METR 80% time horizon
Task length a model completes with 80% success on the same suite. The band applies to a log-linear doubling-time fit over the 80% series from 2024 onward (derived metric horizon_doubling_days_80_since_2024), the same window as the 50% horizon.
Reason through thisThe reading
Fast, but lagging the 50% horizon; the lag is the reliability gap the products bucket has to close.
26 readings drawn from metr.*.horizon_80.pt; 26 observations in the evidence.
Source: METR time horizons (v1.1). METR, Measuring AI Ability to Complete Long Tasks (Horizon v1.1).
The numbers
| as of | reading | value | record |
|---|---|---|---|
| 2026-04-07 | claude mythos preview early inspect | 3.1 h (1.6 h–6.6 h) | measured |
| 2026-03-05 | gpt 5 4 | 53.9 min (24.0 min–1.8 h) | measured |
| 2026-02-19 | gemini 3 1 pro | 1.5 h (52.0 min–2.6 h) | measured |
| 2026-02-05 | gpt 5 3 codex | 54.7 min (22.4 min–2.0 h) | measured |
| 2026-02-05 | claude opus 4 6 inspect | 1.2 h (27.0 min–2.8 h) | measured |
| 2025-12-11 | gpt 5 2 | 1.1 h (31.7 min–2.2 h) | measured |
| 2025-11-24 | claude opus 4 5 inspect | 49.4 min (20.7 min–1.7 h) | measured |
| 2025-11-19 | gpt 5 1 codex max inspect | 50.6 min (27.3 min–1.5 h) | measured |
| 2025-11-18 | gemini 3 pro | 54.1 min (25.7 min–1.7 h) | measured |
| 2025-08-07 | gpt 5 2025 08 07 inspect | 38.3 min (19.2 min–1.2 h) | measured |
| 2025-08-05 | claude 4 1 opus inspect | 23.5 min (10.5 min–47.8 min) | measured |
| 2025-05-22 | claude 4 opus inspect | 20.4 min (8.48 min–41.9 min) | measured |
| 2025-04-16 | o3 inspect | 30.0 min (15.1 min–57.8 min) | measured |
| 2025-02-24 | claude 3 7 sonnet inspect | 12.1 min (4.58 min–28.9 min) | measured |
| 2024-12-05 | o1 inspect | 7.09 min (3.04 min–16.6 min) | measured |
| 2024-10-22 | claude 3 5 sonnet 20241022 inspect | 2.60 min (0.89 min–7.36 min) | measured |
| 2024-09-12 | o1 preview | 4.42 min (2.01 min–8.91 min) | measured |
| 2024-06-20 | claude 3 5 sonnet 20240620 inspect | 1.67 min (0.58 min–4.57 min) | measured |
| 2024-05-13 | gpt 4o inspect | 1.27 min (0.56 min–3.02 min) | measured |
| 2024-04-09 | gpt 4 turbo inspect | 0.93 min (0.43 min–2.20 min) | measured |
| 2024-03-04 | claude 3 opus inspect | 0.64 min (0.19 min–2.12 min) | measured |
| 2023-11-06 | gpt 4 1106 inspect | 0.78 min (0.28 min–2.36 min) | measured |
| 2023-03-14 | gpt 4 | 0.89 min (0.34 min–2.52 min) | measured |
| 2022-03-15 | gpt 3 5 turbo instruct | 0.26 min (0.14 min–0.49 min) | measured |
| 2020-05-28 | davinci 002 | 0.056 min (0.035 min–0.11 min) | measured |
| 2019-02-14 | gpt2 | 0.013 min (0.0016 min–0.069 min) | measured |
Why it matters
80% is closer to what a product needs than 50%. If the reliable horizon lags badly, benchmarks flatter products.
What cuts against it
Our fit has no confidence interval and excludes points METR marks unreliable; the 2023-onward window gives ~125 days, which would read emerging. The 80% horizon is still an order of magnitude below the 50% horizon, which is the reliability gap the products bucket has to close.
How it is measured
- series
metr.*.horizon_80.pt - source METR time horizons (v1.1) · default tier 1 · METR; analysis code MIT at https://github.com/METR/eval-analysis-public
- proxy benchmark · cadence per release
Same bands and same 2024-onward window as the 50% horizon, applied to our own fit because METR publishes doubling times for the 50% series only. The 2023-onward fit (horizon_doubling_days_80) sits in the gap between the bands.
Applied to metric:horizon_doubling_days_80_since_2024.
Latest readings
- 2026-04-07 claude_mythos_preview_early_inspect · 3.1 h(1.6 h–6.6 h)as of 2026-04-07
- 2026-03-05 gpt_5_4 · 53.9 min(24.0 min–1.8 h)as of 2026-03-05
- 2026-02-19 gemini_3_1_pro · 1.5 h(52.0 min–2.6 h)as of 2026-02-19
- 2026-02-05 gpt_5_3_codex · 54.7 min(22.4 min–2.0 h)as of 2026-02-05
- 2026-02-05 claude_opus_4_6_inspect · 1.2 h(27.0 min–2.8 h)as of 2026-02-05
- 2025-12-11 gpt_5_2 · 1.1 h(31.7 min–2.2 h)as of 2025-12-11
Why this status, and its history
Same 2024-onward window as the 50% horizon: our fit gives 107 days, inside the fast band; the 2023-onward fit is 125 days, in the gap. No confidence interval on our fit; the 80% horizon remains roughly 6x below the 50% horizon.
- 2026-09-10emerging to faster than normalconf 70 → 70 · initial seed (model-drafted)
Same 2024-onward window as the 50% horizon: our fit gives 107 days, inside the fast band; the 2023-onward fit is 125 days, in the gap. No confidence interval on our fit; the 80% horizon remains roughly 6x below the 50% horizon.
- 2026-09-10unmeasured to emergingconf — → 70 · initial seed (model-drafted)
Log-linear fit over the 80% horizon, 2023 onward, gives a doubling time of 124.7 days, just above the fast band ceiling (122). No CI on our fit. Initial seed; pending Alex's review.
Why this confidence
Rests on 26 observations from METR time horizons (v1.1); best evidence tier 1, grade A; current.
Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.
Predictions that lean on this
- HappeningThe length of task AI can finish keeps doubling about every four months. · AI Futures Project (Kokotajlo, Lifland, Jurkovic et al.)
- Slower than saidIf the trend speeds up, by March 2027 AI could reliably do software tasks that take people years. · AI Futures Project (Kokotajlo, Lifland, Jurkovic et al.)
- Slower than said
- Too early to tellAn AI that codes better than the best people arrives in 2027. · AI Futures Project
- Too early to tell
More predictions that lean on this
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell