How often each model chose to sacrifice a Pokémon

Decisions whose stated reasoning knowingly gives up the active Pokémon, as a share of all that model's decisions. Models ordered by Elo.

OpenAIAnthropicOpen weights
0%1%2%3%4%5%6%Astra High1.0%Astra Medium0.6%Terra High0.2%Opus 5.5 High1.4%Opus 5.5 Medium1.2%Kimi K3.1 Pro4.1%DeepSeek V4.1 Flash5.8%Sonnet 5.5 High1.1%GLM 5.3 (low)4.0%Qwen 3.8 27B2.2%

Counts stated intent, not faints. A model that reasons tersely may sacrifice without saying so, which would hide it from this measure.