Skip to main content
Artificial Intelligence

Act I

The Acceleration

Capability is compounding. Compute grows five-fold each year, autonomous work horizons double every four months, and benchmarks designed to hold for a decade are falling in months.

Capabilities Index

The Race No One Agreed To Run

Epoch AI's Capabilities Index folds more than fifty benchmarks into one score per model, so it keeps telling models apart long after any single test is maxed out. The frontier has climbed from 122 with GPT-4 in March 2023 to 174 today, and the record has changed hands between labs 12 times along the way.

Each line tracks a lab's most capable model over time. Frontier shows the single best model at each date above every other model Epoch has scored, the same curve as the Capabilities Frontier on the AI timeline.

Data: Epoch AI, Epoch Capabilities Index (CC BY). Last updated:

The index fits one capability score per model from its results on many benchmarks, estimating each benchmark's difficulty from the models that took several of them. The scale is linear but arbitrary: Claude 3.5 Sonnet is fixed at 130 and GPT-5 at 150, so only the distance between two models means anything, and the vertical axis does not start at zero. Hover a model to see Epoch's 95% range. Epoch refits the whole index as models and benchmarks are added, so earlier scores can shift a few points between refreshes. Method: Epoch AI, A Rosetta Stone for AI Benchmarks (2025). Until November 2026 this chart used the Artificial Analysis Intelligence Index; its version 5 re-scored only the newest models, so it could no longer show the race from 2023.

Raw scores are one thing. But the real question is: what can AI do on its own? Not answer a question. Not write a paragraph. Work: complex tasks that would take a skilled person hours, finished alone, half the time. That horizon is doubling every four months.

Autonomous task length

Autonomy Horizon

How long a task takes a skilled human before AI agents complete it with 50% success. In 2019, AI could manage a 2-second task. By April 2026, the frontier had reached tasks of roughly 17 hours of skilled work, though METR notes that measurements above 16 hours are unreliable with the current task suite. The horizon is doubling every 129 days.

This measures the length of real software engineering task, in skilled human time, that AI agents finish on their own half the time. It is not how long an agent runs. The y-axis is logarithmic: each gridline represents a tenfold increase in task length.

Data: METR — Horizon Benchmark v1.1. Cite: arXiv:2503.14499 (NeurIPS 2025). Used with attribution for educational purposes. · Data as of 7 Apr 2026

The p50 horizon length is the task duration (for a skilled human) at which AI agents succeed 50% of the time. Shaded bands show 95% confidence intervals. The frontier line connects state-of-the-art models at time of release. The dashed trend line shows the 129-day doubling rate from 2023 onward.

One lab, seen from inside. October 2026.

When the work feeds itself

The chart above asks how large a task an agent can finish alone. A second question sits behind it: is AI already speeding up the making of AI? From August to October 2026, METR, an independent evaluator, went inside Anthropic to look, with internal data, benchmark scores for unreleased models and staff interviews.

  • ~2× pace of progress

    Anthropic is likely improving its models on well-specified tasks at about twice its rate of March 2024 to October 2025. METR would not be surprised by 1.3× or 3×.

  • ≈80% of that speedup due to AI

    METR's own guess for October 2026, with substantial uncertainty. Other causes likely played a large role in late 2025. Faster growth in training compute was ruled out.

  • Assisted not yet autonomous

    The speedup comes mainly from AI helping researchers. AI cannot yet make highly significant, fully autonomous contributions to the most important parts of model training, though researchers describe "signs of life".

How far the estimates spread. 2× means two years of earlier progress made in one.
METR's overall judgment ~2× (1.3 to 3)
Benchmark trend, Anthropic models 1.9× (1.5 to 2.8)
Anthropic's own system card 1.7×
Benchmark trend, all frontier models 1.4×
Across alternative analyses 1.1× to 2.9×

1× means no speedup. Adding other labs' models lowers the trend to 1.4×, and METR reads that two ways: closer to the truth if Anthropic trailed the frontier in 2025, too low if earlier models from other labs were more tuned to these benchmarks. Bars on the first two rows show METR's range and a 95% confidence interval.

Why it matters

Progress that feeds on itself can compound. METR saw no clear bottleneck over the next 12 months in algorithms, compute or data, and calls the risk of even faster acceleration significant. Its view of those bottlenecks changed several times along the way, and it may be missing some.

What cannot be checked

This is one lab and a subjective estimate. Anthropic redacted most of the evidence METR found most informative, saying it could accelerate other developers. METR's own verdict: the public has to rely on its judgment much more than would be ideal.

Source: METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, 5 November 2026.

The benchmark trend uses Anthropic's version of the Epoch Capabilities Index, which scores models across many benchmarks. Anthropic's system card figure is from Haiku 5.5, and METR's is higher mainly because it includes newer unreleased models. Anthropic also dropped a planned check of how it divides compute between capabilities and safety, before METR reached conclusions on it.

The horizon is not just rising — it is multiplying. AI agents — systems that take autonomous actions in the world, calling tools, navigating web pages, completing multi-step tasks — are advancing on two axes at once. Demand for them in the labor market has exploded; their measured capability on real-world benchmarks is not far behind. Both are happening simultaneously, and the frameworks meant to govern them are still being written.

AI-agent demand & capability

The Horizon Extends

AI agents are advancing on two axes at once. The labor market reveals demand: a skill cluster that did not meaningfully exist before 2024 — postings for "agentic AI," "AI agents," "LangGraph" — exploded across 2025 and now appears in roughly 48,000 US job postings, overtaking decade-old categories like neural networks, robotics, and visual recognition. And the agents themselves are gaining capability: on four real-world benchmarks where the best models scored 1–20% as recently as 2024, they now hit 66–93%, closing in on human performance across domains as distinct as web navigation, terminal commands, computer use, and cybersecurity.

Demand view: each line tracks the share of US job postings mentioning an AI-related skill cluster, 2010–2025. The cyan AI agent line emerged from zero in 2024 — note how quickly it overtook visual recognition, robotics, and neural networks. Capability view: each line tracks the best AI model on a real-world agent benchmark; dashed lines mark human performance. Terminal-Bench is monthly; the others are annual.

Data: Stanford AI Index 2026 (CC BY-ND 4.0). Demand: Economy chapter, Fig 4.4.3 (Lightcast 2010–2025) and Fig 4.4.7 (agent-skill breakdown, 2024 vs 2025). Capability: Technical Performance chapter, Fig 2.5.2 (Terminal-Bench 2.0), Fig 2.6.2 (OSWorld), Fig 2.6.3 (WebArena), Fig 2.6.5 (Cybench).

Demand intermediate-year values were read off Stanford's published chart to ±0.05%; 2025 endpoints are exact. Capability values are all exact (printed on Stanford's charts).

Scores on a benchmark are one thing. But benchmarks are supposed to be hard — designed to test the limits of what machines can do. What happens when the limits keep moving? When problems designed to take a decade are solved in six months?

Frontier benchmarks

Closing the Gap

Seven benchmarks designed to measure the frontier of AI capability. Each line tracks the best model score over time. FrontierMath went from under 2% to 52%. ARC-AGI-1 from 18% to 98.5%. SWE-bench Verified from 31% to 83%. Humanity's Last Exam — scored just 3% at launch — has already been pushed to 27% in months. Problems designed to resist progress for years are falling in months.

Each line tracks AI performance on a different standardized test. When a line approaches the top of the chart, AI has matched or surpassed human expert performance on that task.

Data: Epoch AI Benchmark Hub (CC BY 4.0). Individual sources: SWE-bench, Epoch FrontierMath, ARC Prize, GPQA, SimpleQA, MATH, Humanity's Last Exam. · Data as of 10 Nov 2026

For one score that combines these tests, see the Frontier view of The Race No One Agreed To Run.

Behind every model is a number: the raw computation used to train it — a proxy for ambition. It has been growing at five times per year since 2020. There is no sign it is slowing.

Training compute

The Engine Room

Training compute for frontier AI models has grown by over 10,000× since 2020 — doubling every 9 months. Each point on this chart represents a decision by an organization to invest more computational resource than any previous model in history.

Each dot represents an AI model. The vertical axis shows the computing power used to train it — note the logarithmic scale, where each step represents a 10× increase. Training compute has grown roughly 4× per year since 2010.

Data compiled from Epoch AI research, published papers, and industry reports. Training compute estimates carry uncertainty — see Epoch AI documentation for methodology. CC BY 4.0.

Hardware Scaling: transistor counts from Our World in Data (Moore's Law), GPU FLOPS from manufacturer specifications (NVIDIA, AMD). Moore's original observation: Gordon Moore, "Cramming More Components onto Integrated Circuits," Electronics, 1965.

Training compute is only one axis of acceleration. AI development has entered a self-reinforcing feedback loop — a cycle in which AI systems are themselves accelerating the development of more powerful AI systems. Each stage operates on a shorter timescale than the last.

AI building AI

The Feedback Loop

AI systems now help build the next AI systems. The end point people worry about has a name, recursive self-improvement: a model that builds its successor with no one in the loop. That has not happened. But in 2026 a frontier lab began publishing measurements of how close it is, and the numbers moved fast.

The cycle, in theory

1Infrastructure Years AI success drives investment, which buys better hardware, which enables better AI Ongoing, massive

AI success drives investment, which buys better hardware, which enables better AI. This is the broadest loop — but it is already the most powerful driver of progress. Hundreds of billions of dollars are flowing into AI infrastructure on the basis of near-term commercial returns.

2Model Development 1–2 Years Human-led research with AI assistance produces better models Active across labs

Human researchers use AI assistants to design better architectures, run experiments faster, and explore larger search spaces. All major AI labs now use AI extensively in their own research pipelines.

3Data Generation Months AI systems generating synthetic training data Beginning phase

AI systems generate synthetic training data for the next generation of models. This partially addresses the ‘data wall’ — the exhaustion of available high-quality human-generated training data.

4Tool Development Days–Weeks AI systems creating their own scaffolding and tools Early experiments

AI systems creating their own scaffolding, tools, and software infrastructure — writing the code that enhances their own capabilities. Limited by software integration time, but this stage operates on a timescale of days to weeks.

5Network Self-Improvement Hours AI collectives innovating their own organisational structures Unknown

Groups of AI systems innovating their own social and organisational structures to work more effectively as collectives. Limited by inference rate. This stage has no confirmed examples but is a logical extension of current agentic AI development.

6Recursive Improvement Unknown AGI autonomously self-improving Not yet possible

The final stage: AGI or superintelligent systems autonomously improving themselves — designing better neural networks, better training methods, better hardware. This is the ‘intelligence explosion’ scenario. Once started, the rate-limiting factor becomes the laws of physics, not human engineering capacity.

Each stage feeds the next, and each runs on a faster clock: from years for new hardware to, potentially, hours for AI collectives improving themselves. Colour follows the order of the stages. Open one to read more. Framework: Anthony Aguirre, Keep the Future Human (Future of Life Institute, 2025), Table 2.

Measured inside one lab

Who does the work of building AI

Anthropic broke the work of making its models into 378 kinds of task, from designing experiments to fixing a broken training run, and rated each month how much of it Claude did. The scale comes from Epoch AI and runs from 0, no AI, to 5, fully automated.

In August 2025 most work sat at "AI assists". A year later, more than 90% of it was done with AI collaborating or leading, and Claude led 26%: it took a high-level request, did most of the task, and a person approved the result. No part of the work was fully automated. Anthropic checked its Claude-based ratings against staff: they agreed exactly 59% of the time and were within one level 97% of the time, where two staff members rating the same task agreed exactly only 35% of the time.

The work, task by task

What AI already leads, and what people still decide

Epoch AI mapped the work of researchers and engineers at frontier labs into 62 tasks in six kinds of work, and rated how much of each AI typically handles today. Each square is one task. Hover or tap a square to read it.

Writing code, generating training data, launching runs and reviewing code are already led by AI. Choosing what to work on, deciding what ships and allocating compute stay at level 1: AI helps people think, people decide. No task reaches level 5. Epoch calls these ratings best-guess starting points for discussion, describing how the work is usually done rather than the most AI could do.

Keeping watch

Thirty thousand agents, and who checks them

  • ~30,000 AI agents at work at once

    Doing research and engineering on Anthropic's most-used internal platform in August 2026. Over that month they made more than a billion decisions.

  • 1 in 47,000 actions blocked before they run

    Every action passes a live monitor first, built to stop fast, irreversible harm such as an agent copying model weights. People review each blocked action within a week.

  • ~50 a week reach a human reviewer

    A second monitor reads everything afterwards and flags about 100,000 transcripts a week, one or two in every thousand, looking for slow harms such as signs of misalignment. Classifiers sort them, and the most serious go to people.

  • 6% of R&D compute went to safety

    In one week of July 2026, and 12% of the compute used by AI doing the research itself. Work that helps capabilities as much as safety counted as capabilities, so these shares are on the low side. One week shows the method works, not a trend.

Does it speed things up?

Two slopes

METR, an independent evaluator, fitted the benchmark scores of Anthropic's models, released and internal, over time. Before October 2025 they rose about 15 points a year on Anthropic's version of the Epoch Capabilities Index. Since then, about 29: progress at 1.9 times its earlier pace, with a 95% interval of 1.5 to 2.8 times. METR's wider assessment is above.

If the slope holds

What the forecasters made of it

The AI Futures Project wrote the AI 2027 scenario. On 10 November 2026 its six authors fed METR's finding into their own model. If progress runs at about twice its old pace and about 80% of the gain comes from AI, AI alone is speeding things up about 1.7 times. That is roughly five months ahead of their own AI 2027 story. Three months earlier, even the most bullish of them gave a pace this fast less than a 4% chance.

All six now expect a superhuman AI researcher at some point in 2027: an AI that matches the research taste of the best people at an AI company and can do all of its coding. The next milestone is top-expert-dominating AI, AI at least as good as the best human experts at almost every cognitive task. They give it a 35 to 65% chance before 2028. These are forecasts for a world where no one deliberately slows down.

Pick a pace for November 2026. How likely is each milestone before 2028?

METR: about 2×, anywhere from 1.3× to 3×. About 1.7× from AI alone.

The pace compares AI progress with its 2024 to mid-2025 rate, so 2× means two years of earlier progress in one. Results come from the AI Futures Model runs that match each pace, with two authors' settings: Daniel Kokotajlo's and Eli Lifland's. At about 1.25× and at 3× or more the authors give only rounded figures, for Kokotajlo's settings alone.

Their forecasts for AI as good as top experts at almost everything, from AI 2027 to now

The expected year is the median: the year by which each author thinks the milestone is as likely to have happened as not. Until August 2026 they gave it to the year only, shown as a bar spanning that year. Grey dots are the other four authors, who first gave forecasts in November.

Years of progress in one calendar year. 1 means a year at the 2024–25 pace.
2025 1.0
2026 1.4
2027, Kokotajlo's settings 4.9 (2.7–25)
2027, Lifland's settings 4.2 (2.6–12)

The white line marks 2 years: what 2027 holds if today's ~2× pace simply stays where it is.

Two readings fit the same data. If the speedup was a one-time jump and today's pace simply holds, 2027 brings about two years of progress. If it is a loop, where each gain speeds the next, the authors' model gives four to five, with a long tail. METR's report cannot yet tell the two apart, as the Scheme writers point out. Ranges are 50% intervals from the model runs that match METR's pace. Log scale.

What they conclude

If a superhuman AI researcher is coming in 2027, the authors say it is incredibly urgent to pause or substantially slow frontier development. Ideally that means an industry-wide pause followed by an international agreement. In their view, no company is on track to keep such a system aligned with human intentions. They also think China would probably be able to steal the training recipe and could then follow within a few months.

What could make this wrong

Everything rests on one evaluator's judgment of one lab, and Anthropic redacted most of the evidence METR found most informative. The speedup was measured on well-specified tasks, and messier work may lag behind. The 80% share due to AI is METR's rough guess. The authors call their analysis highly preliminary and expect some errors in the modelling. The Scheme writers note that a one-time jump fits the data as well as a self-feeding loop. If the jump came from one bigger training run and the real speedup is 1.2 to 1.5×, they put superintelligence around 2030 rather than 2027.

For how AI 2027's earlier predictions have held up, see AI 2027, Checked So Far.

Sources

All of these measurements come from one lab, which built the measuring tools and partly relies on its own models to apply them. Anthropic says so, and asks for regular publication, outside verification and shared definitions across labs.

Read the full story About & Sources