What is an AI time horizon? 12 hours at 50%, 70 min at 80%
An AI time horizon is the length of task, measured by how long a human expert needs, that an AI model completes with a given probability. METR, a US nonprofit that evaluates AI systems, introduced it in a March 2025 paper and reports it at 50% and at 80% success. It does not mean the AI works alone for that long. It expresses task difficulty in human time. In METR public data read on November 20, 2026, Claude Opus 4.6 has a 50% horizon of about 12 hours and an 80% horizon of about 1 hour 10 minutes, and the 95% confidence interval on the 50% figure runs from 5 hours 17 minutes to 60 hours 34 minutes. The trend has doubled about every 6 to 7 months since 2019 and about every 4 months since 2023. The tasks lean heavily toward software, machine learning and security, and METR says measurements above 16 hours are unreliable with its current task suite.
The three lines
- Meaning — the human-expert time of tasks a model finishes with 50% (or 80%) probability; a unit of task difficulty, not of how long the AI runs
- Numbers — in METR data read November 20, 2026, Claude Opus 4.6 is at about 12 hours at 50% and about 1 hour 10 minutes at 80%; the trend has doubled about every 4 months since 2023
- Pitfalls — 50% success is not working reliability, the tasks are mostly software, and intervals are wide (5 to 60 hours); METR calls measurements above 16 hours unreliable
Key questions
- what is an AI time horizon
- **It is the length of task, measured in the time a human expert needs, that an AI model completes with a given probability. METR introduced it in March 2025. It measures task difficulty, not how long the AI works.** | Step | What happens | |---|---| | Task suite | 228 software, machine-learning and security tasks (Time Horizon 1.1) | | Human baseline | Experts attempt the same tasks; the times of successful attempts are combined by geometric mean | | AI runs | 6 independent runs per task, each scored as success or failure | | Curve fit | A logistic curve is fitted to success probability against the log of human time | | Read the horizon | The human time where the curve crosses 50% or 80% is the time horizon |
- METR time horizon by model
- **In METR public data (Time Horizon 1.1, read November 20, 2026), GPT-4 is at 4 minutes, GPT-5 at 3 hours 23 minutes and Claude Opus 4.6 at 11 hours 59 minutes. At 80% the figures are far shorter: 53 seconds, 38 minutes and 1 hour 10 minutes.** | Model (release) | 50% horizon | 80% horizon | |---|---|---| | GPT-4 (March 2023) | 4 min | 53 sec | | GPT-4o (May 2024) | 7 min | 1 min 16 sec | | Claude 3.7 Sonnet (February 2025) | 1 hr | 12 min | | GPT-5 (August 2025) | 3 hr 23 min | 38 min | | Claude Opus 4.5 (November 2025) | 4 hr 53 min | 49 min | | Claude Opus 4.6 (February 2026) | 11 hr 59 min | 1 hr 10 min | | GPT-5.4 (March 2026) | 5 hr 42 min | 54 min |
- AI time horizon doubling every 7 months
- **The original METR paper of March 2025 said about every 7 months since 2019. With the larger Time Horizon 1.1 task suite, the all-time figure is about 6 to 6.5 months and the figure since 2023 is about 4 months.** | Period | Time Horizon 1.0 | Time Horizon 1.1 (announced January 2026) | 1.1 current data (read November 20, 2026) | |---|---|---|---| | All-time (2019 on) | 195.8 days | 196.5 days | 187.8 days | | Since 2023 | 165.3 days | 130.8 days | 128.7 days (104 to 158) | | Since 2024 | 108.9 days | 88.6 days | Not in the data file |
An AI time horizon is the length of task, measured by how long it takes a human expert, that an AI model completes with a given probability. METR, a US nonprofit that evaluates AI systems, introduced it in a March 2025 paper. Saying a model has a 2-hour time horizon does not mean it works for two hours. It means the model succeeds about half the time on tasks that take a human expert about two hours. A benchmark score such as 84% is hard to translate into what kind of work a model can do. The time horizon converts that into a unit people already understand, and its chart has become the most cited picture in arguments and forecasts about how fast AI is improving.
This article explains how a time horizon is calculated, why the 50% and 80% horizons are so far apart, where the claim of doubling every 7 months came from and how it has since been revised, and what to watch for when reading the number. It ends with two figures from November 2026, an internal OpenAI statistic and a forecast from the AI Futures Project, and sets out how each differs from a time horizon. Every figure carries the date it was measured or checked.
1. What a time horizon is: definition and calculation
The METR definition: the task-completion time horizon is the task duration at which an AI agent is predicted to succeed with a given level of reliability, where duration is measured by human expert completion time. The calculation has five steps.
| Step | What happens | Detail |
|---|---|---|
| Task suite | Tasks of different lengths are assembled | RE-Bench, HCAST and a set of short novel tasks, mostly software engineering, machine learning and cybersecurity. 170 tasks in the March 2025 version, 228 in the January 2026 version (Time Horizon 1.1) |
| Human baseline | A human time is set for each task | Contracted experts attempt the task with the same instructions and tools; the times of successful attempts are combined by geometric mean |
| AI runs | The model is given tools and the same tasks | By the METR description, 6 independent runs per task, with human review for reward hacking |
| Curve fit | Success probability is modeled against human time | A logistic curve with the log of human time on the horizontal axis and success probability on the vertical |
| Read the horizon | The points where the curve crosses 50% and 80% are read off | The human time at those points is the 50% horizon and the 80% horizon |
The fourth step is the core. A model solves nearly all short tasks and almost none of the long ones. The METR post of March 2025 said models then had near-100% success on tasks under about 4 minutes for a human and under 10% on tasks over about 4 hours. Somewhere in between the success rate passes 50%, and the time horizon marks that point as a task length.
Three clarifications. First, it has nothing to do with how long the AI actually takes. METR says agents are typically several times faster than humans on tasks they complete. Second, a task is a coherent unit of work that cannot be trivially split. Solving 1,000 one-hour math problems in a row is a one-hour task done 1,000 times, not a 1,000-hour task. Third, a 2-hour horizon does not mean every 2-hour task has a 50% success rate. METR explains that on tasks taking a human 90 minutes to 3 hours, GPT-5 always succeeded on about a third, always failed on about a third, and sometimes succeeded on the rest.
2. The 50% and 80% horizons: measurements by model
METR publishes both thresholds. The 50% horizon tolerates failing half the time. The 80% horizon requires four successes in five. The values below are from the Time Horizon 1.1 data file published by METR, downloaded on November 20, 2026. Figures in parentheses are bootstrap confidence intervals calculated by METR. Release dates are as recorded in the data file.
| Model | Released | 50% horizon (interval) | 80% horizon (interval) |
|---|---|---|---|
| GPT-2 | February 14, 2019 | About 3 sec | About 1 sec |
| GPT-3.5 Turbo Instruct | March 15, 2022 | About 36 sec | About 15 sec |
| GPT-4 | March 14, 2023 | 4 min (1.9 to 8 min) | 53 sec |
| GPT-4o | May 13, 2024 | 7 min (4 to 13 min) | 1 min 16 sec |
| Claude 3.5 Sonnet (new) | October 22, 2024 | 21 min (10 to 41 min) | 2 min 36 sec |
| o1 | December 5, 2024 | 39 min (21 min to 1 hr 5 min) | 7 min |
| Claude 3.7 Sonnet | February 24, 2025 | 1 hr (33 min to 1 hr 44 min) | 12 min |
| o3 | April 16, 2025 | 2 hr (1 hr 15 min to 3 hr 11 min) | 30 min |
| GPT-5 | August 7, 2025 | 3 hr 23 min (1 hr 53 min to 6 hr 46 min) | 38 min |
| Gemini 3 Pro | November 18, 2025 | 3 hr 44 min (2 hr 20 min to 6 hr 19 min) | 54 min |
| Claude Opus 4.5 | November 24, 2025 | 4 hr 53 min (2 hr 42 min to 10 hr 24 min) | 49 min |
| GPT-5.2 | December 11, 2025 | 5 hr 52 min (3 hr 18 min to 13 hr 35 min) | 1 hr 6 min |
| Claude Opus 4.6 | February 5, 2026 | 11 hr 59 min (5 hr 17 min to 60 hr 34 min) | 1 hr 10 min (27 min to 2 hr 50 min) |
| Gemini 3.1 Pro | February 19, 2026 | 6 hr 24 min (3 hr 54 min to 11 hr 35 min) | 1 hr 30 min |
| GPT-5.4 | March 5, 2026 | 5 hr 42 min (3 hr 7 min to 12 hr 49 min) | 54 min |
| Claude Mythos Preview (early) | April 7, 2026 | 17 hr 25 min (8 hr 29 min to 55 hr 4 min) | 3 hr 6 min (1 hr 37 min to 6 hr 39 min) |
Three things stand out.
First, the 80% horizon is usually a quarter to an eighth of the 50% horizon. For Claude Opus 4.6 the gap is tenfold. A model that handles 12-hour tasks at the 50% standard can be trusted only up to 70-minute tasks at the 80% standard. Both statements about the same model are true: it does more than half a day of work, and it does a little over an hour of work. That is why it matters which number is being quoted.
Second, the intervals are wide. The point estimate for the 50% horizon of Opus 4.6 is about 12 hours, and the interval runs from 5 hours 17 minutes to 60 hours 34 minutes, more than an elevenfold spread. The stronger a model gets, the fewer tasks it fails, and the less data there is to pin down the right-hand end of the curve.
Third, the table stops here. When METR added Mythos Preview (early) on May 8, 2026, it attached a warning that measurements above 16 hours are unreliable with the current task suite. On September 8, 2026 it stated that the page is no longer actively updated. Time horizons for later models are not on the page. METR also says the list is not a complete record of the most capable models, and that limited capacity meant some were measured long after release and others skipped.
3. How fast it has grown: doubling every 7 months, and after
The metric became famous for its trend more than for any single value. On March 19, 2025, METR reported that the length of tasks generalist AI agents complete at 50% reliability had been doubling approximately every 7 months for six years. The same post noted that fitting only the 2024 and 2025 data gave a steeper trend, bringing the projected arrival of month-long tasks forward by about 2.5 years. The widely repeated extrapolation, that AI could carry out month-long projects on its own by the end of the decade if the trend held, comes from here.
On January 29, 2026, METR released Time Horizon 1.1, expanding the suite from 170 to 228 tasks: 73 added, 15 removed and 53 updated. Tasks taking a human 8 hours or more rose from 14 to 31, and evaluation moved from an in-house platform to Inspect, the open-source framework from the UK AI Security Institute. Doubling times changed as follows.
| Period | Time Horizon 1.0 | Time Horizon 1.1 (announced January 29, 2026) | 1.1 current data (read November 20, 2026) |
|---|---|---|---|
| All-time (2019 on) | 195.8 days (about 6.4 months) | 196.5 days (about 6.5 months) | 187.8 days (about 6.2 months) |
| Since 2023 | 165.3 days (about 5.4 months) | 130.8 days (about 4.3 months) | 128.7 days (about 4.2 months; interval 104 to 158 days) |
| Since 2024 | 108.9 days (about 3.6 months) | 88.6 days (about 2.9 months) | Not in the data file |
The all-time trend barely moved. Since 2023 it sped up to a doubling about every 4 months. Changing the tasks pushed older models down and newer models up. In the 1.1 announcement, the 50% horizon of GPT-4 (1106) fell 57%, from 8.5 to 3.6 minutes, and that of GPT-5 rose 55%, from 138 to 214 minutes. The models did not change. The ruler did, and the numbers moved that much, which says something about the nature of the metric.
The current data differ again from the 1.1 announcement because METR corrected a regularization mistake on March 3, 2026. According to the METR research note of March 20, 2026, the fix lowered the 50% horizons of recent models by up to 20%. The doubling times in the current data exclude measurements whose central estimate is above 16 hours.
4. Pitfalls when reading the number
| Pitfall | Detail | Basis |
|---|---|---|
| 50% is not working reliability | Half the attempts fail. To know what can be handed off and left alone, look at the 80% horizon or higher, and the 80% horizon is a quarter to a tenth of the 50% one | Table in section 2. METR says it does not report a 99% horizon because that would need many more tasks |
| The tasks are mostly software | Software, machine-learning and security tasks dominate, and they are well specified and automatically scored. METR says it found similar exponential trends in other domains with very different absolute values | Task description and FAQ on the METR page |
| The human baseline is a person without context | The baseliners are professionals with about 5 years of experience on average, seeing the task for the first time. METR suggests thinking of a 2-hour task as what a new hire or freelance contractor could do in 2 hours. Someone who knows the project would be faster, so human times are probably overestimates | METR page |
| Long tasks often have no measured human time | Of the 31 tasks of 8 hours or more in 1.1, only 5 have human baselines. The UC Berkeley paper notes that measured human times were unobtainable or invalid for 29% of tasks, where an expert estimate was used | 1.1 announcement, arXiv 2610.12466 |
| Intervals are wide | The 50% interval for Opus 4.6 runs from 5 to 60 hours. METR calls measurements above 16 hours unreliable | METR data file and page |
| Results depend on modeling choices | Reasonable alternative curve shapes cut the 50% horizons of then-recent models by up to 35% and raise 80% horizons by up to 100%. Dropping the public tasks takes the 50% horizon of Opus 4.6 from 11 hours 59 minutes to 7 hours 11 minutes | METR research note (March 20, 2026) |
| Clean tasks are not real work | METR found that agents do worse on messier tasks, and that performance drops substantially when work is scored holistically, not algorithmically. It gives this as a reason an 8-hour horizon does not mean jobs can be automated | FAQ on the METR page |
There is outside statistical criticism as well. In a paper posted to arXiv on October 8, 2026, On the estimation and validity of AI time horizons, Drew T. Nguyen and William Fithian of the UC Berkeley Department of Statistics reanalyzed the public METR data (228 tasks, 26 AIs). Their target is the assumption that the difficulty of a task for an AI rises linearly with the log of human time. Relaxing it with splines and item-response theory, they found the curve nearly flat between about 2 and 30 minutes of human time. Their reading: a jump in time horizon from 3 to 30 minutes is much easier than one from 30 minutes to 5 hours, though both are tenfold. A doubling is not the same amount of progress everywhere on the scale.
The paper does not dispute the exponential trend itself. The authors write that the observed exponential growth is empirical and that they do not challenge it. Their proposal is to stop reading a time horizon as a single point and to read it alongside diagnostic plots showing how human time maps to AI difficulty. It is a preprint and has not been peer reviewed. METR has acknowledged the underlying problem too: its March 2026 note says that as the task suite saturates, the results become more sensitive to analysis choices.
5. Two figures from November 2026: are they the same measurement?
Two things brought talk of AI doing tasks that take hours back into circulation in November 2026. Both sound like time horizons. They measure different things.
The internal OpenAI statistic. In a post published on November 18, 2026, OpenAI was reported to have said that the share of internal research tasks taking a human 2 to 4 hours that agents completed without human intervention rose from 23% in March 2026 to 46% in October. For 4-to-8-hour tasks the figures were 13% and 35%. Results varied sharply by type of work: 79% for monitoring training runs, 25% for writing inference and serving code. The OpenAI post itself could not be opened, so this rests on three outlets that agree.
The AI Futures Project forecast. The authors of the AI 2027 scenario wrote in a quarterly update dated November 10, 2026 that all of their forecasters now expect a superhuman AI researcher is likely to arrive some time in 2027, assuming no substantial deliberate slowdown. Eli Lifland gives a median of August 2027. That Daniel Kokotajlo moved his median from December 2028 to May 2027 was confirmed in secondary coverage only. The main evidence for the update is not time horizons. It is a separate METR report estimating that AI R&D at Anthropic is progressing at about twice its historical pace (range 1.3 to 3 times), with roughly 80% of that acceleration attributed to AI. The authors say most of the informative evidence was redacted and call their analysis highly preliminary. Their forecasting model has, however, used METR time horizon research as one of its inputs since its December 2025 version.
| METR time horizon | Internal OpenAI statistic (published November 18, 2026) | AI Futures forecast (November 10, 2026) | |
|---|---|---|---|
| What is measured | Success rate is fixed at 50% or 80% and the task length is solved for | Task length is fixed (2 to 4 hours and so on) and the zero-intervention completion rate is counted | Not a measurement. A probability distribution over when AI of a given capability arrives |
| Tasks | A fixed suite of 228 tasks; every model gets the same ones | Tasks researchers actually delegated to agents at work; the mix changes over time | Not applicable |
| Human time | Measured by experts attempting the task (estimated for some) | An estimate of human time; the method could not be confirmed | Not applicable |
| Success criterion | Automatic scoring defined per task | Finished with no human intervention; how this was judged could not be confirmed | Not applicable |
| Who measures | A third-party nonprofit, with data and code public | Company figures, with no reproduction package | An independent research group |
| Weak point | Limited to clean software tasks; unreliable above 16 hours | Selection effects. By the coverage, OpenAI itself cautions that researchers may have changed what they delegate | Depends on model assumptions and redacted evidence |
The two numbers cannot be read across. Translated loosely into time-horizon terms, 46% on 2-to-4-hour tasks sounds like a 50% horizon of around 3 hours, yet on the clean METR tasks public models from early 2026 were already measured at 50% horizons of about 6 to 12 hours. Real work comes with context and mess, and the bar is not passing a scorer but never needing a person to step in. The OpenAI figure shows performance in an actual workplace, which a time horizon cannot. The time horizon holds every model to the same ruler, which internal figures cannot. They are better read as filling each other's gaps.
6. Frequently asked questions
Does a 12-hour time horizon mean the AI works alone for 12 hours? No. It means the model succeeds about half the time on tasks that take a human expert about 12 hours. The time the AI actually takes is a separate matter, and METR says agents are typically several times faster than humans on tasks they complete.
Which should I look at, the 50% or the 80% horizon? It depends on the question. For where the frontier of capability sits, the 50% horizon. For the range of work that needs less human checking, the 80% horizon. METR says the trends for the two are very similar: the levels differ and the growth rate is the same.
Does an 8-hour horizon mean AI can replace a day of work? METR says no, for three reasons. The human baseline is a person without context. The tasks lean toward software. And most jobs are not made of well-specified tasks that can be scored automatically.
Why is there no 99% horizon? METR explains that measuring it accurately would take many more tasks, that those tasks would have to be very short and so harder to baseline reliably, and that the result would be much more sensitive to a single broken task.
If the trend continues, when does AI do month-long work? The original March 2025 paper extrapolated to around the end of the decade if the trend held. That is an extension of a line, not a measurement. The current task suite runs out at about 16 hours, and measuring beyond that needs longer tasks and people who have actually completed them. Whether the trend has bent or steepened above that point cannot be told with the current ruler.
What should be read alongside a time horizon? That one model can score very differently on different tests is already well documented for coding benchmarks. Tests built from real user change requests, completion rates that companies publish from internal use, and the share of AI-written changes that maintainers actually accept are different kinds of evidence, and they belong side by side.
7. What remains open or unconfirmed
| Item | Confirmed | Not confirmed |
|---|---|---|
| Definition and method | METR page and the March 2025 paper | Nothing outstanding |
| Measurements by model | METR data file (Time Horizon 1.1), downloaded and read on November 20, 2026 | Time horizons for models released after May 2026, which are not on the METR page |
| Mythos Preview (early) | 50% horizon about 17 hours 25 minutes, 80% horizon about 3 hours 6 minutes | The 50% figure is above the 16 hours METR calls unreliable |
| Doubling time | About 7 months in the original paper; 196.5, 130.8 and 88.6 days in the 1.1 announcement; 187.8 and 128.7 days in current data | The current value of the trend since 2024; the trend above the 16-hour limit |
| METR page no longer updated | Stated on the page on September 8, 2026 | The reason, and any schedule for a successor task suite |
| Statistical criticism | The abstract and opening pages of arXiv 2610.12466; the flat region from 2 to 30 minutes | Peer review; the full set of revised estimates by model |
| Internal OpenAI statistic | 23% to 46% and 13% to 35%, reported consistently by three outlets | The post itself could not be opened; measurement period, how human time was estimated, how completion was judged |
| AI Futures forecast | The November 10, 2026 post, the statement that all forecasters expect 2027, the Lifland median of August 2027 | The size of the move for Kokotajlo comes from secondary coverage; the later part of the post was not read |
| Time horizons in other domains | METR says it found exponential trends in other domains | Absolute values by domain, which were not checked for this article |
The debate over AI accelerating AI research is covered in our October 16, 2026 explainer on recursive self-improvement. Why one model gets different scores on different tests is in our August 17, 2026 explainer on coding benchmarks and our November 10, 2026 report on AI coding agent success rates. How a model with tools carries out a task is in our August 26, 2026 explainer on AI agents, and the definitions of the goal these metrics point toward are in our September 4, 2026 explainer on AGI. The internal OpenAI figures are covered in detail in our separate November 20, 2026 report on the OpenAI research acceleration post.
Sources
- METR — Task-Completion Time Horizons of Frontier AI Models (last updated 2026-09-08)
- METR — Time Horizon 1.1 raw results (benchmark_results_1_1.yaml, read 2026-11-20)
- METR — Time Horizon 1.0 raw results (benchmark_results_1_0.yaml, read 2026-11-20)
- METR — Measuring AI Ability to Complete Long Tasks (2025-03-19)
- arXiv 2503.14499 — Measuring AI Ability to Complete Long Software Tasks (2025-03-18)
- METR — Time Horizon 1.1 (2026-01-29)
- METR — Impact of modelling assumptions on time horizon results (2026-03-20)
- arXiv 2610.12466 — On the estimation and validity of AI time horizons: a statistical look at the METR plot (2026-10-08)
- AI Futures Project — Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All (2026-11-10)
- Crypto Briefing — AI 2027 authors now expect a superhuman AI researcher in 2027 (2026-11-11)
- Wikipedia — AI Futures Project (last edited 2026-11-01)
- AI News Weekly — OpenAI Says Coding Agents Finish 46% of 2-4 Hour Research Tasks (2026-11-19)
- explainx.ai — OpenAI Agents Hit 46% on 2–4h Research Tasks (2026-11-19)
- Blockchain.News — OpenAI Reports Uneven Progress in AI-Driven Research Acceleration (2026-11-19)
- AI 2027 Tracker — METR doubling (last updated 2026-11-07)