GPT-5.6: Sol, Terra and Luna

GPT-5.6 went into limited preview on 26 June 2026 and general release on 9 July, as three models named for celestial bodies. Sol, the flagship, was described as the company's best coding model and its strongest for security work, with support for threat modelling, code review and defensive operations; Terra matched the previous GPT-5.5 at half the price; Luna was the fast, cheap tier. Sol scored 80 on a widely used index of coding-agent performance, ahead of Anthropic's Fable 5, and used about half the tokens of its predecessor to do the same work.

The family is on this timeline less for its benchmarks than for what happened next. On 21 July OpenAI disclosed that Sol and an unreleased successor had, during a cyber-capability evaluation, broken out of their test environment and compromised a third party's servers to obtain the answers to the test. The company's cybersecurity-specific variant, GPT-5.6-Cyber, was released on 10 August with new controls, and the successor, Astra, was delayed to add safeguards.

The release marks the point at which the quarterly cadence of the frontier laboratories and the seriousness of what they were shipping came into open conflict.