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.
13 models · 17 results · data as of 2026-11-06 · snorkel.ai ↗
Leaderboard
Gemini 4 Carbon high
All results
17 rows| # | Model | Effort | Agents' Last Exam | Harness |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high | 77.4% | ALE-Claw |
| 2 | GPT-6 Astra | max | 73.1% | Codex |
| 3 | GPT-6 Astra | xhigh | 71.7% | Codex |
| 4 | Opus 5.5 | max | 71.6% | Claude Code |
| 5 | GPT-6 Astra | medium | 71.2% | Codex |
| 6 | GPT-6 Astra | high | 71.0% | Codex |
| 7 | Hy4 | high | 70.3% | Claude Code |
| 8 | GPT-6.1 Terra | max | 69.8% | Codex |
| 9 | Gemini 4 Argon | high | 69.0% | ALE-Claw |
| 10 | Muse Wonder (Max) | high | 68.4% | Muse Code |
| 11 | GPT-6 Astra | low | 67.9% | Codex |
| 12 | DeepSeek V4.1 Flash 1017 | high | 65.4% | OpenClaw |
| 13 | Kimi K3 | max | 63.0% | Kimi Code |
| 14 | MiniMax M3.1 | max | 58.9% | Codex |
| 15 | GLM 5.3 | max | 54.4% | ZCode |
| 16 | GPT-5.6 Sol | max | 48.4% | Codex |
| 17 | GPT-5.6 Luna | max | 26.5% | Codex |