FIRE-Bench: leaderboard

Metric: Claim-level F1 (%; agent conclusions and the source paper's findings are split into atomic claims and matched by semantic entailment with a GPT-5.2 judge; mean over the 30 core LLM-behaviour rediscovery tasks and three runs; agent runs full research cycles from the research question in a GPU sandbox). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.

Top models

#ModelScore
1Claude Sonnet 4 (Claude Code)46.7
2GPT-5 (Codex CLI, medium)41.9
3GPT-5 (OpenHands)37.9
4o4-mini (OpenHands)31.9

Interactive version: theaggregate.ai/benchmark?slug=fire-bench · How It Works · Data refreshed daily, snapshot 2026-09-26.