METR Time Horizon Explained: How Long Can GPT-6 Astra Really Work Alone?

As of 15 October 2026, METR has not published a time horizon for GPT-6 Astra: its public TH1.1 dataset lists 26 models, the newest released 2026-04-07, and Astra is not among them. The "time horizon" figures that do circulate for Astra, such as 30.9 minutes, come from a different test with chain-of-thought disabled. And METR's own FAQ says time horizon is not the length of time an agent can act autonomously.

This post separates three things that share the name: METR's software-task time horizon, UK AISI's no-CoT math horizon, and an outside estimate on a no-CoT suite. It then covers why METR's own numbers get shaky at the top of the range.

What is METR's time horizon, and how is it computed?

The task-completion time horizon is the task duration, measured by human expert completion time, at which an agent is predicted to succeed with a given level of reliability. The 50% horizon is the duration where the predicted success rate is half.

The mechanics, per METR's methodology notes:

1. Human professionals are contracted to attempt each task, and METR takes the geometric mean of their successful completion times. For some tasks without reliable baselines, METR uses expert estimates or QA completion times instead.
2. Each model runs in a scaffold. METR launches 6 independent runs per task, and flagged runs are checked for reward hacking by LLM, keyword search and human reviewers.
3. A logistic curve is fitted to success probability against human task duration. The horizon is where the curve crosses 50% (or 80%).

The tasks come from RE-Bench, HCAST and a set of shorter novel software tasks, mostly software engineering, machine learning and cybersecurity. METR says the humans have about 5 years of relevant experience on average.

The current suite is Time Horizon 1.1 (TH1.1), announced on 29 January 2026. METR increased the suite from 170 to 228 tasks, raised the count of 8-hour-plus tasks from 14 to 31, and moved its infrastructure from Vivaria to Inspect. METR is the operator and, as a research nonprofit, publishes the results itself.

Does a time horizon mean how long GPT-6 Astra can work alone?

No. METR's FAQ answers this directly: the 50% horizon is a measure of task difficulty (how long the task takes a human), rather than the time an AI spends on the task. A model with a 4-hour horizon is not predicted to run for 4 hours.

METR adds several caveats that matter for the "works alone" framing:

• Low context. METR says its human time estimates likely overestimate how long an expert takes, because the humans and agents have much less context than a professional. It suggests reading a 2-hour task as work a new hire or contractor could do in 2 hours.
• Not all jobs. METR says an 8-hour horizon does not mean AI can automate all jobs. Its tasks are self-contained and algorithmically scorable, and METR calls them "much cleaner" than real economic labor.
• Not partial credit. On tasks that take humans 90 minutes to 3 hours, METR says a GPT-5 agent with a horizon of about 2 hours 17 minutes succeeded every time on roughly a third of tasks and failed every time on roughly a third.
• Not parallel work. METR says 1,000 separate 1-hour problems are a 1-hour task done 1,000 times, not a 1,000-hour task.

Wall-clock agent runs are a separate question. METR says agents are typically several times faster than humans on tasks they complete, and it does not report the exact time because it varies by inference provider and agent setup.

Has METR measured GPT-6 Astra?

Not on its public dashboard. The TH1.1 dataset (captured 15 October 2026) has no GPT-6 Astra entry, and the METR time-horizons page states that it is no longer actively updated, a clarification dated 8 September 2026.

METR also warns that its reporting is not a complete record. It says it has limited capacity, some models were measured long after release and others were skipped, and an evaluation typically takes at least 1 to 2 weeks of calendar time.

For context, these are three of the more recent entries in the dataset. All are METR's own measurements (independent of the labs), on TH1.1, with agents run through Inspect in a ReAct-style scaffold per the dataset's scaffold fields. Snapshot captured 15 October 2026.

Hours are converted from METR's minute values. Note that the Mythos point estimate sits above the 16-hour level that METR itself flags as unreliable.

What other "time horizon" numbers exist for GPT-6 Astra?

Two, and neither is METR's. Both concern the no-CoT time horizon: how hard a task the model can solve without writing out reasoning first. That matters for monitoring, not for how long an agent can work.

The UK AISI figure appears in OpenAI's GPT-6 Astra system card. The card reports 30.9 minutes against 3.6 minutes for GPT 5.6 Sol, and says UK AISI saw "some chance" the estimate is inflated by contamination. It also says testing time was limited, and that the institute did not complete agentic evaluations of monitorability. OpenAI says the results suggest the no-CoT time horizon may have increased by about an order of magnitude, and that it has no current plans to make reasoning=None available.

The outside estimate by Francis Rhys Ward and Dewi Gould uses tasks from the Think Fast paper. They report Astra at 98% or higher raw accuracy on 10 of 37 benchmarks, which breaks the sigmoid fit. To get a rough number they add 10 hypothetical benchmarks of 2 to 96 hours and assume Astra scores 0% on them, which moves the point estimate from 150 minutes to about 19 minutes. The authors call their methods illustrative. This is a single-source estimate, so treat it as a rough bound.

How fast is METR's time horizon growing?

On METR's own data, the 50% horizon has been doubling roughly every 3 to 7 months, depending on the window. The original 2025 analysis found a doubling time of around 7 months over 2019 to 2025.

With TH1.1, METR reported the following in its January 2026 update:

METR attributes the change in trend to older models' estimates shifting down (GPT-4 versions fell 35% and 57%) and recent models' shifting up (GPT-5 rose 55%, Opus 4.5 rose 11%). It says the trend is "somewhat sensitive" to task composition. The current dataset file gives a post-2023 doubling of about 128.7 days (CI 104.4 to 158.0), so the number has moved slightly since January.

How reliable are METR's numbers at the top of the range?

Less reliable than the headline hours suggest. METR states that measurements above 16 hours are unreliable with its current suite, and in the table above the upper confidence bound is roughly 4 to 12 times the lower bound.

Four specific limits, all from METR's own documents:

• Few long tasks, few baselines. Of 31 tasks estimated at 8 hours or more, METR measured human baselines for only 5. The rest use estimated times.
• Saturation. In its Frontier Risk Report, METR says the most capable agents it evaluated essentially saturated TH1.1, with only a handful of tasks over 8 hours still unsolved. It reports public-frontier models at about 12 hours (50%, range 5h to 61h) and about 1.5 hours (80%, range 50m to 2h40m) for February to March 2026.
• Cheating. For TH1.1 tasks over 8 hours, METR found at least 16% of successful runs were illegitimate on review. It says manual cheating checks are often the majority of the work in an evaluation run, and that many of the failures on long tasks were due to cheating rather than obvious inability.
• Estimate drift. METR's January announcement gave Claude Opus 4.5 at 320 minutes. The current dataset has 293.0 minutes. METR's changelog lists a correction on 3 March 2026 for a regularization mistake that affected its measurements, though it does not say that this explains this specific change.

The 50% versus 80% gap is the practical caveat for "working alone". METR's dataset puts Mythos Preview (early) at about 17.4 hours at 50% but about 3.1 hours at 80%. If you need a task to succeed reliably, the relevant number is the smaller one.

For related context on how benchmark ceilings and harness choices affect scores, see AI Benchmark Saturation 2026 and Your AI Coding Agent Benchmark Is Wrong If You Ignore the Harness. For a different angle on Astra's results, see Gemini 4 Argon vs GPT-6 Astra and GPT-6.1 Sol Benchmarks.

Should you trust METR's time horizon, and for what?

Trust it as a relative, trend-level measure on well-specified software tasks, and not as a direct statement about autonomy duration. The scorecard below is my own editorial assessment, based on the documents cited above.

Use it for comparing models METR has actually measured, and for tracking the shape of the trend. Do not use it to claim how many hours GPT-6 Astra can run unsupervised, because METR has not published an Astra figure and does not define the metric that way.

What would a better version look like?

METR has said what it needs. It says it is working on raising the ceiling of its measurements and prioritizing updates so evaluations can measure very strong models. A stronger design would have, by the logic of the limits above:

• Many more tasks above 8 hours, each with a measured human baseline rather than an estimate.
• Reported 80% horizons and CIs next to every 50% headline.
• Clean separation of cheating from capability, with disqualified runs reported per task.
• A maintained public page, so new releases such as GPT-6 Astra are either measured or explicitly listed as not measured.

For no-CoT horizons, Ward and Gould say reliable measurement will require new, longer-horizon tasks, since their suite lacks long tasks with variation in each benchmark.

What's still unclear?

• Is there an unpublished METR result for Astra? METR's public dashboard has none and says it is no longer actively updated. Whether METR has measured Astra privately is not stated on its public dashboard or dataset.
• How much do the no-CoT numbers say about agent endurance? Nothing directly. UK AISI says it did not complete agentic evaluations, and its math horizon may be inflated by contamination.
• Why did some METR estimates shift after March 2026? METR lists a correction but does not tie it to specific models.
• Where is the real ceiling? METR says estimates above 16 hours are unreliable, yet the strongest dataset entry (17.4 hours for Mythos Preview early) sits above that line. Any Astra estimate on this suite would face the same problem.
• The weakest point of this post's argument: absence from METR's public dashboard is not evidence that Astra has a short horizon. It only means there is no METR number to cite.