GPT-5.6 reaches general availability with gains in code, cybersecurity and efficiency

OpenAI has made the GPT-5.6 family generally available following its limited preview: Sol, the flagship; Terra, for everyday work; and Luna, the most affordable option. They arrive today in ChatGPT, Codex and the API, with a gradual rollout over the next 24 hours.

The company highlights efficiency as the core of the launch. On Agents' Last Exam, which evaluates long-running professional workflows, Sol beats Claude Fable 5 by 13.1 points, while Terra and Luna match it at a fraction of the cost. In programming, it sets the best score yet on the Artificial Analysis Coding Agent Index. Sol also makes a leap in interface design: it can inspect and visually refine its own results, not just generate the underlying code. It also improves the quality of documents, spreadsheets and presentations, following reference templates and formats more faithfully than GPT-5.5. It debuts two new settings: max, which extends reasoning time, and ultra, which coordinates four agents in parallel for complex tasks.

In science, it improves results in biology and chemistry compared with GPT-5.5. In cybersecurity, it goes from GPT-5.5's 47.9% to 73.5% on ExploitBench. OpenAI notes that it remains more useful for fixing vulnerabilities than for carrying out attacks, and that the model does not cross the "Critical" threshold in either cybersecurity or biology. It reserves its most sensitive capabilities for verified users through the Trusted Access for Cyber program, following 700,000 hours of automated testing.

Inside OpenAI, researchers doubled their daily token usage during GPT-5.6's internal testing compared with GPT-5.5's peak. Sol is available on paid ChatGPT plans, while free plans access Terra in ChatGPT Work and Codex. In the API, Sol costs $5 per million input tokens and $30 per million output tokens; Terra, $2.50 and $15; and Luna, $1 and $6.

Key points

• GPT-5.6 (Sol, Terra and Luna) is now available in ChatGPT, Codex and the API.
• Sol beats Claude Fable 5 on Agents' Last Exam at a lower cost.
• Best score yet on the Artificial Analysis Coding Agent Index, in programming.
• Leap in interface design and better quality in documents and presentations.
• New modes: max, more reasoning time; ultra, four agents in parallel.
• Improves biology and chemistry results compared with GPT-5.5.
• In cybersecurity, rises from 47.9% to 73.5% on ExploitBench.
• Does not cross the "Critical" threshold in either cybersecurity or biology.
• Most sensitive capabilities reserved for verified users (Trusted Access for Cyber).
• Sol on paid ChatGPT plans; Terra on free plans.