PM-Bench (Optional Heartbeat): leaderboard

Metric: Set-F1 (%; F1 of the prospective actions the agent selects against the actions actually due, accumulated over the 80 steps and 81 scored tasks of the released simulated week, lures included; single agent that may switch on a 30- or 60-minute periodic monitoring heartbeat; one run per model). Source: arxiv.org. Saturation forecast: Around December 2026. 8 models tracked.

Top models

#ModelScore
1GPT-5.479.1
2GPT-5.3 Codex76.6
3Mistral-Large-3-675B-Instruct-251273.7
4Qwen 3 8B71.9
5Llama 3.3 70B Instruct68.1
6Mistral Small 3.253
7Qwen 3 14B50
8Qwen 3 32B48.5

Interactive version: theaggregate.ai/benchmark?slug=pm-bench-optional-heartbeat · How It Works · Data refreshed daily, snapshot 2026-09-29.