Skip to content
veri—ableMy research
All visual stories

CAPABILITIES / STORY 01

How much can an AI agent do reliably?

Change the success threshold. Watch the task horizon change with it.

4 min readEvidence dated 8 May 2026
THE RELIABILITY GAP METR · TH 1.1

A higher bar.
A shorter horizon.

1.2 hrClaude Opus 4.6 · 80% success

Task difficulty, measured in human-expert time Log scale · model release date →

Task horizons at 80% predicted successModels released since 2023. Higher points indicate harder software tasks. Use the model selector or data table below for values and intervals.1 min10 min1.0 hr8.0 hr16 hr2023202420252026Above 16 hr: limited measurement reliabilityGPT-4: 53 secGPT-4 (Nov 2023): 47 secClaude 3 Opus: 38 secGPT-4 Turbo: 56 secGPT-4o: 1 minClaude 3.5 Sonnet (Jun): 2 mino1-preview: 4 minClaude 3.5 Sonnet (Oct): 3 mino1: 7 minClaude 3.7 Sonnet: 12 mino3: 30 minClaude Opus 4: 20 minClaude Opus 4.1: 23 minGPT-5: 38 minGemini 3 Pro: 54 minGPT-5.1 Codex Max: 51 minClaude Opus 4.5: 49 minGPT-5.2: 1.1 hrClaude Opus 4.6: 1.2 hrGPT-5.3 Codex: 55 minGemini 3.1 Pro: 1.5 hrGPT-5.4: 54 minClaude Mythos Preview (early): 3.1 hrSelect GPT-4Select GPT-4 (Nov 2023)Select Claude 3 OpusSelect GPT-4 TurboSelect GPT-4oSelect Claude 3.5 Sonnet (Jun)Select o1-previewSelect Claude 3.5 Sonnet (Oct)Select o1Select Claude 3.7 SonnetSelect o3Select Claude Opus 4Select Claude Opus 4.1Select GPT-5Select Gemini 3 ProSelect GPT-5.1 Codex MaxSelect Claude Opus 4.5Select GPT-5.2Select Claude Opus 4.6Select GPT-5.3 CodexSelect Gemini 3.1 ProSelect GPT-5.4Select Claude Mythos Preview (early)

Reported interval: 27 min–2.8 hr. Selected model released 2026-02-05.

Human task time, not AI runtime
Source: METR Data updated 8 May 2026 · incomplete frontier coverage

THE SHORT ANSWER

Capability is only half the story.

  1. METR measures task difficulty using the time a human expert needs. Its fitted horizons describe 50% or 80% predicted success on the evaluated software tasks.
  2. Requiring greater reliability reduces the estimated task horizon. Compare the same model and setup at both thresholds.

Keep the limits in view.

This measures human-time task difficulty, not uninterrupted AI running time or all jobs. Coverage is incomplete. METR flags estimates above 16 hours as unreliable with this task suite.

Keep asking.

Ask about this evidence

Collection reviewed 30 September 2026. Initial research edition. Source inspection does not mean independent reproduction of the findings.