OpenAI ships GPT-5.6, strong on agents but not the coding leader

OpenAI has released GPT-5.6, its new model family, in three sizes: Luna, Terra, and Sol. All three carry a one million token context window, a February 2026 knowledge cutoff, and up to 128,000 output tokens, and they are live in ChatGPT, Codex, and the API. Pricing runs from Luna at $1 and $6 per million input and output tokens up to Sol at $5 and $30. The more interesting story is where the models are strong and where they are not.

Sol is built for agent work. On the Agents' Last Exam benchmark it scores 53.6, a new high that edges past Claude Fable 5, and OpenAI says the smaller models reach Fable 5's level at roughly one-sixteenth the cost. On the Artificial Analysis Coding Agent Index, Sol sets the top score at 80 while using less than half the output tokens of its nearest rival. But it does not win everywhere. On SWE-Bench Pro, Fable 5 still leads clearly, 80 percent against Sol's 64.6, and OpenAI responded by publishing its own critique of that benchmark's reliability. Simon Willison, who had early access, calls Sol "definitely very competent" but says it has not beaten Fable on the harder coding tasks he tried.

For developers the new API features may matter more than the headline scores: programmatic tool calling that lets the model orchestrate tools in JavaScript, native support for spawning parallel subagents, and explicit prompt cache breakpoints. OpenAI also shipped ChatGPT Work, an enterprise assistant for documents, spreadsheets, and slides.

Why it matters

If you build agents or coding tools, the choice between GPT-5.6 and Fable 5 now depends on your exact workload rather than one leaderboard. Sol looks strong and cheaper for agent orchestration, but test it on your own coding tasks before switching, because Fable still wins some of them.