A model can lead on one game statistic and lose on the actual objective. In OpenRouter's 30-game battle-royale experiment, OpenAI's GPT-5.4 recorded the most eliminations, 38, but won twice. xAI's Grok 4.1 Fast recorded 30 eliminations and won 13 times.
Eleven models chose actions through tools in the same simulated environment and could keep notes between matches. The scoring rewarded survival and placement as well as combat. The results illustrate how the chosen objective changes a model comparison; they do not isolate training choices as the cause of behavior or establish safety around people.
Terms in this piece · Glossary
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
