Benchmark explainer
What is METR Time Horizons?
Autonomous agent capability as the length of task (in human completion time) an agent can complete at a given reliability (50%/80% time horizon).
METR Time Horizons as reported in Epoch AI's Capabilities Index CSV.
How it's scored
- Metric
- normalized performance in Epoch CSV, rendered as accuracy (%)
- Score ceiling
- 100
- Construction
- Composite of 3 task suites: RE-Bench (ML research engineering), HCAST (challenging SWE), SWAA (small SWE ops); ~8 runs per model per task with logistic regression on log-time.
- Human baseline
- Human-expert completion time per task is the calibration baseline (logistic regressor input).
How trustworthy
Discrimination
100/100Does it still separate models?
Saturation headroom
95.4/100How far from ceiling / clustered at the top?
Contamination resistance
41.2/100Public vs held-out; training-leak risk.
Harness comparability
50/100Is it apples-to-apples, or mixed / vendor-optimized?
Freshness
88.3/100How old is the benchmark?
Contamination history: HCAST and SWAA tasks are public on GitHub — models trained on task text may have seen instances.
Limitations: Task suite skews toward software/ML engineering (domain-bounded); selection of agents (Claude Code, OpenHands, Codex) affects scoring; horizon doubles ~7 months — extrapolation risk.
Who leads METR Time Horizons
| Model | Score | Evidence |
|---|---|---|
| Claude Opus 4.6 | 78.86 | unverified |
| Gemini 3.1 Pro | 77.03 | self-reported |
| GPT-5.2 | 75.29 | unverified |
| Claude Opus 4.5 | 74.97 | unverified |
| GPT-5.3 Codex | 74.54 | unverified |
Compare the top METR Time Horizons scorers
Frequently asked questions
METR Time Horizons: METR Time Horizons as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 4.6 leads METR Time Horizons at 78.86. The full leaderboard above lists every recorded measurement, not just the headline number.
34 models have recorded scores on METR Time Horizons, spanning a score spread of 49.38.
tensor.news grades METR Time Horizons B for integrity (score 81/100), ranking #42 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — METR Time Horizons still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Only partly — scores come from a mix of harnesses and protocols, so a head-to-head on METR Time Horizons is not fully apples-to-apples.
Every METR Time Horizons measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites metr.org/time-horizons.