The new GPT-5.6 family: Luna, Terra, Sol

According to Simon Willison's Weblog, OpenAI has released its new GPT-5.6 family, available this morning in three sizes: Luna, Terra, and Sol, from smallest to largest. Pricing per million input and output tokens runs from one dollar and six dollars for Luna up to five dollars and thirty dollars for Sol. OpenAI's headline claim is around long-running agentic work: it says Sol scores fifty-three point six on the Agents' Last Exam benchmark, topping Claude Fable 5 by thirteen point one points, with the smaller models beating Fable at a fraction of the cost. Notably, Fable 5 still outperformed the GPT-5.6 lineup on the SWE-Bench Pro coding benchmark, eighty percent to Sol's sixty-four point six percent. OpenAI, meanwhile, published a piece the prior day arguing that roughly thirty percent of SWE-Bench Pro tasks are broken. Willison, who had early access, notes Sol is very competent but hasn't yet struck him as better than Fable on complex coding. The release also brings new API features, including programmatic tool calling, built-in multi-agent subagents, and explicit prompt cache breakpoints.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton