How Much Longer Can AI Work Alone in 2026 Than in 2019?

This is not a measure of how long an AI takes to think. It's a difficulty scale, measured in human time. METR gives each model a large set of test tasks — some short enough that a skilled person finishes in seconds, some long enough to take a full day — and finds the task length where the model starts failing about half the time. That length, expressed in the minutes a skilled human professional would need, is the model's time horizon.

So when Claude Opus 4.6 is described as having a roughly 12-hour time horizon, the claim is specific: give it something that would take a skilled person about 12 hours, and it succeeds roughly half the time. Give it something a person would need two full days for, and it will usually fail. That's a ceiling on the task difficulty a model can be trusted with unsupervised — not a measurement of how fast the model itself works.

METR builds the time horizon from three task suites: HCAST (real software engineering work), RE-Bench (machine learning research tasks), and SWAA (short, atomic actions used to anchor the low end of the scale). Every task in those suites also has a human baseline — how long a paid contractor with the relevant skill actually took to finish it. A model's 50% time horizon is the task length, on that human-minutes scale, where the model's success rate crosses 50%.

The design choice that matters most is the human baseline. It converts an abstract pass/fail benchmark score into something with a real unit: minutes of skilled human work. That is what makes GPT-2's 2 seconds and Claude Opus 4.6's roughly 12 hours comparable numbers on the same axis, rather than two different scores on two different scales. METR published the full methodology in "Measuring AI Ability to Complete Long Software Tasks" (Kwa et al., arXiv:2503.14499, March 2025), later presented at NeurIPS 2025, and has since expanded the leaderboard with a January 2026 update called Time Horizon 1.1.

GPT-2, released in February 2019, managed roughly 2 seconds of task length at 50% reliability — about as long as it takes to read a single sentence, and essentially the floor of this entire dataset. Four years later, GPT-4 (the March 2023 checkpoint) reached about 4 minutes — enough to answer a quick factual question or fix a one-line bug — a jump of two orders of magnitude that reflects the leap from a text predictor to a genuinely task-capable model.

Claude 3.5 Sonnet (New) followed in October 2024 at about 35 minutes — a short, well-defined coding task a developer could hand off and expect back, more or less finished, without checking in halfway through. It's the first model in this dataset that reads less like a demo of what AI can do and more like a tool someone could actually delegate real, bounded work to.

February 2025 landed two milestones within days of each other: GPT-4.5's preview checkpoint reached about 30 minutes — a short, well-defined task like writing a small script or summarizing a document — and Claude 3.7 Sonnet reached 50 minutes, roughly a focused, coffee-break-length task. Neither model was dramatically ahead of the other; together they marked the point where 'AI task length' stopped being measured in single-digit minutes.

o3, released by OpenAI in April 2025, reached about 110 minutes — close to two hours, and enough to work through most of a morning's focused task before a human would need to check back in.

GPT-5 followed in August 2025 at a 2h17m (137-minute) point estimate — a solid half-day chunk of work. METR's own 95% confidence interval for that figure runs from 65 minutes to 4 hours 25 minutes, a wide band that is typical of estimates this far out on the scale, where fewer test tasks exist to pin the number down precisely.

Claude Opus 4.6, added to METR's leaderboard on February 20, 2026, is the highest point in this dataset by a wide margin: a 50% time horizon of roughly 718 minutes, or about 12 hours. That is more than five times o3's score from ten months earlier, and it means a model can now be trusted, at 50% reliability, to work through a task that would occupy a skilled professional for the better part of a working day.

It is also the least certain point on the chart. METR's own published confidence interval for Opus 4.6 runs from about 5.3 hours to 65.8 hours — a band wide enough that the true value could plausibly sit anywhere from "a long afternoon" to "most of a work week." That is not a flaw unique to this model; it is a structural feature of measuring long time horizons, where very few tasks exist that take that long to complete, so the estimate rests on a thin slice of the test suite.

A separate compilation of the same underlying data put Opus 4.6 closer to 14.5 hours rather than 12. Both figures fall comfortably inside the published 5.3-65.8 hour interval, so this is less a factual contradiction than a reminder of how much room that interval leaves for disagreement about the exact point estimate.

This dataset is eight confirmed points across seven years, not a complete model census. METR's own leaderboard has added several models this piece deliberately leaves out — GPT-4o, o1, Claude Opus 4.5, Claude Sonnet 4.5, Gemini 3 Pro, Grok 4, and the GPT-5.1 through 5.4 line among them — because a minute-level point estimate for each could not be sourced from a reference solid enough to stand behind in this pass. Filling those gaps would meaningfully thicken the middle and right of the timeline, and is flagged as follow-up work rather than guessed at here.

GPT-3 and GPT-3.5 are absent for a related reason: METR itself dropped them from the current leaderboard, citing the need for significant changes to the tool-calling scaffold to evaluate them fairly, and no primary-sourced minute-level figure for either could be confirmed independently.

One further wrinkle worth naming: METR's own sources disagree slightly on Claude 3.5 Sonnet (New), with a preliminary evaluation report putting it at about 35 minutes and the later time-horizon paper placing the broader late-2024 model cohort closer to 40 minutes. Both figures are METR-sourced and the gap likely reflects different task-suite versions rather than an error; this piece uses the 35-minute figure throughout, as the more specific, model-level estimate.

Read as a set of milestones rather than a statistic, the climb is easy to picture: a sentence in 2019, a one-line bug fix in 2023, a real coding task in 2024, a morning's work in April 2025, a half-day chunk by August 2025, and a full unsupervised workday by February 2026. Each step is a task a person would recognize, which is exactly the point of measuring 'time horizon' in human minutes rather than an abstract benchmark score.

It's worth being just as clear about what this dataset doesn't settle. Eight confirmed anchor points is a strong skeleton, not a full census — several 2023-2026 releases are missing because a confident minute-level figure couldn't be sourced for them, and the two newest points, GPT-5 and Claude Opus 4.6, carry the widest published uncertainty bands in the whole series.

What comes next is a question of scale, not a new kind of question: will the frontier keep moving from 'a full workday' toward 'a full week,' and how will METR's own methodology hold up as it tries to measure tasks long enough that almost no test suite has enough of them to test against.