METR 50% time horizon
Length (in human-expert minutes) of software tasks a frontier model completes with 50% success, per METR's Horizon v1.1 suite, plotted by model release date. The band applies to the doubling time of that series over the most recent window METR itself reports (2024 onward), recomputed here from METR's corrected data file; METR's published 2023-onward figure is shown alongside.
Reason through thisThe reading
What models can do is improving fast by any historical standard. The open question the rule answers is whether it is improving as fast as AI 2027, the most aggressive widely read forecast, needs.
26 readings drawn from metr.*.horizon_50.pt, one of 5 series listed under How it is measured; 80 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 (disputed) | 17.4 h (8.5 h–55.1 h) | measured |
| 2026-03-05 | gpt 5 4 | 5.7 h (3.1 h–12.8 h) | measured |
| 2026-02-19 | gemini 3 1 pro | 6.4 h (3.9 h–11.6 h) | measured |
| 2026-02-05 | gpt 5 3 codex | 5.8 h (3.2 h–13.6 h) | measured |
| 2026-02-05 | claude opus 4 6 inspect | 12.0 h (5.3 h–60.6 h) | measured |
| 2025-12-11 | gpt 5 2 | 5.9 h (3.3 h–13.6 h) | measured |
| 2025-11-24 | claude opus 4 5 inspect | 4.9 h (2.7 h–10.4 h) | measured |
| 2025-11-19 | gpt 5 1 codex max inspect | 3.7 h (2.2 h–6.6 h) | measured |
| 2025-11-18 | gemini 3 pro | 3.7 h (2.3 h–6.3 h) | measured |
| 2025-08-07 | gpt 5 2025 08 07 inspect | 3.4 h (1.9 h–6.8 h) | measured |
| 2025-08-05 | claude 4 1 opus inspect | 1.7 h (59.3 min–2.7 h) | measured |
| 2025-05-22 | claude 4 opus inspect | 1.7 h (60.0 min–2.7 h) | measured |
| 2025-04-16 | o3 inspect | 2.0 h (1.2 h–3.2 h) | measured |
| 2025-02-24 | claude 3 7 sonnet inspect | 1.0 h (33.0 min–1.7 h) | measured |
| 2024-12-05 | o1 inspect | 38.8 min (21.2 min–1.1 h) | measured |
| 2024-10-22 | claude 3 5 sonnet 20241022 inspect | 20.5 min (10.1 min–40.8 min) | measured |
| 2024-09-12 | o1 preview | 20.3 min (11.7 min–33.4 min) | measured |
| 2024-06-20 | claude 3 5 sonnet 20240620 inspect | 11.4 min (5.49 min–22.4 min) | measured |
| 2024-05-13 | gpt 4o inspect | 6.99 min (4.00 min–12.9 min) | measured |
| 2024-04-09 | gpt 4 turbo inspect | 3.73 min (1.98 min–6.74 min) | measured |
| 2024-03-04 | claude 3 opus inspect | 3.95 min (1.71 min–8.76 min) | measured |
| 2023-11-06 | gpt 4 1106 inspect | 4.04 min (1.87 min–8.44 min) | measured |
| 2023-03-14 | gpt 4 | 3.99 min (1.93 min–8.00 min) | measured |
| 2022-03-15 | gpt 3 5 turbo instruct | 0.60 min (0.26 min–1.12 min) | measured |
| 2020-05-28 | davinci 002 | 0.14 min (0.093 min–0.22 min) | measured |
| 2019-02-14 | gpt2 | 0.054 min (0.010 min–0.14 min) | measured |
Why it matters
The cleanest single measure of how fast the models themselves are improving. Narayanan and Kapoor's argument is that fast improvement here does not, on its own, make firms, jobs or productivity change quickly.
What cuts against it
On the 2023-onward window (METR's published 128.7 days, our fit 125) the status would be emerging, not faster than normal; the status depends on the window. METR notes measurements above 16 hours are unreliable with the current task suite, so the newest points carry wide intervals and are flagged disputed, and human baselines exist for only 5 of the 31 tasks over 8 hours. Horizon is a software-task construct and may not transfer to other work.
How it is measured
- series
metr.*.horizon_50.pt - series
metr.suite.doubling_time_days.pt - series
metr_blog.suite.doubling_time_days_since_2024.pt - series
metr_blog.suite.doubling_time_days_since_2023.pt - series
metr_blog.suite.doubling_time_days_2019_2025.pt - source METR time horizons (v1.1) · default tier 1 · METR; analysis code MIT at https://github.com/METR/eval-analysis-public
- source METR research posts · default tier 6 · METR
- proxy benchmark · cadence per release
Normal = doubling time ≥ 7 months (≈213 days, METR's 2019–2025 trend); fast = ≤ 4 months (≈122 days), the AI 2027 assumption. Between the two is `emerging`. The bands are about the current pace, so the input is a log-linear fit over the 2024-onward window, the most recent window METR reports (89 days in the TH1.1 post of 29 Jan 2026), recomputed from METR's corrected data. On the 2023-onward window METR's published figure is 128.7 days, in the gap between the bands: the reading is window-sensitive and both numbers are shown.
Applied to metric:horizon_doubling_days_50_since_2024.
| fit | estimate | 95% interval | n | R² | AIC | inputs |
|---|---|---|---|---|---|---|
horizon_doubling_days_50_since_2024 | 108 days | 99 days–119 days | 20 | 0.966 | -44.8 | sources12+23 more |
horizon_doubling_days_50_since_2023 | 125 days | 111 days–143 days | 22 | 0.933 | -31.0 | sources12+23 more |
horizon_blowup_year_50_since_2024 | 2025.8 | 2025.4–2026.8 | 20 | 0.645 | 78.1 | sources12+23 more |
Latest readings
- 2026-04-07 claude_mythos_preview_early_inspect · 17.4 h(8.5 h–55.1 h)⚑disputed: METR: measurements above 16 hours are unreliable with the current task suiteas of 2026-04-07
- 2026-03-05 gpt_5_4 · 5.7 h(3.1 h–12.8 h)as of 2026-03-05
- 2026-02-19 gemini_3_1_pro · 6.4 h(3.9 h–11.6 h)as of 2026-02-19
- 2026-02-05 gpt_5_3_codex · 5.8 h(3.2 h–13.6 h)as of 2026-02-05
- 2026-02-05 claude_opus_4_6_inspect · 12.0 h(5.3 h–60.6 h)as of 2026-02-05
- 2025-12-11 gpt_5_2 · 5.9 h(3.3 h–13.6 h)as of 2025-12-11
Why this status, and its history
Band input switched to the most recent window METR reports (2024 onward), per the updated rationale: our log-linear fit on METR's corrected v1.1 data gives a doubling time of 108 days, inside the fast band (<=122). METR's own TH1.1 post (29 Jan 2026) reports 89 days since 2024. On the 2023-onward window the figures are 128.7 (METR) and 125 (ours), which would read emerging; the status is window-sensitive and both are shown on the page.
- 2026-09-10emerging to faster than normalconf 80 → 80 · initial seed (model-drafted)
Band input switched to the most recent window METR reports (2024 onward), per the updated rationale: our log-linear fit on METR's corrected v1.1 data gives a doubling time of 108 days, inside the fast band (<=122). METR's own TH1.1 post (29 Jan 2026) reports 89 days since 2024. On the 2023-onward window the figures are 128.7 (METR) and 125 (ours), which would read emerging; the status is window-sensitive and both are shown on the page.
- 2026-09-10unmeasured to emergingconf — → 80 · initial seed (model-drafted)
METR's published 2023-onward doubling time is 128.7 days (CI 104-158), between the normal band (>=213 days) and the fast band (<=122 days). Initial seed; the brief expected faster_than_normal. Pending Alex's review.
source1
Why this confidence
Rests on 80 observations from METR research posts, 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.)
- Happening
- Happening
- Happening
- Slower than said
More predictions that lean on this
- Both sides expect this
- Too early to tellAn AI that codes better than the best people arrives in 2027. · AI Futures Project
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tellAI improving itself cannot get past limits that sit outside the lab. · Narayanan & Kapoor
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tellAI costs too much for the problems it can solve today. · Jim Covello (Goldman Sachs)