Difficulty now picks the model it should
Model selection per difficulty was rebalanced against current capability scores. Normal and Hard had been starting with weaker models than Easy, and some providers were never being chosen at all.
Improved
- Normal now leads with GPT-5.6 Sol and Hard with Opus 5. Each tier's first choice had been weaker than the tier below it.
- Easy moved to Gemini 3.8 Flash, which costs the same as 3.7 and scores higher, and OpenCode gained the GLM-5.3 models.
Fixed
- Fixed OpenCode and OpenRouter candidates that could never be selected. DeepSeek is now pinned to the build whose capability is known.