Humanity's Last Exam
2,500 expert-written questions across more than 100 academic subjects that need deep, specialised knowledge.
- Unit
- % correct
- Results
- 56 results40 Tokscale models matched, 46 in the source
- Last update
- Oct 7, 2026Newest model release
- Retrieved
- Oct 21, 2026Source data through Oct 15, 2026
Conditions. Externally reported results collected by Epoch AI. The reasoning effort is often not reported.
Attribution. Humanity's Last Exam results, collected by Epoch AI, ‘Capabilities & benchmarking’, epoch.ai/benchmarks (CC BY 4.0)
Results
One row per model: its best reported result. The rank is the position in this list, by score; the source's own ranking can differ.
Tokscale usage counts accounts that used the model in public Tokscale submissions, Oct 14 – Nov 12, 2026. It is shown next to the score and does not explain it.
| Rank | Model | Score (% correct) | Conditions | Date | Tokscale usage (30 days) | Source link |
|---|---|---|---|---|---|---|
| 1 | GPT-6 AstraListed as GPT 6 Astra | 54.8%Standard error ±1.9 pts | Not reported | Sep 3, 2026Model released | 562active users | Source for GPT-6 Astra (opens in a new tab) |
| 2 | Claude Fable 5.1Listed as Fable 5.1 (xhigh) | 46.5%Standard error ±2.0 pts | XHigh effort | Sep 1, 2026Model released | 361active users | Source for Claude Fable 5.1 (opens in a new tab) |
| 3 | Gemini 3.1 Pro PreviewListed as Gemini 3.1 Pro Preview (thinking high) | 46.4%Standard error ±2.0 pts | Not reported | Feb 19, 2026Model released | 5active users | Source for Gemini 3.1 Pro Preview (opens in a new tab) |
| 4 | Claude Sonnet 5.5Listed as Sonnet 5.5 (xhigh) | 46.1%Standard error ±1.9 pts | XHigh effort | Sep 28, 2026Model released | 555active users | Source for Claude Sonnet 5.5 (opens in a new tab) |
| 5 | 44.5%Standard error ±2.0 pts | Not reported | Sep 2, 2026Model released | 66active users | Source for Gemini 3.8 Flash (opens in a new tab) | |
| 6 | 44.3%Standard error ±1.9 pts | Not reported | Mar 5, 2026Model released | No Tokscale usage in the last 30 days | Source for GPT-5.4 Pro (opens in a new tab) | |
| 7 | Muse SparkNot matched to a Tokscale model | 40.6%Standard error ±1.9 pts | Not reported | Apr 8, 2026Model released | Not matched to a Tokscale model | Source for Muse Spark (opens in a new tab) |
| 8 | Gemini 3 Pro PreviewListed as gemini-3-pro-preview | 37.5%Standard error ±1.9 pts | Not reported | Nov 18, 2025Model released | 1active users | Source for Gemini 3 Pro Preview (opens in a new tab) |
| 9 | GPT-5.4Listed as gpt-5.4-2026-03-05 (xhigh thinking) | 36.2%Standard error ±1.9 pts | XHigh effort | Mar 5, 2026Model released | 1active users | Source for GPT-5.4 (opens in a new tab) |
| 10 | 36.2%Standard error ±1.9 pts | Not reported | Apr 16, 2026Model released | 27active users | Source for Claude Opus 4.7 (opens in a new tab) | |
| 11 | Claude Opus 4.6Listed as claude-opus-4-6-thinking-max | 34.4%Standard error ±1.9 pts | Max effort | Feb 5, 2026Model released | 52active users | Source for Claude Opus 4.6 (opens in a new tab) |
| 12 | GPT-5 ProListed as gpt-5-pro-2025-10-06 | 31.6%Standard error ±1.8 pts | Not reported | Oct 7, 2025Model released | No Tokscale usage in the last 30 days | Source for GPT-5 Pro (opens in a new tab) |
| 13 | 30.1%Standard error ±1.8 pts | Not reported | Oct 7, 2026Model released | 477active users | Source for Claude Haiku 5.5 (opens in a new tab) | |
| 14 | GPT-5.2Listed as gpt-5.2-2025-12-11 | 27.8%Standard error ±1.8 pts | Not reported | Dec 11, 2025Model released | No Tokscale usage in the last 30 days | Source for GPT-5.2 (opens in a new tab) |
| 15 | GPT-5 2025 08 07Listed as gpt-5-2025-08-07 | 25.3%Standard error ±1.7 pts | High effort | Aug 7, 2025Model released | No Tokscale usage in the last 30 days | Source for GPT-5 2025 08 07 (opens in a new tab) |
| 16 | Claude Opus 4.5Listed as claude-opus-4-5-20251101-thinking | 25.2%Standard error ±1.7 pts | Not reported | Nov 24, 2025Model released | 4active users | Source for Claude Opus 4.5 (opens in a new tab) |
| 17 | Kimi K2.5Listed as kimi-k2.5 | 24.4%Standard error ±1.8 pts | Not reported | Jan 27, 2026Model released | No Tokscale usage in the last 30 days | Source for Kimi K2.5 (opens in a new tab) |
| 18 | GPT-5.1Listed as gpt-5.1-thinking | 23.7%Standard error ±1.7 pts | Not reported | Nov 13, 2025Model released | No Tokscale usage in the last 30 days | Source for GPT-5.1 (opens in a new tab) |
| 19 | Gemini 2.5 Pro Preview 06 05Listed as gemini-2.5-pro-preview-06-05 | 21.6%Standard error ±1.6 pts | Not reported | Jun 5, 2025Model released | No Tokscale usage in the last 30 days | Source for Gemini 2.5 Pro Preview 06 05 (opens in a new tab) |
| 20 | o3Listed as o3 (high) (April 2025) | 20.3%Standard error ±1.6 pts | High effort | Apr 16, 2025Model released | No Tokscale usage in the last 30 days | Source for o3 (opens in a new tab) |
| 21 | GPT-5 MiniListed as gpt-5-mini-2025-08-07 | 19.4%Standard error ±1.6 pts | Not reported | Aug 7, 2025Model released | 1active users | Source for GPT-5 Mini (opens in a new tab) |
| 22 | Gemini 2.5 Pro Exp 03 25Listed as Gemini 2.5 Pro Experimental (March 2025) | 18.2%Standard error ±1.5 pts | Not reported | Mar 25, 2025Model released | No Tokscale usage in the last 30 days | Source for Gemini 2.5 Pro Exp 03 25 (opens in a new tab) |
| 23 | o4-miniListed as o4-mini (high) (April 2025) | 18.1%Standard error ±1.5 pts | High effort | Apr 16, 2025Model released | No Tokscale usage in the last 30 days | Source for o4-mini (opens in a new tab) |
| 24 | Gemini 2.5 Pro Preview 05 06Listed as Gemini 2.5 Pro Preview (May 06 2025) | 17.8%Standard error ±1.5 pts | Not reported | May 6, 2025Model released | No Tokscale usage in the last 30 days | Source for Gemini 2.5 Pro Preview 05 06 (opens in a new tab) |
| 25 | Claude Sonnet 4.5Listed as claude-sonnet-4-5-20250929-thinking | 13.7%Standard error ±1.4 pts | Not reported | Sep 29, 2025Model released | 23active users | Source for Claude Sonnet 4.5 (opens in a new tab) |
| 26 | Gemini 2.5 Flash (April 2025)Not matched to a Tokscale model | 12.1%Standard error ±1.3 pts | Not reported | Apr 17, 2025Model released | Not matched to a Tokscale model | Source for Gemini 2.5 Flash (April 2025) (opens in a new tab) |
| 27 | Claude Opus 4.1Listed as claude-opus-4-1-20250805-thinking | 11.5%Standard error ±1.3 pts | Not reported | Aug 5, 2025Model released | No Tokscale usage in the last 30 days | Source for Claude Opus 4.1 (opens in a new tab) |
| 28 | Gemini 2.5 Flash Preview 05 20Listed as Gemini 2.5 Flash Preview (May 2025) | 11.0%Standard error ±1.2 pts | Not reported | May 20, 2025Model released | No Tokscale usage in the last 30 days | Source for Gemini 2.5 Flash Preview 05 20 (opens in a new tab) |
| 29 | Claude Opus 4Listed as Claude Opus 4 (Thinking) | 10.7%Standard error ±1.2 pts | Not reported | May 22, 2025Model released | No Tokscale usage in the last 30 days | Source for Claude Opus 4 (opens in a new tab) |
| 30 | Gemini 3.1 Flash LiteListed as gemini-3.1-flash-lite-preview | 8.6%Standard error ±1.1 pts | Not reported | Mar 3, 2026Model released | 5active users | Source for Gemini 3.1 Flash Lite (opens in a new tab) |
| 31 | GLM-4.5Listed as glm-4p5 | 8.3%Standard error ±1.1 pts | Not reported | Aug 3, 2025Model released | No Tokscale usage in the last 30 days | Source for GLM-4.5 (opens in a new tab) |
| 32 | GLM-4.5 AirListed as glm-4p5-air | 8.1%Standard error ±1.1 pts | Not reported | Jul 20, 2025Model released | No Tokscale usage in the last 30 days | Source for GLM-4.5 Air (opens in a new tab) |
| 32 | o1-proListed as o1 Pro | 8.1%Standard error ±1.1 pts | Not reported | Mar 19, 2025Model released | No Tokscale usage in the last 30 days | Source for o1-pro (opens in a new tab) |
| 34 | Claude 3.7 SonnetListed as Claude 3.7 Sonnet (Thinking) | 8.0%Standard error ±1.1 pts | Not reported | Feb 24, 2025Model released | No Tokscale usage in the last 30 days | Source for Claude 3.7 Sonnet (opens in a new tab) |
| 35 | o1Listed as o1 (December 2024) | 8.0%Standard error ±1.1 pts | Not reported | Dec 17, 2024Model released | No Tokscale usage in the last 30 days | Source for o1 (opens in a new tab) |
| 36 | Claude Sonnet 4Listed as Claude Sonnet 4 (Thinking) | 7.8%Standard error ±1.1 pts | Not reported | May 22, 2025Model released | No Tokscale usage in the last 30 days | Source for Claude Sonnet 4 (opens in a new tab) |
| 37 | Gemini 2.0 Flash Thinking (January 2025)Not matched to a Tokscale model | 6.6%Standard error ±1.0 pts | Not reported | Jan 21, 2025Model released | Not matched to a Tokscale model | Source for Gemini 2.0 Flash Thinking (January 2025) (opens in a new tab) |
| 38 | Llama 4 Maverick 17B 128e InstructListed as Llama 4 Maverick | 5.7%Standard error ±0.9 pts | Not reported | Apr 5, 2025Model released | No Tokscale usage in the last 30 days | Source for Llama 4 Maverick 17B 128e Instruct (opens in a new tab) |
| 39 | GPT 4.5 PreviewNot matched to a Tokscale model | 5.4%Standard error ±0.9 pts | Not reported | Feb 27, 2025Model released | Not matched to a Tokscale model | Source for GPT 4.5 Preview (opens in a new tab) |
| 40 | GPT-4.1 2025 04 14Listed as GPT-4.1 | 5.4%Standard error ±0.9 pts | Not reported | Apr 14, 2025Model released | 5active users | Source for GPT-4.1 2025 04 14 (opens in a new tab) |
| 41 | Gemini-1.5-Pro-002Not matched to a Tokscale model | 4.6%Standard error ±0.8 pts | Not reported | Sep 24, 2024Model released | Not matched to a Tokscale model | Source for Gemini-1.5-Pro-002 (opens in a new tab) |
| 42 | Mistral Medium 2505Listed as Mistral Medium 3 | 4.5%Standard error ±0.8 pts | Not reported | May 7, 2025Model released | No Tokscale usage in the last 30 days | Source for Mistral Medium 2505 (opens in a new tab) |
| 43 | 4.4%Standard error ±0.8 pts | Not reported | Dec 3, 2024Model released | No Tokscale usage in the last 30 days | Source for Nova Pro (opens in a new tab) | |
| 44 | Claude 3.5 Sonnet (October 2024)Not matched to a Tokscale model | 4.1%Standard error ±0.8 pts | Not reported | Oct 22, 2024Model released | Not matched to a Tokscale model | Source for Claude 3.5 Sonnet (October 2024) (opens in a new tab) |
| 45 | 3.6%Standard error ±0.7 pts | Not reported | Dec 3, 2024Model released | No Tokscale usage in the last 30 days | Source for Nova Lite (opens in a new tab) | |
| 46 | GPT-4o 2024 11 20Listed as GPT-4o (November 2024) | 2.7%Standard error ±0.6 pts | Not reported | Nov 20, 2024Model released | 1active users | Source for GPT-4o 2024 11 20 (opens in a new tab) |
Showing 46 of 46 models
Score by release date
Each dot is one model's best result on Humanity's Last Exam, placed at the model's release date. The dashed line is the best score among models released up to that date.
- Used on Tokscale in the last 30 days
- No Tokscale usage in the last 30 days
- Best score so far
GPT-6 Astra: 54.8% · released Sep 3, 2026 · 562 active users on Tokscale in the last 30 days
40 models, released Nov 20, 2024 to Oct 7, 2026. Highest: GPT-6 Astra, 54.8%, released Sep 3, 2026. Use the left and right arrow keys to read each model in release order.0%20%40%60%
Show data
| Model | Released | Score | Tokscale users, 30 days |
|---|---|---|---|
| Claude Haiku 5.5 | Oct 7, 2026 | 30.1% | 477 |
| Claude Sonnet 5.5 | Sep 28, 2026 | 46.1%XHigh effort | 555 |
| GPT-6 Astra | Sep 3, 2026 | 54.8% | 562 |
| Gemini 3.8 Flash | Sep 2, 2026 | 44.5% | 66 |
| Claude Fable 5.1 | Sep 1, 2026 | 46.5%XHigh effort | 361 |
| Claude Opus 4.7 | Apr 16, 2026 | 36.2% | 27 |
| GPT-5.4 | Mar 5, 2026 | 36.2%XHigh effort | 1 |
| GPT-5.4 Pro | Mar 5, 2026 | 44.3% | No Tokscale usage in the last 30 days |
| Gemini 3.1 Flash Lite | Mar 3, 2026 | 8.6% | 5 |
| Gemini 3.1 Pro Preview | Feb 19, 2026 | 46.4% | 5 |
| Claude Opus 4.6 | Feb 5, 2026 | 34.4%Max effort | 52 |
| Kimi K2.5 | Jan 27, 2026 | 24.4% | No Tokscale usage in the last 30 days |
| GPT-5.2 | Dec 11, 2025 | 27.8% | No Tokscale usage in the last 30 days |
| Claude Opus 4.5 | Nov 24, 2025 | 25.2% | 4 |
| Gemini 3 Pro Preview | Nov 18, 2025 | 37.5% | 1 |
| GPT-5.1 | Nov 13, 2025 | 23.7% | No Tokscale usage in the last 30 days |
| GPT-5 Pro | Oct 7, 2025 | 31.6% | No Tokscale usage in the last 30 days |
| Claude Sonnet 4.5 | Sep 29, 2025 | 13.7% | 23 |
| GPT-5 Mini | Aug 7, 2025 | 19.4% | 1 |
| GPT-5 2025 08 07 | Aug 7, 2025 | 25.3%High effort | No Tokscale usage in the last 30 days |
| Claude Opus 4.1 | Aug 5, 2025 | 11.5% | No Tokscale usage in the last 30 days |
| GLM-4.5 | Aug 3, 2025 | 8.3% | No Tokscale usage in the last 30 days |
| GLM-4.5 Air | Jul 20, 2025 | 8.1% | No Tokscale usage in the last 30 days |
| Gemini 2.5 Pro Preview 06 05 | Jun 5, 2025 | 21.6% | No Tokscale usage in the last 30 days |
| Claude Sonnet 4 | May 22, 2025 | 7.8% | No Tokscale usage in the last 30 days |
| Claude Opus 4 | May 22, 2025 | 10.7% | No Tokscale usage in the last 30 days |
| Gemini 2.5 Flash Preview 05 20 | May 20, 2025 | 11.0% | No Tokscale usage in the last 30 days |
| Mistral Medium 2505 | May 7, 2025 | 4.5% | No Tokscale usage in the last 30 days |
| Gemini 2.5 Pro Preview 05 06 | May 6, 2025 | 17.8% | No Tokscale usage in the last 30 days |
| o4-mini | Apr 16, 2025 | 18.1%High effort | No Tokscale usage in the last 30 days |
| o3 | Apr 16, 2025 | 20.3%High effort | No Tokscale usage in the last 30 days |
| GPT-4.1 2025 04 14 | Apr 14, 2025 | 5.4% | 5 |
| Llama 4 Maverick 17B 128e Instruct | Apr 5, 2025 | 5.7% | No Tokscale usage in the last 30 days |
| Gemini 2.5 Pro Exp 03 25 | Mar 25, 2025 | 18.2% | No Tokscale usage in the last 30 days |
| o1-pro | Mar 19, 2025 | 8.1% | No Tokscale usage in the last 30 days |
| Claude 3.7 Sonnet | Feb 24, 2025 | 8.0% | No Tokscale usage in the last 30 days |
| o1 | Dec 17, 2024 | 8.0% | No Tokscale usage in the last 30 days |
| Nova Lite | Dec 3, 2024 | 3.6% | No Tokscale usage in the last 30 days |
| Nova Pro | Dec 3, 2024 | 4.4% | No Tokscale usage in the last 30 days |
| GPT-4o 2024 11 20 | Nov 20, 2024 | 2.7% | 1 |
Method and matching
How results are run
- Externally reported results collected by Epoch AI. The reasoning effort is often not reported.
Matching to Tokscale models
- 50 of 56 results are matched to 40 Tokscale models: 24 by exact id, 26 by normalized name.
- No match is marked as reviewed by hand yet.
- 6 results from 6 models are not matched. They stay in the table without a Tokscale link.
- Reasoning effort, thinking budget and agent are conditions of a result. They never change which model a result belongs to.
License and attribution
- CC BY 4.0 (opens in a new tab), retrieved Oct 21, 2026.
- Humanity's Last Exam results, collected by Epoch AI, ‘Capabilities & benchmarking’, epoch.ai/benchmarks (CC BY 4.0)
Sources and attribution
Each score keeps its own source, unit and run conditions. Tokscale does not combine benchmarks into one score, and a missing score means the model was not evaluated, not that it scored zero.