Skip to content
Weights Not Walls
Menu

METR Time Horizons

Reported

No description yet: we only describe an eval set when we can do it correctly.

Kind
Published by a third party's leaderboard, collected by Epoch AI. The leaderboard is named and linked.
Link
https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
Score shown
Epoch's column "average_score", as a percentage, higher is better; random guessing scores about 0.0%. The source gives no error bars.
Results
49 in all: 9 for open-weight models, 40 for closed ones. 8 linked to a model page here.

Best open-weight result: moonshotai/Kimi-K2-Thinking, 59.2%, number 24 of 49 results.

METR Time Horizons results
#ModelWeightsSettingScoreDateSource
1claude-mythos-preview-earlyAnthropiccloseddefault85.2%not datedmetr.org
2claude-opus-4-6_unknownAnthropicclosedeffort unknown78.9%not datedvia Epoch AI
3gemini-3.1-pro-previewGoogle DeepMindcloseddefault77.0%not datedmetr.org
4gpt-5.2-2025-12-11_highOpenAIclosedeffort high75.3%not datedvia Epoch AI
5claude-opus-4-5-20251101_16KAnthropiccloseddefault75.0%not datedMETR - Measuring AI Ability to Complete Long Tasks
6gpt-5.3-codexOpenAIcloseddefault74.5%not datedvia Epoch AI
7gpt-5.4-2026-03-05_xhighOpenAIclosedeffort xhigh74.3%not datedmetr.org
8claude-opus-4-5-20251101Anthropiccloseddefault73.0%not datedvia Epoch AI
9gemini-3-pro-previewGoogle DeepMindcloseddefault71.0%not datedvia Epoch AI
10gpt-5.1-codex-maxOpenAIcloseddefault70.8%not datedmetr.org
11gpt-5-2025-08-07_mediumOpenAIclosedeffort medium69.6%not datedMETR - Measuring AI Ability to Complete Long Tasks
12gpt-5-2025-08-07_highOpenAIclosedeffort high69.4%not datedvia Epoch AI
13claude-sonnet-4-5-20250929_16KAnthropiccloseddefault67.4%not datedMETR - Measuring AI Ability to Complete Long Tasks
14claude-opus-4-1-20250805_16KAnthropiccloseddefault66.8%not datedMETR - Measuring AI Ability to Complete Long Tasks
15grok-4-0709xAIcloseddefault66.6%not datedMETR - Measuring AI Ability to Complete Long Tasks
16o3-2025-04-16_mediumOpenAIclosedeffort medium65.4%not datedMETR - Measuring AI Ability to Complete Long Tasks
17claude-opus-4-20250514_16KAnthropiccloseddefault63.9%not datedMETR - Measuring AI Ability to Complete Long Tasks
18o4-mini-2025-04-16_mediumOpenAIclosedeffort medium63.9%not datedMETR - Measuring AI Ability to Complete Long Tasks
19o3-2025-04-16_unknownOpenAIclosedeffort unknown63.6%not datedvia Epoch AI
20claude-sonnet-4-20250514_16KAnthropiccloseddefault62.0%not datedMETR - Measuring AI Ability to Complete Long Tasks
21claude-opus-4-1-20250805Anthropiccloseddefault61.6%not datedvia Epoch AI
22claude-opus-4-20250514_unknownAnthropicclosedeffort unknown61.5%not datedvia Epoch AI
23claude-3-7-sonnet-20250219_16KAnthropiccloseddefault60.0%not datedMETR - Measuring AI Ability to Complete Long Tasks
24moonshotai/Kimi-K2-ThinkingEpoch: kimi-k2-thinking, Moonshotopendefault59.2%not datedMETR - Measuring AI Ability to Complete Long Tasks
25openai/gpt-oss-120bEpoch: gpt-oss-120b, OpenAIopendefault56.6%not datedMETR - Measuring AI Ability to Complete Long Tasks
26claude-3-7-sonnet-20250219Anthropiccloseddefault55.8%not datedvia Epoch AI
27gemini-2.5-pro-preview-06-05Google DeepMindcloseddefault55.4%not datedMETR - Measuring AI Ability to Complete Long Tasks
28deepseek-ai/DeepSeek-R1-0528Epoch: DeepSeek-R1-0528, DeepSeekopendefault53.8%not datedMETR - Measuring AI Ability to Complete Long Tasks
29deepseek-ai/DeepSeek-R1Epoch: DeepSeek-R1, DeepSeekopendefault51.9%not datedMETR - Measuring AI Ability to Complete Long Tasks
30o1-2024-12-17_mediumOpenAIclosedeffort medium51.1%not datedmetr.org
31deepseek-ai/DeepSeek-V3-0324Epoch: DeepSeek-V3-0324, DeepSeekopendefault49.6%not datedMETR - Measuring AI Ability to Complete Long Tasks
32deepseek-ai/DeepSeek-V3Epoch: DeepSeek-V3, DeepSeekopendefault47.4%not datedMETR - Measuring AI Ability to Complete Long Tasks
33claude-3-5-sonnet-20241022Anthropiccloseddefault45.2%not datedmetr.org
34o1-preview-2024-09-12OpenAIcloseddefault45.1%not datedmetr.org
35gpt-4o-2024-05-13OpenAIcloseddefault40.8%not datedMETR - Measuring AI Ability to Complete Long Tasks
36claude-3-5-sonnet-20240620Anthropiccloseddefault40.2%not datedmetr.org
37gpt-4-turbo-2024-04-09OpenAIcloseddefault36.7%not datedMETR - Measuring AI Ability to Complete Long Tasks
38gpt-4-0314OpenAIcloseddefault36.1%not datedMETR - Measuring AI Ability to Complete Long Tasks
39Qwen/Qwen2.5-72B-InstructEpoch: qwen2.5-72b-instruct, Alibabaopendefault35.8%not datedMETR - Measuring AI Ability to Complete Long Tasks
40gpt-4-0125-previewOpenAIcloseddefault35.2%not datedMETR - Measuring AI Ability to Complete Long Tasks
41gpt-4o-2024-05-13OpenAIcloseddefault33.8%not datedvia Epoch AI
42Qwen/Qwen2-72B-InstructEpoch: qwen2-72b-instruct, Alibabaopendefault29.9%not datedMETR - Measuring AI Ability to Complete Long Tasks
43claude-3-opus-20240229Anthropiccloseddefault29.5%not datedmetr.org
44gpt-4-0314OpenAIcloseddefault29.3%not datedvia Epoch AI
45gpt-4-1106-previewOpenAIcloseddefault28.9%not datedmetr.org
46gpt-4-turbo-2024-04-09OpenAIcloseddefault27.2%not datedvia Epoch AI
47gpt-3.5-turbo-instructOpenAIcloseddefault21.5%not datedmetr.org
48davinci-002OpenAIcloseddefault16.2%not datedmetr.org
49gpt2-xlOpenAIno page matchopendefault10.1%not datedmetr.org

Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method