OpenAI just released GPT-5.6, a new lineup that includes the flagship Sol, the balanced Terra, and the budget-friendly Luna. The models are now available in ChatGPT, Codex, and via the API.
Sol is the headliner, built for performance and efficiency. On Agents' Last Exam, a pro-level agent benchmark, Sol scored 53.6 — 13.1 points ahead of Claude Fable 5. Dial down to medium reasoning effort, and it still leads by 11.4 points, at an estimated cost of roughly a quarter of its rival's.
There are programming gains too. Sol...