OpenAI launches GPT-5.6 series, Sol flagship beats rivals in coding

OpenAI has released its GPT-5.6 series with three models: Sol (flagship), Terra (balanced), and Luna (speed-focused). For now, only a few 'trusted partners' picked by the US government can preview them. Sol costs $5 per million input tokens and $30 per million output tokens, while Terra and Luna are cheaper. They also improved prompt caching to save costs.

In coding tests, Sol scored 88.8% in standard mode on Terminal-Bench 2.1, beating Claude Mythos 5's 88.0%. With Ultra mode, it hit 91.9%. In biology tasks (GeneBench v1), Sol used fewer tokens but performed better than GPT-5.5.

For security, Sol did much better on long tasks like finding bugs. In ExploitBench, it used about one-third of the output tokens to match Mythos Preview's performance. OpenAI plans to make all three models public in the coming weeks, and Sol will be on Cerebras in July, reaching up to 750 tokens per second.