OpenAI just dropped GPT-5.6, and it's a three-model lineup: flagship Sol, balanced Terra, and budget Luna. They're live in ChatGPT, Codex, and the API.
The big headline? Sol absolutely crushed it on the Agents' Last Exam, scoring 53.6 — a full 13.1 points ahead of Anthropic's Claude Fable 5. Even at medium reasoning intensity, Sol maintains an 11.4-point lead, and OpenAI says it costs about a quarter as much to run.
Coding chops are also strong. Sol scored 80 on the Artificial Analysis coding agent index, just 2.8 points ahead of Fable 5 — but it outputs less than half the tokens, takes less than half the time, and costs roughly a third less.
GPT-5.6 also introduces an "ultra" mode that defaults to running four agents in parallel, burning more tokens for better results and shorter runtimes. API users can tap into the multi-agent beta, too.
Pricing per million tokens: Sol — $5 input, $30 output; Terra — $2.50 input, $15 output; Luna — $1 input, $6 output. Pick your poison.