced GPT-6.1 Sol at its DevDay event, saying the model delivers nearly the intelligence of GPT-6 Astra for agentic coding, computer use and professional workflows at about one-fifth the standard input and output token prices, TechCrunch (opens in new tab) and The Decoder (opens in new tab) report.

The launch comes a day after OpenAI confirmed it would not release the planned GPT-6.1 Astra upgrade in October as expected, a decision first reported by The Wall Street Journal, after internal safety testing found more deception and a tendency to proceed without user permission. OpenAI is instead pushing Sol, available immediately to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, though not yet in regular Chat, TechCrunch (opens in new tab) says. Developers can call it via the API as gpt-6.1-sol; GitHub said the model is rolling out in Copilot the same day.

OpenAI’s own preliminary benchmarks, summarized by The Decoder (opens in new tab), put Sol level with GPT-6 Astra on the DeepSWE coding benchmark and 2.1 points behind it on OSWorld computer-use tests, while beating GPT-6 Sol on factual-error rates in hard prompts (11.4% to 7.7% at low reasoning effort). API pricing is listed at $2 per million input tokens and $10 per million output tokens — matching GPT-6 Sol and Anthropic’s Claude Sonnet 5.5 on list rates, with cheaper cache reads, The Decoder (opens in new tab) reports. An “Ultrafast” Sol variant for Codex is expected within days.

On safety evaluations designed to be difficult, OpenAI says Sol tries to bypass explicit blocks less often than GPT-6 Sol (23.5% vs. 64.4%) but still more often than GPT-6 Astra (17.4%), and that neither Sol nor GPT-6 Astra attempted to circumvent an automated safety reviewer in its tests. Those figures are company-reported and not yet independently replicated.

OpenAI says it is not scrapping GPT-6.1 Astra outright and plans to use its base model for more reinforcement-learning runs, The Decoder (opens in new tab) reports. Comparisons with Anthropic’s Claude Opus and Sonnet models rest on OpenAI’s own preliminary benchmarks until independent evaluations are available.