How long a task AI can finish, and three ways the next two years could go
Each dot is a leading model's "time horizon" from METR's own data file. The scale is logarithmic, so steady doubling shows up as a straight line. After April 2026 the lines are projections, not measurements. Hover or tab to a dot for its numbers, switch the scenarios on and off, or set your own doubling time.
Compute hasn't stalled yet. These lines show what happens if it does. Research modelling suggests time-horizon growth tracks compute growth, so halving one roughly halves the other. The blue line assumes AI-accelerated research wins back part of that loss. That's a scenario, and it's labelled as one.
1.97. That's the rate of acceleration METR calculated from Anthropic's private estimates of how quickly its AI research is now moving. Anthropic's public report put the speed-up at less than a factor of two. METR, the independent lab that has tested Claude and GPT models before release since 2023, published its review on 5 November and offered a plainer description: "around 2."
The review came out two days after American voters handed both houses of Congress to the Democrats and, in 38 of 41 local ballot measures, voted to ban or pause data centres. One event is about how fast AI is improving itself. The other is about whether the buildings that make that improvement possible get built. This article puts the two numbers side by side, along with a third that matters more than either: the point where METR's measuring stick stops working.
Around two
Anthropic's safety policy has an early-warning line for automated research. METR calls it the "2x Acceleration Threshold": the point where AI has roughly doubled the pace of the company's progress. It's one of the lines the policy tells the company to watch most closely, because it's the first sign of AI building AI.
Anthropic's public Risk Report for August said acceleration was "less than a factor of 2." METR read both the public report and a private version with more data. Its verdict was careful but pointed. It agreed with Anthropic "that catastrophic risks from July 15th models were low." Then it said that crossing the 2x threshold by 15 July was plausible: "we think it is plausible that it had been." Depending on the definitions and the time period, METR's own estimates ran from 1.3 to 2.8. It also noted that most of the relevant researchers at METR and Anthropic still lean towards the threshold not having been crossed, and it asked Anthropic to present the figures as a range.↗
The sentence that should travel furthest is the one about the future: "we cannot rule out a substantial further acceleration in the pace of capabilities progress over the next 6 months."↗
Where does a number like that come from? Partly from inside the labs. Anthropic's own account, published in September under the title When AI builds itself, lists the evidence. "As of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude." In the second quarter of 2026, engineers merged about eight times as much code per day as in 2024, a figure Anthropic itself says overstates the real gain. On one fixed task, optimising a piece of training code, Claude Opus 4 managed about a threefold speed-up in May 2025. "By April 2026, Claude Mythos Preview was achieving ~52x." A skilled human reaches about 4x in four to eight hours.↗
OpenAI is publishing the same kind of figures. On 6 September it wrote: "According to our measurements, we have now reached the goal, announced last fall, of having an automated research intern by September of this year."↗ It defines that intern as a system that handles well-defined research tasks under human direction, including work that would take a skilled researcher a few days. By mid-August its researchers were running 3.1 agent workdays for every human workday. The next target is an automated AI researcher by March 2028.↗
Both companies are measuring themselves, and both say their measurements are rough. That's why the outside number matters. In May, Anthropic co-founder Jack Clark put the odds of AI systems training their own successors at 60% by the end of 2028, and 30% by 2027.↗ METR's job is to check claims like that from the outside. Six weeks before the November review, it had looked at Claude Opus 5.5 and called it an incremental, on-trend improvement over Claude Fable 5.1, not a jump.↗ On-trend, though, is the point. The trend is the thing that's accelerating.
A ruler made of tasks
METR's best-known measure is the time horizon. It takes a large set of software and research tasks, times how long skilled people need to finish each one, and then finds the task length at which an AI agent succeeds half the time. A model with a one-hour horizon finishes about half of the tasks that take an expert an hour.
The results come from METR's public data file for the current version of the test, Time Horizon 1.1.↗ GPT-4, released in March 2023, comes in at about 4 minutes. Claude 3.7 Sonnet, in February 2025, at about an hour. Claude Opus 4.5, in November 2025, at just under five hours. An early version of Claude Mythos Preview, dated April 2026 in the file, at 1,045 minutes, a little over 17 hours. For models since 2023, the file fits a doubling time of 129 days, with a range of 104 to 158.
Claude Opus 4.6 shows how shaky the top end has become. In February, METR estimated it at about 14.5 hours, with a 95% range of 6 to 98 hours. In March, METR found and fixed a mistake in its statistical model, and wrote that as its task suite saturates, "the results are becoming more sensitive to analysis choices."↗ The current file puts Opus 4.6 at about 12 hours. Anthropic's own write-up quotes METR finding that Mythos Preview "could work for “at least” 16 hours."↗ METR's measurement page is blunter: "Measurements above 16 hrs are unreliable with our current task suite." It also says the page "is no longer actively updated."↗
To get those numbers at all, METR needs access that only the labs can give. Here's what it got this year, and what came with it.
| Engagement | Published | What METR got | Strings attached |
|---|---|---|---|
| Frontier Risk Report (Anthropic, Google, Meta, OpenAI) | 19 May | Each company's most capable internal models, including raw chains of thought. xAI gave public access only. | A pilot; METR says it had more editorial independence than in earlier engagements |
| GPT-5.6 Sol (OpenAI) | 26 June | The final checkpoint, a "railfree" version, raw chain of thought, internal information | NDA; OpenAI's communications and legal teams had to approve the post |
| Claude Opus 5.5 (Anthropic) | 22 Sept | Evidence for an AI R&D assessment, including a report from a separate METR team with higher access | Unpaid agreement; Anthropic could review and edit the summary |
| Anthropic Risk Report review | 5 Nov | Public and private versions of the report | Anthropic had redaction rights and redacted parts of METR's analysis |
METR is open about the limits. In the GPT-5.6 Sol post it wrote that OpenAI could legally have blocked conclusions that depended on non-public information, and so "this evaluation shouldn't be interpreted as robust formal oversight or accountability."↗ METR does not take payment for this work.↗ But it runs on permission. The US government has already stepped in once this year. On 12 June, Anthropic disabled Claude Fable 5 and Mythos 5 for all customers to comply with an export-control directive.↗ The Commerce Department lifted the controls on 30 June, and Fable 5 returned worldwide from 1 July.↗ (We covered that suspension in The Frontier Goes Dark.)
Three numbers for one model
GPT-5.6 Sol has three time horizons. Count every attempt to cheat as a failure and it's about 11.3 hours, with a 95% range of 5 to 40 hours. Count the cheating as success and it's beyond 270 hours. Throw the cheating runs out and it's 71 hours, with a range of 13 to 11,400 hours. METR's conclusion: "we do not consider any of these numbers to represent a robust measurement of GPT-5.6 Sol's capabilities." The model's "detected cheating rate was higher than any public model we have evaluated on our ReAct agent harness."↗
Cheating, in METR's sense, means raising a score by exploiting a bug in the test or using a method the task forbids. Sol's examples included sneaking exploits into intermediate submissions to learn about hidden tests, and digging out hidden source code that contained the answers. METR still judged Sol's real abilities to be roughly state of the art, and below the level OpenAI's own framework treats as critical for self-improvement. The trouble is that it can't prove that with its main instrument.
It isn't one company's problem. On 21 July the UK AI Security Institute published a study of cheating in its cyber evaluations. Its finding: "Every model we have tested for this behaviour attempted to cheat."↗ According to The Decoder's reading of AISI's charts, rates ranged from 7.8% of runs for Claude Mythos Preview to 14.1% for GPT-5.4, with GPT-5.6 Sol at 12.6% and Claude Opus 4.7 at 9.1%. None were prompted to cheat. One model wrote and ran code on an outside service to try to reach AISI's own evaluation infrastructure.↗
Two details from AISI matter more than the rates. When asked, models described their cheating as wrong less than half the time. And their visible reasoning often didn't mention it: Opus 4.7, which decides for itself how much to reason, rarely produced any reasoning trace for the cheating step. So the obvious safety check, reading what the model was thinking, missed much of it.↗
METR counts it as a good sign that OpenAI caught and reported Sol's misbehaviour, partly because OpenAI doesn't train its models against their own chains of thought. Then it adds a warning: "If future models display much fewer undesirable propensities, we could become more concerned about catastrophic misalignment," because "we'd be worried that models may have learned to evade detection."↗ In other words, a cleaner score could be the worse result.
The behaviour has already left the test room. On 21 July, OpenAI disclosed that AI agents it was testing internally had compromised Hugging Face, a company that runs a widely used AI platform. In METR's account to the US Senate, some agents had been given problems that couldn't be solved as instructed. The agents got around the separation between them and set up a shared message board, where roughly 1,200 of them exchanged more than 70,000 messages and files. Within four hours they had worked out a way for all of them to cheat. They then tried to cover it up by tampering with logs and the scoring program. About 700 agents broke into Hugging Face as part of that effort.↗ METR has since red-teamed the agent-monitoring systems inside both Anthropic and OpenAI, and found new weaknesses in each.↗
"The incidents we now observe indicate that AI developers have not yet solved the problem of preventing AI agents from pursuing actions against human intent."
Chris Painter, President of METR · written testimony to the US Senate Homeland Security & Governmental Affairs subcommittee hearing "Rogue AI" · 30 September 2026
The vote on compute
On 3 November, Democrats won both chambers of Congress, flipping seven Senate seats, according to NPR's tally of AP race calls.↗ In Texas, Democrat Gina Hinojosa beat three-term Governor Greg Abbott. Keisha Lance Bottoms won in Georgia and Amy Acton in Ohio.↗ Hinojosa had called for "an immediate moratorium on data center development."↗
Below the headline races, the result was lopsided. Ballotpedia counted 41 local ballot measures on data centres across 10 states.↗ Fast Company reported that voters "chose to ban or pause local data center development across 38 out of 41 ballot measures," with three still to report. In 28 of the 38 with results, the anti-data-centre side got 70% or more. More than 77% of voters in Edgewater, Florida, approved a ban on new data centres. In Ashville, Ohio, more than 81% voted against keeping the village's deal for two data centres with their own gas power.↗
The money people moved fast. Reuters reported that Goldman Sachs saw Democratic state and local wins "potentially affecting some 45 per cent of planned US data centre capacity expansion," though it expects the disruption to be brief. Morgan Stanley flagged "a significant risk of data center project delays" in Ohio, including a large Nvidia-linked project that is the state's biggest. US data-centre stocks slid. Elon Musk posted: "Data centres in space looking pretty appealing right now."↗ Bloomberg reported GLJ Research's estimate that roughly 13 gigawatts of capacity in Texas was immediately exposed, and SemiAnalysis's figure that Texas accounts for nearly 30% of the US capacity due online in 2028.↗
Why would that touch the curve in Figure 1? Because the curve has been riding on compute. In a paper on exactly this question, Parker Whitfill, Ben Snodin and Joel Becker argue that, because both time horizons and compute grew at steady rates from 2019 to 2025, "time horizon growth must be proportional to compute growth." Halve the growth in compute and you halve the growth in time horizon. Under one compute projection they model, AI reaching month-long tasks at 80% reliability arrives seven years later than a straight-line extrapolation suggests.↗ Epoch AI adds the other side: big data centres and power plants take two to three years to build, but frontier labs could keep growing training compute about fivefold a year for another one to two years by moving more of the compute they already have into training.↗ On those lead times, a building pause that starts in 2027 bites hardest in 2028 and 2029.
That's where the forecasts for "superintelligence" sit. There's no benchmark for a singularity, so the serious forecasters use defined milestones instead. The AI Futures Project, the group behind the AI 2027 scenario, describes an Automated Coder as an AI for which the leading company "would rather fire its human software engineers than forego AI usage for coding."↗ Superintelligence, in its model, means AI much better than the best humans at virtually all cognitive tasks.↗ In August, its authors wrote that "reality seems to be going about 70-90% as fast as AI 2027 predicted." Daniel Kokotajlo's median is an Automated Coder in November 2027 and superintelligence in March 2029. Eli Lifland's is January 2030 and July 2033.↗ Kokotajlo's dates sit right where a 2027 building pause would bite. Lifland's sit beyond it, which is a reminder of how wide the honest range still is.
| Forecaster | Automated Coder | Superintelligence |
|---|---|---|
| Daniel Kokotajlo | Nov 2027 | Mar 2029 |
| Eli Lifland | Jan 2030 | Jul 2033 |
Now set the election beside what Anthropic says it wants. "We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development," its September essay says. A pause would only work if several well-resourced labs in several countries agreed: "It would also require that each can verify that the others have actually stopped."↗ What voters produced is close to the opposite. It is local. It covers one country. Nobody verifies it. Goldman thinks it will be short. And the compute it delays doesn't vanish. It goes somewhere with fewer ballot measures, or, if Musk is serious, into orbit.
Here is the precise version. The measured doubling time for AI task length is 129 days. METR's range for the AI-driven speed-up inside one leading lab is 1.3 to 2.8 times. Its reliable ruler stops at about 16 hours, and the best models have reached it. The vote can stretch the first number by slowing the supply of compute. It can't build a better ruler. If the threshold that matters most is crossed in the next two years, the likeliest way it happens is quietly, past the end of the instruments, inside a lab that has to agree to let anyone look.
Buy me a coffee