Benchmarks · agentic
Agents' Last Exam
Snorkel AI's benchmark of long-horizon, economically valuable professional workflows across 55 sub-industries, run in each model's agent harness with verifiable outcomes. Score is the share of rubric points earned (the pass rate, all checks met, is lower). Best harness per model and effort.
10 models · 14 results · data as of 2026-10-29 · snorkel.ai ↗
Leaderboard
Gemini 4 Carbon high
All results
14 rows| # | Model | Effort | Agents' Last Exam | Harness |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high | 77.4%launch | |
| 2 | GPT-6 Astra | max | 73.1% | Codex |
| 3 | GPT-6 Astra | xhigh | 71.7% | Codex |
| 4 | GPT-6 Astra | medium | 71.2% | Codex |
| 5 | GPT-6 Astra | high | 71.0% | Codex |
| 6 | GPT-6 Astra | low | 67.9% | Codex |
| 7 | Opus 5.5 | max | 67.5% | Claude Code |
| 8 | Muse Wonder (Max) | high | 66.7% | Muse Code |
| 9 | Gemini 4 Argon | high | 63.8% | ALE-Claw |
| 10 | Kimi K3 | max | 62.5% | Kimi Code |
| 11 | MiniMax M3.1 | max | 58.2% | Codex |
| 12 | GLM 5.3 | max | 53.2% | ZCode |
| 13 | GPT-5.6 Sol | max | 48.4% | Codex |
| 14 | GPT-5.6 Luna | max | 26.5% | Codex |