Agents' Last Exam
Multi-domain agentic reasoning benchmark spanning professional long-horizon workflows. OpenAI's GPT-5.6 launch reports standard-mode scores; preserve harness and mode labels because code-mode and standard-mode rows are not directly comparable.
Models ranked
12
tracked on this benchmark
Score band
75.8 – 26.3
best → lowest tracked
Snapshot trend
+32.60
Oct 23 → Nov 18 · 1 models
Leaderboard
Tracked models ranked by Score (higher is better).
Configuration: Agents' Last Exam v1.1; Antigravity-CLI 1.4.8; 122-task; 5-hour window; safety filters enabled; averaged rubric score
Notes: Google DeepMind model card results table (as of November 2026): Agents' Last Exam v1.1 75.8%. Evals methodology records the self-computed harness above. Timestamp is retrieval date.
Configuration: gpt-6.1-terra (Max)
Notes: OpenAI launch post chart 'Agents’ Last Exam v1' (Max 0.600; Low 0.559, Medium 0.576, High 0.572, Xhigh 0.585). Chart is labelled v1; the official ALE board shows v1.1 and did not list GPT-6.1 Terra at capture. Retrieved 2026-10-26.
Notes: Official GPT-5.6 GA launch benchmark row.
Notes: Moonshot self-reported launch table, row 'Agents' Last Exam (ALL)', column 'Kimi K3.1 Pro (max)'. Settings: reasoning effort max, temperature 1.0, top-p 0.95 (blog footnote).
Configuration: Gemini 4 Argon
Notes: Blog table + methodology page: binary pass rate, ALE-Claw default harness, 5-hour window, safety filters on. Retrieved 2026-10-17.
Configuration: minimax-m3-1-flash (Max)
Notes: Official ALE leaderboard (benchmark owner; headless-Chrome DOM): row 'Claude Code | minimax-m3-1-flash | Max | 29.6% | 55.0%' (Pass Rate 29.6%, Score 55.0%). Matches MiniMax blog (29.6, 'official Overall setting'). A Codex-harness row for the same model shows 27.0% (not packed; highest official entry used, same as blog). Retrieved 2026-10-19.
How to read this benchmark
This benchmark scores models where higher is better. Use scores for directional filtering and shortlisting, not universal quality ranking; then validate pricing, context window, provider availability, and fit for your workload.
Trust this score when
- There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
- The model list covers the same version family you can actually deploy today.
- Top candidates overlap with your required routing and feature requirements.
Be cautious when
- There is only one benchmark snapshot or the dataset appears stale.
- The benchmark metric direction is opposite of your decision objective.
- The score difference between options is narrow and likely within implementation variance.
Related benchmarks
Last reviewed: Jul 9, 2026