Elo by model family

Bradley–Terry Elo with approximate 95% intervals. Families run from weakest to strongest; effort variants share a column.

OpenAIAnthropicOpen weightsHigh / default effortMedium effort
10001200140016001800Qwen 3.827BGLM 5.3(low)Sonnet5.5DeepSeekV4.1 FlashKimiK3.1 ProOpus5.5GPT-6.1TerraGPT-6Astrahighmediumhighmedium

Only Opus 5.5 and GPT-6 Astra were run at two effort levels. GLM 5.3 was run at low effort throughout because higher settings stalled.