Agents' Last Exam leaderboard
As of November 8, 2026, GPT-6 Astra (OpenAI) leads Agents' Last Exam at 59.3%, ahead of GPT-6 Sol at 56.4% and GPT-5.6 Sol at 53.6%. 20 systems ranked. Every score here is published by the model's own lab; each row links to the page it comes from.
What is Agents' Last Exam? Agents' Last Exam measures whether AI agents can finish long, economically valuable professional workflows with a checkable result, across 55 sub-fields in 13 industry clusters. Made by Berkeley RDI with 300+ industry experts; leaderboard hosted by Snorkel AI, June 2026.
Top models on Agents' Last Exam, November 2026
OpenAI's reported metric for its own models.
| # | Model | Score | $ / task | Price / 1M | Context | Source |
|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra OpenAI | max effort | $6.23 | $10 / $50 | 1.05M | Vendor-reported OpenAI |
| 2 | GPT-6 Sol OpenAI | max effort | $2.93 | $2 / $10 | 1.05M | Vendor-reported OpenAI |
| 3 | GPT-5.6 Sol OpenAI | xhigh effort | $5.08 | $4 / $20 | 1.05M | Vendor-reported OpenAI |
| 4 | GPT-6 Luna OpenAI | max effort | $0.15 | $0.1 / $0.5 | 1.05M | Vendor-reported OpenAI |
| 5 | GPT-5.6 Luna OpenAI | max effort | $2.57 | $0.2 / $1.2 | 1.05M | Vendor-reported OpenAI |
Agents' Last Exam under other settings
Labs report Agents' Last Exam under different settings, and a score from one setting is not comparable with a score from another. Each setting has its own table, ranked on its own.
Strict pass rate
Share of tasks fully completed.
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | Gemini 4 Argon Google DeepMind | pass rate | $2 / $10 | 1M | Vendor-reported Google DeepMind |
| 2 | Qwen3.8 Max Alibaba (Qwen) | pass rate | $2 / $6 | 1M | Vendor-reported Alibaba (Qwen) |
| 3 | Gemini 3.7 Flash Deprecated Google DeepMind | pass rate | $0.75 / $3.75 | 1M | Vendor-reported Google DeepMind |
| 4 | Gemini 3.6 Flash Google | pass rate | $0.75 / $3.75 | 1M | Vendor-reported Google DeepMind |
| 5 | Qwen3.7 Max Alibaba (Qwen) | pass rate | $2.5 / $7.5 | 1M | Vendor-reported Alibaba (Qwen) |
CLI split
The Linux-only subset for command-line agents.
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | Kimi K3.1 Pro Moonshot (Kimi) | max effort, CLI split | $2 / $10 | 1M | Vendor-reported Moonshot AI |
| 2 | DeepSeek V4.1 Flash DeepSeek · listed as “DeepSeek V4.1 Flash (1017)” | as reported | $0.3 / $1.2 | 1M | Vendor-reported DeepSeek |
| 3 | GLM-5.3 Zhipu (Z.ai / GLM) | CLI split | $1.4 / $4.4 | 1M | Vendor-reported Z.ai (Zhipu) |
| 4 | DeepSeek V4 Pro DeepSeek | pass@1 | $1.32 / $3.96 | 1M | Vendor-reported DeepSeek |
| 5 | DeepSeek V4 Flash Deprecated DeepSeek | pass@1 | $0.44 / $1.32 | 1M | Vendor-reported DeepSeek |
| 6 | GLM-5.2 Zhipu (Z.ai / GLM) | CLI split | $1.4 / $4.4 | 1M | Vendor-reported Z.ai (Zhipu) |
ALL split
The full benchmark, including Windows tasks.
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | Kimi K3.1 Pro Moonshot (Kimi) | max effort, ALL split | $2 / $10 | 1M | Vendor-reported Moonshot AI |
Split not stated
| # | Model | Score | Price / 1M | Context | Source |
|---|---|---|---|---|---|
| 1 | MiMo-V2.6-Pro Xiaomi MiMo | split not stated | $0.435 / $0.87 | 1M | Vendor-reported Xiaomi |
| 2 | MiniMax M3.1 Flash MiniMax | split not stated | $0.15 / $0.6 | 1M | Vendor-reported MiniMax |
| 3 | Kimi K3 Moonshot (Kimi) | max effort, split not stated | $3 / $15 | 1M | Vendor-reported Moonshot AI |
| 4 | MiMo-V2.6-Flash Xiaomi MiMo | split not stated | $0.14 / $0.28 | 1M | Vendor-reported Xiaomi |
Agents' Last Exam: score against price
What more reasoning buys
Some publishers ran the same model at several reasoning-effort levels and printed the score and the cost per task for each. Higher effort spends more tokens per task; this is what it bought on Agents' Last Exam.
| Model | Score by effort level ($ per task) | Source |
|---|---|---|
| GPT-6 Astra | low 53.4% $3.07 medium 57.6% $4.10 high 57.8% $4.64 xhigh 58.3% $5.40 max 59.3% $6.23 | OpenAI |
| GPT-6 Sol | low 48.7% $0.86 medium 53.1% $1.27 high 52.6% $1.53 xhigh 55.4% $1.67 max 56.4% $2.93 | OpenAI |
| GPT-5.6 Sol | low 45.1% $1.63 medium 52.1% $3.41 high 52.4% $3.79 xhigh 53.6% $5.08 max 52.8% $7.13 | OpenAI |
| GPT-6 Luna | low 36.3% $0.025 medium 46.8% $0.11 high 43.6% $0.11 xhigh 47.9% $0.11 max 50.9% $0.15 | OpenAI |
| GPT-5.6 Luna | low 31% $0.18 medium 38.5% $0.38 high 46.1% $0.84 xhigh 49.4% $1.55 max 50.4% $2.57 | OpenAI |
Can you trust Agents' Last Exam scores?
No independent leaderboard publishes Agents' Last Exam results for these models yet, so every number on this page is the lab's own report of its own model. Labs choose the harness, the effort level and the subset they report, and they rarely publish failures. Read the setting under each score before comparing two rows.
- Labs quote four different numbers for this benchmark: the v1 score, the strict pass rate, the full ALL split and the Linux-only CLI split. They are separate tables here because they are not comparable.
- The official leaderboard is scored on held-out private tasks, and the public subset is rotated over time, so the version matters.
- Runs are long and expensive, so some published scores rest on only a few runs.
How Agents' Last Exam works
The task
Each task is a real professional workflow written with industry experts. A computer-use or command-line agent works through it in a sandboxed virtual machine, and its output is checked against a hidden reference.
The score
Graders compare the result with the hidden reference. Results are published as a strict pass rate (the share of tasks fully completed) and as a mean score with partial credit, so labs quote different numbers.
Test set
1,500+ tasks, toward a 5,000-task target, as published by the authors.
Made by
Berkeley RDI with 300+ industry experts; leaderboard hosted by Snorkel AI, first published June 2026.
Official sources
- agents-last-exam.org
- Paper: arxiv.org/abs/2606.05405
- Definition read on 2026-11-08
Sources and freshness
22 published scores from 11 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.
- Vendor-reported OpenAI
- Vendor-reported DeepSeek
- Vendor-reported Alibaba (Qwen)
- Vendor-reported Google DeepMind
- Vendor-reported Moonshot AI
- Vendor-reported Xiaomi
- Vendor-reported Z.ai (Zhipu)
- Vendor-reported DeepSeek
- Vendor-reported Google DeepMind
- Vendor-reported MiniMax
- Vendor-reported Moonshot AI
Agents' Last Exam FAQ
Answers derived from the 22 published scores on this page and the current price index.