Summary: METR added GPT-5.2 to its time-horizons tracker on February 4, 2026 and reported a record 50% time horizon of about 6.6 hours with high reasoning effort. The measure reflects task length in human expert hours that an agent completes half the time. The confidence interval runs from 3 hours 20 minutes to 17 hours 30 minutes, and the benchmark suite is dominated by engineering tasks.
METR, the AI evaluation nonprofit, entered GPT-5.2 into its time-horizons tracker on February 4, 2026 — Gemini 3 Pro and GPT-5.1 Codex Max had joined the list just one day earlier. The figure arrived as a record on the spot: with “high” reasoning effort selected, GPT-5.2 posts a 50% time horizon of about 6.6 hours (95% confidence interval from 3 hours 20 minutes to 17 hours 30 minutes), the strongest reading this benchmark has produced.
The question behind the metric is concrete: hand a job to an agent, and how long is the task — counted in the hours a human expert would need for the same job — that it finishes half the time? At 6.6 hours, handing off an entire afternoon of work sits inside dependable range for the first time. Set that against mid-2025 models stuck near one hour, and the line is clear: agents have crossed over from demos to work you can actually delegate.
The Metric, Defined: Where the Number Comes From
METR’s definition: the time horizon is how long a task an agent is predicted to complete at a stated reliability level, with duration reckoned in human expert completion time. The procedure runs like this — professionals trained for the job estimate how long each task takes (the individual estimates are combined as a geometric mean), a logistic regression is fitted with success rate against the logarithm of human time, and the 50% and 80% crossings are read off the fitted curve. Every model takes each task roughly eight times.
The tasks are drawn from three suites — RE-Bench (ML research engineering), HCAST (software engineering with a wider span), and SWAA (shorter software-operation items) — passing one hundred in total. One caveat deserves emphasis: engineering dominates the mix. A time horizon is a statement about how long engineering work can run, not a gauge of general office capability.
Reading GPT-5.2’s 6.6 Hours Properly
Three qualifications belong next to the headline. For one, the 6.6-hour figure comes from “high” reasoning effort, not the harsher xhigh tier. For another, the confidence interval stretches from 3 hours 20 minutes out to 17 hours 30 minutes — a wide band, and the lone point estimate carries real error with it. Finally, METR says outright that the current task suite cannot measure anything beyond 16 hours reliably — task difficulty is closing in on the ceiling of the instrument, and that ceiling is why the suites keep gaining new items.
Side by Side With the Models Before It
Put the models on one chart and the slope of the connecting line tells the story:
- GPT-5 (measured August 2025): roughly 2 hours 17 minutes
- GPT-5.1 Codex Max: roughly 2 hours 42 minutes
- Claude Opus 4.5 (added December 2025): roughly 4 hours 49 minutes
- GPT-5.2 (added February 4, 2026, high reasoning mode): roughly 6.6 hours — first place the moment it landed
In short, OpenAI needed one generation to lift its own record from 2 hours 17 minutes to 6.6 hours, close to a tripling, and both Gemini 3 Pro and GPT-5.1 Codex Max, inducted the same week, finished underneath it. The model that led until then, Anthropic’s Claude Opus 4.5, saw its 4-hour-49-minute mark last about six weeks.
Exponential Climb: A Doubling Every 4–7 Months
Across METR’s accumulated history, the 50% time horizon doubles roughly every 4 to 7 months, and when candidate curves are fitted to the data, the exponential tracks it better than the linear or the hyperbolic — nothing in the record shows the pace easing off. GPT-5.2’s score is broadly read as evidence that the doubling continues to accelerate. Push the trajectory forward and tasks of “reliable workday length” (4 to 8 hours) become routine inside 2026; build agent architectures on that assumption from the start — the same position we took in our 2026 opening outlook.
What Engineering Teams Should Change
Three grounded takeaways. One: the task-splitting doctrine needs a rewrite — once six-hour tasks clear the halfway reliability mark, the habit of shredding work into 30-minute pieces and reassembling them yields to delegating the whole assignment and verifying at checkpoints. Two: monitoring deserves a larger share of attention — with success at 50%, half of everything delegated goes wrong, which makes acceptance tests and rollback paths part of the architecture itself rather than extras bolted on afterward. Three: choose models by time horizon rather than any single score — a SWE-bench number cannot answer whether a model can sustain an afternoon of continuous work.