Skip to content
AI Tracker
← Back to Agents & METR

METR's last task-length reading: about 12 hours for top public AI, then its tests ran out

Milestone

In early 2026 the best public AI succeeded half the time on tasks a skilled person needs 12 hours for. METR's tests can't measure further.

~12 htask length the best public AI finished half the time, Feb–Mar 2026

Length of tasks AI agents can complete on their own

METR time horizon: the length of software tasks (measured in skilled-human time) an AI agent completes with 50% or 80% success. Log scale.

What this measures: How big a job an AI agent can finish on its own. A task's "length" is how long it takes a skilled human; the chart shows the longest tasks an AI completes half the time (blue) or reliably (orange). Tasks are software, machine-learning and cybersecurity work.

  • 50% success
  • 80% success
  • Exponential trend
1 s10 s1 min10 min1 h8 h40 h201920202021202220232024202520262027

Dashed line: exponential fit since 1 Jan 2023, doubling roughly every 4.0 months.

50% success, No longer updated · last data Apr 2026: METR stopped updating this measure on 8 September 2026. The newest models finish almost every task in its test suite, and METR says results above about 16 hours are unreliable, so its last data point is Claude Mythos Preview (April 2026), and no GPT-6, Claude 5.x or Gemini 4 model has a score. METR has moved to measuring how much AI speeds up AI research inside labs. Now watch: METR's AI R&D acceleration audits (e.g. Anthropic, Nov 2026) and Epoch AI/METR's MirrorCode weeks-long coding tasks.

80% success, No longer updated · last data Apr 2026: METR stopped updating this measure on 8 September 2026. The newest models finish almost every task in its test suite, and METR says results above about 16 hours are unreliable, so its last data point is Claude Mythos Preview (April 2026), and no GPT-6, Claude 5.x or Gemini 4 model has a score. METR has moved to measuring how much AI speeds up AI research inside labs. Now watch: METR's AI R&D acceleration audits (e.g. Anthropic, Nov 2026) and Epoch AI/METR's MirrorCode weeks-long coding tasks.

METR Time Horizon 1.1 values (after METR's March 2026 correction), plotted at model release dates. Tasks are well-specified software, ML and cybersecurity tasks. Confidence intervals are wide (e.g. Claude Opus 4.6: 317–3,634 min). METR stopped updating this dataset in September 2026 and says measurements above ~16 hours are unreliable with its current tasks, so models released after spring 2026 are not shown.

Source: METRData as of 9 Nov 2026
Show data table and sources
SeriesDateModelValue
50% success14 Feb 2019GPT-26 s
50% success28 May 2020davinci-002 (GPT-3)6 s
50% success15 Mar 2022GPT-3.5 Turbo Instruct36 s
50% success14 Mar 2023GPT-44 min
50% success6 Nov 2023GPT-4 11064 min
50% success4 Mar 2024Claude 3 Opus4 min
50% success9 Apr 2024GPT-4 Turbo3.7 min
50% success13 May 2024GPT-4o7 min
50% success20 Jun 2024Claude 3.5 Sonnet (Jun)11 min
50% success12 Sept 2024o1-preview20 min
50% success22 Oct 2024Claude 3.5 Sonnet (Oct)21 min
50% success5 Dec 2024o139 min
50% success24 Feb 2025Claude 3.7 Sonnet1 h
50% success16 Apr 2025o32 h
50% success22 May 2025Claude Opus 41.7 h
50% success5 Aug 2025Claude Opus 4.11.7 h
50% success7 Aug 2025GPT-53.4 h
50% success18 Nov 2025Gemini 3 Pro3.7 h
50% success19 Nov 2025GPT-5.1-Codex-Max3.7 h
50% success24 Nov 2025Claude Opus 4.54.9 h
50% success11 Dec 2025GPT-5.25.9 h
50% success5 Feb 2026Claude Opus 4.612 h
50% success5 Feb 2026GPT-5.3-Codex5.8 h
50% success19 Feb 2026Gemini 3.1 Pro6.4 h
50% success5 Mar 2026GPT-5.45.7 h
50% success7 Apr 2026Claude Mythos Preview (limited access)17 h
80% success15 Mar 2022GPT-3.5 Turbo Instruct18 s
80% success14 Mar 2023GPT-454 s
80% success6 Nov 2023GPT-4 110648 s
80% success13 May 2024GPT-4o1.3 min
80% success20 Jun 2024Claude 3.5 Sonnet (Jun)1.7 min
80% success12 Sept 2024o1-preview4.4 min
80% success5 Dec 2024o17.1 min
80% success24 Feb 2025Claude 3.7 Sonnet12 min
80% success16 Apr 2025o330 min
80% success7 Aug 2025GPT-538 min
80% success18 Nov 2025Gemini 3 Pro54 min
80% success11 Dec 2025GPT-5.21.1 h
80% success5 Feb 2026Claude Opus 4.61.2 h
80% success19 Feb 2026Gemini 3.1 Pro1.5 h
80% success7 Apr 2026Claude Mythos Preview (limited access)3.1 h

METR measures AI agents by the length of tasks they can finish: how long a task takes a skilled person, at the point where the agent succeeds half the time (its 50% time horizon). Its Frontier Risk Report (May 2026, covering February–March 2026) put the best public models at about 12 hours (95% CI 5–61 h). For tasks finished reliably (80% of the time), the figure was about 1.5 hours. Labs’ internal models were likely at 16 hours or more, roughly two months ahead.

By then the most capable agents had “essentially saturated” METR’s tasks: the suite had too few harder tasks left to measure them well. Claude Opus 4.6 (February 2026) measured 12 hours. An early version of the limited-access Claude Mythos Preview measured 17.4 hours, but METR said results above 16 hours are unreliable with its current tasks. For GPT-5.6 Sol (June 2026), METR estimated about 11 hours if cheating attempts count as failures, and over 270 hours if they count as successes. It said none of these numbers were robust.

On 8 September 2026 METR marked its time-horizon page as no longer actively updated. Models released since spring 2026 have no METR time horizon.

What it means

The best-known measure of how long AI agents can work on their own has run out of room at the top, so newer models have no like-for-like figure. That does not mean progress stopped, only that this measure can no longer show it. Claims about newer models’ task lengths need other evidence.

Caveats

  • The 12-hour figure is very uncertain: METR’s 95% range runs from 5 to 61 hours.
  • Above about 16 hours the suite has too few long tasks with human timings (5 of 31 eight-hour-plus tasks).
  • Cheating matters: on review, at least 16% of successful runs on the longest tasks were illegitimate (reward hacking, gaming the scoring instead of doing the task).
  • METR notes its reporting is not a complete record of the most capable models.
Suggest a correction