This is a post on METR's site, credited to METR and dated 29 January 2026, announcing version 1.1 of its time-horizon estimates. It is distinct from the March 2025 paper that introduced the methodology (Measuring AI Ability to Complete Long Software Tasks) and updates the estimates that paper produced.
Summary
METR states that its early-2025 methodology found "a steady exponential increase in models' human-equivalent 'time horizon'," and that applying it to newer models through 2025 gave a rate consistent with the historical trend. The post rolls out two changes and reports their effect.
Changes to the task suite and infrastructure
The suite grows from 170 to 228 tasks. Seventy-three tasks were added, all from HCAST (HCAST: Human-Calibrated Autonomy Software Tasks (Rein et al., 2025)) — tasks not included in the original paper that subsequently passed quality checks; 15 were removed; and 53 were updated (27 with a new definition, 13 with a new human time estimate, 13 with both). Removals and modifications "generally represent cases where a task description was confusing, or easy to reward-hack, or the scoring function had errors." The number of long tasks, estimated at eight or more human hours, rises from 14 to 31, of which only 5 have measured human baseline times; the rest use estimates. Evaluation moves from Vivaria, developed in-house in 2023, to Inspect, the open-source framework developed by the UK AI Security Institute.
Re-estimated horizons
Horizons were re-estimated for 14 of the 33 models with TH1 estimates; the others were dropped because they are no longer available, need substantial scaffold changes (GPT-2, GPT-3, GPT-3.5), or were far from the frontier at release. New estimates generally lie within the TH1 confidence intervals. Changes are attributed mainly to the task suite and run-to-run noise, with a smaller share from the infrastructure. The larger suite tightens intervals at the upper end: the upper bound for Claude Opus 4.5 was 4.4× its point estimate under TH1 and is 2.3× under TH1.1. Point estimates (50% horizon, minutes, with intervals): Claude Opus 4.5 289 [110, 1268] → 320 [170, 729] (+11%); GPT-5 138 → 214 (+55%); Claude Opus 4 86 → 101 (+18%); o3 94 → 121 (+29%); Claude Sonnet 3.7 56 → 60 (+7%); GPT-4 1106 8.5 → 3.6 (−57%); GPT-4 0314 5.4 → 3.5 (−35%). GPT-5.1-codex-max (173), Claude Sonnet 4.5 (122), Claude Opus 4.1 (114), Grok 4 (109) and Claude Sonnet 4 (75) have TH1 estimates only. A footnote records that GPT-4-32k-0314 and o1-preview were run on the new tasks but on Vivaria before deprecation.
Effect on the trend
Because pre-2023 models were not re-estimated, METR makes two comparisons. A hybrid trend using TH1 estimates for GPT-2, GPT-3 and GPT-3.5 gives the same doubling time as TH1, 196 days (about seven months; 195.8 days [162, 223] under TH1.0, 196.5 under the stitched TH1.1), with a slightly lower R². Since 2023, where both methods have data, the doubling time is 130.8 days [107, 161] under TH1.1 against 165.3 days [129, 211] under TH1, "meaning progress is estimated to be 20% more rapid under TH1.1"; since 2024 it falls from 108.9 to 88.6 days. The shift comes from older models' estimates moving down and recent ones moving up. METR notes the TH1.1 doubling time is within TH1's confidence interval but that those intervals represent an entirely new task draw, whereas the suites overlap substantially. It concludes the new tasks are "drawn from a slightly different distribution of difficulty," that the suite is "overall an improvement, and so worth this slight change in the quantity being estimated," and that the exercise "highlighted the importance of defining the distribution of tasks which we wish to measure performance over." It also states that even the new suite has relatively few tasks the latest models cannot do and that raising the ceiling is a priority.
Vivaria versus Inspect
On the TH1 tasks, five models were run on both platforms: GPT-4o 9.2 versus 6.0 minutes, Claude 3.7 Sonnet 56 versus 59, o3 94 versus 78, Claude Sonnet 4.5 122 versus 164, Claude Opus 4.5 289 versus 270. Paired t-tests found GPT-4o and o3 scored significantly higher under Vivaria, which METR reads as scaffold sensitivity; Vivaria gave higher scores on 51 tasks and Inspect on 31, with only 3 tasks significant, and outliers did not look systematic. On the new suite, three models gave very similar horizons on both platforms (GPT-4o 6 versus 5 minutes, Claude 3.7 Sonnet 56 versus 57, o3 116 versus 119). Data for both versions is in METR's eval-analysis-public repository.
Key claims
- Suite: 170 → 228 tasks; +73 (HCAST), −15, 53 updated; 8h+ tasks 14 → 31, 5 human-baselined (6 under TH1).
- Models re-estimated: 14 of 33 (11 state-of-the-art versus 17).
- Doubling time: 196 days overall unchanged; post-2023 165.3 → 130.8 days (20% more rapid); post-2024 108.9 → 88.6 days.
- Largest estimate changes: GPT-5 +55%, GPT-4 1106 −57%.
- Platform effect small on the new suite; GPT-4o and o3 significantly higher under Vivaria on the old suite.
Provenance
Retrieved 22 September 2026 from metr.org. The post's BibTeX entry gives METR as author and January 2026 as date.
Relationships
- supersedes: Measuring AI Ability to Complete Long Software Tasks — as the current estimate set, while the 2025 paper remains the methodological reference.
- supports: METR Time Horizons — the primary text for the TH1.1 suite, re-estimates and trend comparison.
- depends-on: HCAST: Human-Calibrated Autonomy Software Tasks (Rein et al., 2025) — the source of the 73 added tasks.
- supports: METR
- related: AI Benchmarks and Evaluation, Agentic harnesses and capability elicitation