benchmarks.site

Public leaderboard · Agents, tool use & computer use

METR Time Horizons

Measures how long a task (in human working time) an AI can complete about half the time; shows how fast agents are improving.

Leaderboard

Top 3 right now

  1. 1stclaude_mythos_preview_early_inspect1044.8 min (508.9-3304.3)
  2. 2ndclaude_opus_4_6_inspect718.8 min (316.7-3633.8)
  3. 3rdgemini_3_1_pro384.1 min (233.5-694.8)

What it measures

Estimates the length of software and research tasks, measured in human working time, that an AI agent can complete with 50% reliability.

How to read the numbers

50% time horizon, minutes (95% CI), as in METR data file (estimate field; unit assumed minutes) — higher is better.

Results

25 rows · higher is better (50% time horizon (min, CI))CSV
1claude_mythos_preview_early_inspect1044.8 min (508.9-3304.3)2026-04-07
2claude_opus_4_6_inspect718.8 min (316.7-3633.8)2026-02-05
3gemini_3_1_pro384.1 min (233.5-694.8)2026-02-19
4gpt_5_2352.2 min (198.1-815.2)2025-12-11
5gpt_5_3_codex349.5 min (194.9-816.4)2026-02-05
6gpt_5_4341.7 min (186.6-768.8)2026-03-05
7claude_opus_4_5_inspect293.0 min (161.7-623.7)2025-11-24
8gemini_3_pro224.3 min (139.6-379.2)2025-11-18
9gpt_5_1_codex_max_inspect223.7 min (134.3-396.2)2025-11-19
10gpt_5_2025_08_07_inspect203.0 min (112.6-405.6)2025-08-07
11o3_inspect119.7 min (74.6-190.9)2025-04-16
12claude_4_1_opus_inspect100.5 min (59.3-159.5)2025-08-05
13claude_4_opus_inspect100.4 min (60.0-163.5)2025-05-22
14claude_3_7_sonnet_inspect60.4 min (33.0-104.2)2025-02-24
15o1_inspect38.8 min (21.2-65.0)2024-12-05
16claude_3_5_sonnet_20241022_inspect20.5 min (10.1-40.8)2024-10-22
17o1_preview20.3 min (11.7-33.4)2024-09-12
18claude_3_5_sonnet_20240620_inspect11.4 min (5.5-22.4)2024-06-20
19gpt_4o_inspect7.0 min (4.0-12.9)2024-05-13
20gpt_4_1106_inspect4.0 min (1.9-8.4)2023-11-06
21gpt_44.0 min (1.9-8.0)2023-03-14
22claude_3_opus_inspect4.0 min (1.7-8.8)2024-03-04
23gpt_4_turbo_inspect3.7 min (2.0-6.7)2024-04-09
24gpt_3_5_turbo_instruct0.6 min (0.3-1.1)2022-03-15
25davinci_0020.1 min (0.1-0.2)2020-05-28

Note: METR-Horizon-v1.1 data file, top 25 sorted by p50 estimate (rank is mine, not source order). Page says it is no longer actively updated; unit minutes inferred (GPT-5 ~203 min vs. page's ~2h17m).

Source: METR · metr.org · results as of 2026-09-08 · retrieved 2026-11-17

Markdown · More Agents, tool use & computer use