On February 4, 2026, AI evaluation nonprofit METR added GPT-5.2 to its time-horizons tracker — one day after Gemini 3 Pro and GPT-5.1 Codex Max went in. The number landed as a record: in “high” reasoning mode, GPT-5.2 shows a 50% time horizon of roughly 6.6 hours (95% confidence interval from 3 hours 20 minutes to 17 hours 30 minutes), the highest mark the benchmark has recorded.
The metric measures something concrete: the length of task — expressed as the time a human expert would need — that an agent completes successfully about half the time. At 6.6 hours, “hand over an afternoon of work” enters reliable territory for the first time. Compare that with mid-2025 models hovering around the one-hour mark, and this is the watershed between agents as demos and agents as delegable workers.
What a Time Horizon Is and How METR Measures It
METR defines the time horizon as the task duration, measured by human expert completion time, at which an agent is predicted to succeed with a given reliability. The procedure: contracted skilled professionals estimate completion times for each task (geometric mean), METR fits a logistic curve to success probability versus the logarithm of human time, and reads off where the curve crosses 50% (and 80%). Each model is run roughly eight times per task.
The tasks come from three suites — RE-Bench (ML research engineering), HCAST (broader software engineering), and SWAA (shorter software-operation tasks) — more than a hundred in total. Worth stressing: this is engineering work. A time horizon is a statement about task length in software and research domains, not general office capability.
What GPT-5.2’s 6.6 Hours Actually Says
Three pieces of context belong next to the headline. First, 6.6 hours is the result at “high” reasoning effort — not the more extreme xhigh setting. Second, the confidence interval is wide (3 hours 20 minutes to 17 hours 30 minutes); the point estimate carries real uncertainty. Third, METR states plainly that measurements above 16 hours are unreliable with the current task suite — the frontier is approaching the measurement ceiling, which is why the suites keep expanding.
How It Stacks Up Against Earlier Models
Plot the points and the slope is the story:
- GPT-5 (measured August 2025): around 2 hours 17 minutes
- GPT-5.1 Codex Max: around 2 hours 42 minutes
- Claude Opus 4.5 (added December 2025): around 4 hours 49 minutes
- GPT-5.2 (added February 4, 2026, high reasoning): around 6.6 hours — top of the chart on arrival
In one generation, OpenAI pushed its own record from 2 hours 17 minutes to 6.6 hours — nearly a 3x jump — with Gemini 3 Pro and GPT-5.1 Codex Max, added the same week, sitting below it. The previous leader, Anthropic’s Claude Opus 4.5, held its 4-hour-49-minute record for only about six weeks.
Doubling Every 4–7 Months, With No Slowdown in Sight
METR’s historical data shows the 50% time horizon doubling roughly every 4 to 7 months, and an exponential curve fits the data better than linear or hyperbolic alternatives — no slowdown has been observed. GPT-5.2’s result is widely read as evidence the doubling is, if anything, accelerating. Extrapolate the rhythm and tasks at “reliable workday length” (4 to 8 hours) become routine within 2026. If you are designing agent architectures, that is the baseline worth planning against — the same call we made in our 2026 opening outlook.
Practical Implications for Engineering Teams
Three sober conclusions. First, rewrite your task-decomposition strategy: once six-hour tasks clear 50% reliability, the old “shred everything into 30-minute steps and reassemble” playbook gives way to “delegate the whole chunk, verify at checkpoints.” Second, monitoring matters more than ever — a coin-flip success rate means half of delegated work fails, so acceptance tests and rollback paths are part of the architecture, not extras. Third, select models on time horizon, not single-point scores: a SWE-bench number will not tell you whether a model can sustain an afternoon of continuous work.
Sources
- Task-Completion Time Horizons of Frontier AI Models — METR
- METR estimate for OpenAI GPT-5.2 (high reasoning effort) — METR on LinkedIn
- METR Time Horizons — Epoch AI
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.