TLDR overview
- GPT-5.6 ships in three variants, Sol, Terra & Luna. We ran evaluation for Sol & Terra through the same 4,444-task Java benchmark we used for GPT-5.5, so every number here is directly comparable.
- Sol is the stronger of the two on correctness: 81.99% pass rate, up from GPT-5.5's 78.66%.
- The security profile changed shape. Blocker vulnerabilities dropped by half, from 18 to 9 per mLOC, while critical vulnerabilities moved from 20 to 125 per mLOC.
- Concurrency and threading is now the largest bug category by a wide margin, at 352 per mLOC for Sol.
- Terra takes a different route to a similar result: 12% less code than GPT-5.5, a pass rate close to it, and a code smell density 37% higher.