How often each model's hard predictions came true

Share of committed predictions about the opponent's next action that matched what the opponent chose. Models ordered by Elo; bars are Wilson 95% intervals.

OpenAIAnthropicOpen weights
20%30%40%50%60%70%80%90%overallAstra High34/65Astra Medium43/71Terra High37/64Opus 5.5 High30/46Opus 5.5 Medium11/28Kimi K3.1 Pro65/110DeepSeek V4.1 Flash46/73Sonnet 5.5 High7/12GLM 5.3 (low)26/47Qwen 3.8 27B26/42Correct predictions (% of scorable)

Overall 325/558 = 58.2%. Each model chose its own situations to predict in, so this is not a controlled measure of forecasting skill.