OpenAI releases GPT-5.6

OpenAI released GPT-5.6, the next update to the GPT-5 line it launched in August 2025, as a family of three models ranked by capability: Luna, the fastest and cheapest; Terra, a mid-tier option OpenAI said cost roughly half its predecessor to run; and Sol, the flagship, which the company described as its strongest cybersecurity model yet, built for threat modelling, code review, patching and blue-team work. A limited preview had gone to a small group of partners on 26 June, with reporting attributing the gap before the wider 9 July rollout to a review tied to US government security concerns.

OpenAI said Sol set new marks on coding, knowledge work, cybersecurity and science benchmarks against both its own predecessors and competing frontier models, while using fewer tokens; on the Artificial Analysis Coding Agent Index, Sol scored 80 with maximum reasoning enabled, ahead of Anthropic’s Fable 5, while using under half the output tokens and about a third less cost, per that benchmark’s reporting. OpenAI’s system card rated the Sol and Terra models “High” capability in cybersecurity — the tier below its “Cyber Critical” threshold — noting the models could find vulnerabilities and exploit primitives but did not autonomously produce full working exploit chains against Chromium and Firefox in OpenAI’s own tests.

The release was announced the same day OpenAI named GPT-5.6 the preferred model across Word, Excel, PowerPoint and Copilot Chat, a move that followed reports that Microsoft had begun quietly substituting its own in-house models for OpenAI’s inside some Office products.