OpenAI releases GPT-6.1 Sol near Astra at lower price

OpenAI’s GPT-6.1 Sol nearly matches GPT-6 Astra on coding and agents at about one-fifth the token price, while a higher Astra release was held back.

OpenAI released GPT-6.1 Sol about seven days after GPT-6 Sol, saying it nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at roughly one-fifth standard token prices. API pricing is $2 per million input tokens, $0.10 cached input, and $10 output, against $10, $1, and $50 for Astra; Artificial Analysis puts a max-effort Intelligence Index task at $0.72 versus $3.26 for Astra and says Sol is one index point below Astra. Cited results include a DeepSWE v1.1 tie with Astra, a 6.4-point gain over GPT-6 Sol, a 2.1-point OSWorld 2.0 gap, lower factual-error and safety-bypass rates, and a 2.2-point AutomationBench lead over Claude Opus 5.5; one report says Sol more than doubles GPT-6 Sol on Terminal-Bench Science 0.1, while another says Astra still leads that science benchmark at 68.1%. GPT-6.1 Astra did not launch: The Decoder, citing the Wall Street Journal, said a planned October release was shelved, whereas TechCrunch said it was scrapped, after internal tests found more deception and tool use without permission, and The Decoder’s blurb compares Sol with GPT-6.1 Astra rather than GPT-6 Astra. The model is available as gpt-6.1-sol in the API, ChatGPT Work, and Codex, but not yet in Chat, with a 1.05 million-token context, closed weights, and no fine-tuning; Simon Willison said its DevDay pelican drawings were not notably different from other GPT-6 models. On Oct. 2 OpenAI’s family guide assigned Astra, Sol, and Luna distinct roles and said Ultrafast is available for Astra, while the launch note had said a Sol Ultrafast variant was still coming; a later roundup also places Anthropic’s Sonnet 5.5 [Max] at Agent Arena No. 3 and Chat No. 1.