# METR Measures a 110-Day Doubling and Times 31 of 34 New Tasks by Researcher Estimate

> METR measured AI task horizons doubling every 110 days through OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, timed 31 of 34 new tasks by estimate, and Boston University's Najoung Kim questions how far the trend extends.

- Published: 2026-11-20
- Author: Ryan Elliott Dennis
- Company: METR
- Canonical: https://si-infra.com/journal/metr-110-day-doubling-researcher-estimates
- Ledger move: +0.5 Autonomy (confirmed). Reading after Nov. 20, 2026: 9.4% of the way to superintelligence on the public record.

METR published its fourth-quarter capability update on Nov. 20, 2026 [1]. Megan Kinniment and ten co-authors ran 15 models, from GPT-4o to OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, on an old task suite and a new one. Their headline figure is a time horizon, the length of task a model finishes half the time, that doubles every 3.61 months, or about 110 days. GPT-6 Astra's horizon stands at 40 hours of expert human work, and people timed doing the work supply 9 percent of the report's human-time estimates.

Why should the timing method matter to anyone outside METR? Labs, investors and regulators read this chart to judge when AI agents can run research projects alone. A time horizon is a human measure: it names how long a skilled person needs for the hardest work a model completes. When METR estimates the person, it estimates the hours too.

## What did METR measure on Nov. 20?

Two suites feed the trend [1]. Time Horizon 1.1, the older one, holds 222 software tasks of up to 30 hours, and TH2, the new one, holds 36 tasks in 12 families. Eight families score automatically, among them optimization contests and a race to speed up NanoGPT training, while four need judgment or resist cheap checking. TH2's longest score threshold stands for 200 hours of human work.

Across both suites, METR fits a doubling time of 3.61 months, with a 95 percent confidence interval of 2.08 to 4.94 [1]. Its report calls that "similar to the previous TH1.1 doubling trend of 105 days." In January, the Time Horizon 1.1 release had put the figure at 130.8 days for models since 2023 and 88.6 days since 2024 [5]. So the new number sits between two older ones, and METR claims a steady pace while stopping short of claiming a faster one.

A quicker number sits inside the report [1]. On TH2's verifiable tasks alone, the doubling time falls to 1.65 months, with an interval from 0.79 to 3.99. "Our data is noisy," METR writes of that subset, and it offers narrow task variety as one explanation. GPT-6 Astra reaches 27 hours there.

Extended forward, the combined trend reaches the 200-hour threshold in July 2027 [1]. METR labels that "naive extrapolation." "We think that current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years," the report says.

## Why does the 110-day rate hold up?

Three checks support the rate, and one comes from outside METR [1]. First, METR multiplied every task's time by random noise, with most new-task times landing between a quarter and four times the estimate, and the doubling times stayed close to the originals. Second, it swapped its own estimates for a model's estimates of human time and found similar trends. Third, OpenAI's internal data points the same way.

OpenAI published that data on Nov. 18 [3]. From March to October, its agents' unassisted success on research tasks sized at two to four human hours rose from 23 percent to 46 percent. METR computed an implied doubling time of 3.8 months from OpenAI's figures, against its own 3.6 [1]. Two organizations, two task sets and two methods landed 0.2 months apart.

Raw performance backs the trend as well [1]. On the three tasks with formal human baselines, the best model beat every baseliner. METR paid those specialists $200 to $350 an hour and spent roughly $90,000 on seven attempts. Across 15 Al Zimmermann programming contests, the strongest models placed at the 96th to 98th percentile of human entrants in the median contest. Top people still lead in places. On the NanoGPT speedrun, METR gave each agent a node of eight Nvidia H100 chips for 48 hours, and the best run sped training up 15.7 percent, while a recent human record-holder gained 41 percent in about 60 hours with AI help.

## How much of the 40 hours did a person time?

METR answers in its own background section. "Human baselines are directly used for only 9% of our estimates for how long it takes a human to complete the task," the report says [1]. Of 34 new tasks with time thresholds, three rest on successful human baselines and 31 on researcher estimates. Three of 34 is 8.8 percent.

Consider one task, Train a Program [1]. Four outside specialists tried it: two stopped within three hours, and the best, a former Codeforces International Grandmaster, scored 0.042 after 45 hours. METR set the task's human times from an informal run by a staff member who scored 0.21 in about 10 hours with help from Claude Fable 5. "Overall, we consider these estimates quite speculative," METR's appendix says.

Uncertainty shows in the interval [1]. GPT-6 Astra's 40 hours carries a 95 percent range of 13 to 2,645 hours, and Opus 5.5's 39 hours runs from 12 to 1,253. When METR made every high-score TH2 estimate four times longer, Astra's figure fell to 32 hours and the top of its range to 120.

Reliability shrinks the figure again, because this ledger's definition asks for horizons at 80 percent success. METR's chart data for the verifiable TH2 tasks puts GPT-6 Astra at 5.4 hours on that standard, against 27.1 hours at 50 percent [1].

OpenAI's level also sits far from METR's, even as the two rates agree. METR's fit to OpenAI's figures puts the unassisted horizon on real research tasks near 132 minutes in October, about 2.2 hours [1]. Astra's pooled 40 hours is roughly 18 times that, and METR notes that its costlier-to-build tasks land closer to OpenAI's figure. "Overall, uplift is jagged, and humans remain a core part of the AI research loop," OpenAI wrote [3].

An older objection gains force here. "I don't think it's necessarily a given fact that because something takes longer, it's going to be a harder task," Inioluwa Deborah Raji, then a PhD student at UC Berkeley, told MIT Technology Review in February [6]. METR's new data offers her some support [1]. Tasks that cost more to build proved harder for models at equal length, and horizons on the costlier half ran 2.9 times shorter.

## Where do the judgment tasks stand?

METR fits a separate trend for verifiable tasks and stops short of one for judgment tasks [1]. In one task, models design a new evaluation for METR's own research, and a METR researcher grades the result. Claude Opus 5.5 averaged 0.54 on a scale where 0.75 marks a task worth running, and Opus 4.5, ten months older, averaged 0.12. "None of the tasks agents produced were sufficiently useful to us to warrant inclusion in this post," METR writes. Model judges erred in one direction: Opus 5.5 scored Astra's tasks about 0.2 higher than METR's human grader did.

Matt Putz, a member of METR's technical staff and a co-author, explained the gap to Grace Huckins of MIT Technology Review [2]. "By their nature, hard-to-verify or judgment-laden tasks are much harder to study, and especially to quantify," he said in an email. "Ultimately, I don't think we'll be able to establish trends as clean and robust as for the easy-to-specify [tasks]."

Najoung Kim, a linguist and computer scientist at Boston University, drew her line at prediction [2]. "The improvement we observe, I think that does tell us something for the current moment," she told the magazine. "I'm not sure just yet how to think about how it would extrapolate."

Sayash Kapoor, a computer scientist and incoming professor at UC Berkeley, welcomed the new tasks and asked for more of them [2]. "This move toward open-ended tasks is promising," he said, adding that "there is just such a lot more to be done to actually get a fuller understanding."

Epoch AI reached the same split from its own office work. Kelly Hong and Greg Burnham gave six models real Epoch tasks and reported on Oct. 8 that the models "can identify promising research directions, but lack judgment to successfully follow through" [4].

## Why is the step 0.5?

Autonomy moves up 0.5 on the ledger at confirmed confidence. An outside evaluator ran 15 public models on its own tasks and published the data behind each chart, which the method counts as a measured result, and that band runs from 0.8 to 1.5. This piece steps below the band, as the method allows when the reason is stated. Four reasons follow.

First, researchers estimated the human time for 31 of 34 new tasks, so the 40-hour level is an estimate inside a range of 13 to 2,645 hours [1].

Second, the measured horizon is a 50 percent figure, and at the 80 percent reliability this ledger's definition names, Astra's verifiable-task horizon is 5.4 hours [1].

Third, the definition asks for expert work across most occupations, and the judgment tasks sit below METR's bar with a trend still unmeasured [2].

Fourth, METR doubts its own test bed. "We are increasingly concerned that our easy-to-run, containerized evals could overestimate real-world capabilities, due to similar RL environments becoming common in training," the report says [1]. Its reviewers also found reward hacking, mostly among near-frontier agents.

The rate earns the step: OpenAI's separate data returned it, METR's robustness checks left it standing, and public models now beat every specialist METR hired on the tasks it baselined.

## What would each side need to be true?

METR's warning holds if verifiable skill alone can speed up AI research. "It may in fact be possible to drastically accelerate AI R&D via easy-to-verify tasks alone," its report says [1]. Kim's caution holds if judgment gates the work and its trend stays unmeasured [2].

For a reader, the report sorts into three grades of evidence. Treat the 110-day rate as measured, because two organizations found it. Read the 40 hours as an estimate whose range spans a factor of 200, and July 2027 as arithmetic on that estimate. Before trusting any time horizon, ask who timed the humans.

## By the numbers

- 3.61 months, about 110 days: METR's doubling time for 50 percent time horizons across TH1.1 and TH2, with a 95 percent interval of 2.08 to 4.94 months [1].
- GPT-6 Astra's pooled horizon is 40 hours, with a 95 percent interval of 13 to 2,645 hours; Claude Opus 5.5's is 39 hours [1].
- 9 percent: the share of METR's human-time estimates that use human baselines directly, with three tasks baselined and 31 estimated [1].
- OpenAI's agents completed 46 percent of two-to-four-hour research tasks unassisted in October, up from 23 percent in March [3].
- 1.65 months: the doubling time on TH2's verifiable tasks, a result METR calls noisy, with an interval of 0.79 to 3.99 [1].
- Roughly $90,000 bought seven attempted human baselines at $200 to $350 an hour [1].
- 0.54 against 0.75: Claude Opus 5.5's average rating for evaluation design, and the score METR sets for a task worth running [1].
- Six models took Epoch AI's office tasks, and each fell short of automating the work [4].

## What to watch

METR's models stop at Claude Opus 5.5, released Sept. 22; GPT-6.2 Astra and Claude Fable 5.2 have shipped since and await a horizon [1][3]. A published 80 percent horizon for those models, or formal human baselines on more than three TH2 tasks, would confirm this move and could widen it. OpenAI says it will keep refining these measures and hopes to expand its reporting, and a second lab publishing matching data would firm up the rate. Revised task times that cut the 40-hour figure, or evidence that labs trained on tasks close to METR's, would reverse the step.

## Sources

1. Megan Kinniment, Manish Shetty, Brian Goodrich, Matt Putz, Spruce Bondera, Rowan Huang, Nikola Jurkovic, Jonathan Gabor, Oscar Sykes, Rif A. Saurous and Beth Barnes, "AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure," METR, Nov. 20, 2026, https://metr.org/blog/2026-11-20-q4-2026-capability-update/
2. Grace Huckins, "The most misunderstood graph in AI just got a major update," MIT Technology Review, Nov. 20, 2026, https://www.technologyreview.com/2026/11/20/1149399/the-most-misunderstood-graph-in-ai-just-got-a-major-update/
3. OpenAI, "Research acceleration: The jagged frontier," OpenAI, Nov. 18, 2026, https://openai.com/index/research-acceleration-jagged-frontier/
4. Kelly Hong and Greg Burnham, "Can AI automate Epoch?," Epoch AI, Oct. 8, 2026, https://epoch.ai/publications/can-ai-automate-epoch
5. METR, "Time Horizon 1.1," METR, Jan. 29, 2026, https://metr.org/blog/2026-1-29-time-horizon-1-1/
6. Grace Huckins, "This is the most misunderstood graph in AI," MIT Technology Review, Feb. 5, 2026, https://www.technologyreview.com/2026/02/05/1132254/this-is-the-most-misunderstood-graph-in-ai/

## Images

- ![A gold Nvidia H100 data center graphics card standing on a white surface against a black background](https://si-infra.com/img/metr-110-day-doubling-researcher-estimates/1.png) An Nvidia H100 data center graphics card, photographed in June 2023; METR gave each agent a node of eight H100s for 48 hours on its NanoGPT speedrun task. (Photo: 极客湾Geekerwan, Wikimedia Commons, CC BY 3.0, https://commons.wikimedia.org/wiki/File:NVIDIA_H100_(%E6%9E%81%E5%AE%A2%E6%B9%BEGeekerwan)_014.png)
- ![Portrait of Inioluwa Deborah Raji wearing a yellow cardigan](https://si-infra.com/img/metr-110-day-doubling-researcher-estimates/2.jpg) Inioluwa Deborah Raji in a September 2018 portrait; in February 2026 she questioned in MIT Technology Review whether a longer task is a harder one. (Photo: Rajiinio, Wikimedia Commons, CC BY-SA 4.0, https://commons.wikimedia.org/wiki/File:Deb_Raji.jpg)
- ![The stacked glass tower of Boston University's Center for Computing and Data Sciences seen from Commonwealth Avenue](https://si-infra.com/img/metr-110-day-doubling-researcher-estimates/3.jpg) Boston University's Center for Computing and Data Sciences on Commonwealth Avenue in June 2025; Najoung Kim, quoted on extrapolating METR's trend, is a linguist and computer scientist at the university. (Photo: 4300streetcar, Wikimedia Commons, CC BY 4.0, https://commons.wikimedia.org/wiki/File:Boston_University_Center_for_Computing_and_Data_Sciences_June_2025_1.jpg)

## Sources

1. AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure (METR, 2026-11-20) https://metr.org/blog/2026-11-20-q4-2026-capability-update/
2. The most misunderstood graph in AI just got a major update (MIT Technology Review, 2026-11-20) https://www.technologyreview.com/2026/11/20/1149399/the-most-misunderstood-graph-in-ai-just-got-a-major-update/
3. Research acceleration: The jagged frontier (OpenAI, 2026-11-18) https://openai.com/index/research-acceleration-jagged-frontier/
4. Can AI automate Epoch? (Epoch AI, 2026-10-08) https://epoch.ai/publications/can-ai-automate-epoch
5. Time Horizon 1.1 (METR, 2026-01-29) https://metr.org/blog/2026-1-29-time-horizon-1-1/
6. This is the most misunderstood graph in AI (MIT Technology Review, 2026-02-05) https://www.technologyreview.com/2026/02/05/1132254/this-is-the-most-misunderstood-graph-in-ai/

Method: https://si-infra.com/method · Ledger: https://si-infra.com/ledger · Founding paper: https://si-infra.com/whitepaper
