AI Atlas Charting the road to superintelligence

Loading data

Charting the road to superintelligence

How long a task AI can complete, and how fast that is changing

Each dot is a model, placed on the day it was released. Its height is the length of software task it completes half the time, counted in the time the task takes a skilled person. The line joins the models that set a record.
Longest task estimate 50% success
Doubling time since 2023
Revenue run rate Anthropic + OpenAI
AI chips shipped H100 equivalents

METR's task-length measure

Task length each model completes, by release date

Show these numbers

How to read this measure

When the projection reaches each task lengthProjection

How the models compare on Epoch's index

Epoch Capabilities Index, every scored model

Show these numbers

Many tests are close to solved

Show these numbers

Chips, data centres and power

AI chips shipped since 2022, by designer

Show these numbers

Computing power each lab can use, year end (latest estimate)

Power at the largest AI data centres, built and planned

Where the largest AI data centres are

Largest AI data centres by planned power

Computation used to train notable models

Revenue, investment and valuations

Annualised revenue of AI labs

Show these numbers

Private investment in AI companies, by year

Latest valuation at a closed funding round

The same capability keeps getting cheaper

Cheapest model at or above a fixed score

Show these numbers

Price against capability today: the Pareto frontier

Show these numbers

What a finished task costsScenario

You set every input here. The atlas has no measured success rate for your tasks, and scores from different tests cannot be combined into one. Set the rate you have seen, or try several.

Cost against a person alone, across success rates and review time

Show these numbers

How many people use AI and what they use it for

People using each assistant

New research papers that acknowledge using a language model

Share of each job's tasks people already use Claude for (use, not job losses)


    Open models on Hugging Face

    Where the trends pointProjection

    If each trend simply continuedProjection

    Build a scenario for task lengthScenario

    When each test reached, or would reach, nine-tenths solvedOpen dots are projections

    Under the hood

    Every projection on this page is a line fitted to the measured points and carried forward. Nothing else is assumed. These are the fits and the choices behind them.

    What people disagree about

    Autonomy

    Does a longer task length mean AI can work alone?

    Some read the doubling as a countdown to AI that does whole jobs. METR's own answer is no: the measure is how hard a clean, well-specified software task is, passed half the time. It says models do worse on messier work.

    Jobs

    Is it already costing work?

    Some point to how much of certain jobs now shows up in AI use. Anthropic's own study of that data found no systematic rise in unemployment among exposed workers so far, with early signs of slower hiring of younger workers. The data so far fits both readings.

    Risk and limits

    Should it continue, and can it?

    Researchers disagree on how dangerous more capable systems are and how fast safeguards must follow. Others doubt the trend lasts: power, chips and money are finite, and reliability on real work lags the test scores.

    Why a trend can bend

    Slower

    The inputs run short

    Chips, power and money have grown several-fold a year. If any of them slows, the capability trends that depend on them should slow too, with a lag.

    Faster

    AI speeds up AI research

    Task length is measured on software and machine-learning work. If models take over more of that work, the doubling time could keep shortening, as it did after 2023.

    Unknown

    The ruler runs out

    Terms used on this page

    METR
    A non-profit that tests AI models independently of the labs that build them.
    Epoch AI
    A non-profit research group that collects public data on AI: test scores, chips, data centres, company figures.
    Task length
    How long a task takes a skilled person, used as a measure of how hard it is. Not the time a model runs.
    Benchmark
    A fixed set of questions or tasks with a score, so different models can be compared.
    Record, frontier
    The best result so far. A model "sets a record" if it beats everything released before it.
    Open weights
    A model whose trained files are published, so anyone can download and run it. Closed models are reachable only through the maker's service.
    H100 equivalent
    A common unit for chip power: one Nvidia H100, a top AI chip from 2022. A newer chip counts as several.
    Token
    The unit AI services charge by, roughly three-quarters of a word.
    Pareto frontier
    The set of options nothing else beats on both measures at once, here price and capability.
    Run rate
    The latest month or quarter of revenue multiplied up to a full year.
    Post-money valuation
    What investors agreed a company was worth right after a funding round, including the new money.
    Preprint
    A research paper posted publicly before a journal has reviewed it. arXiv and bioRxiv are the main sites.

    Earlier projections, kept as they were

    The task-length projection in the Outlook chapter moves whenever METR publishes. A copy is saved with the date, the data it was built on and the method, and is not changed afterwards. Each milestone is then marked as observed, not yet observed, or not measurable with current data.

    • Loading

    What changed

    New observations and revisions in the figures this page tracks, found by comparing each daily refresh with the one before.

    • Loading

    Sources and freshness

      Data from Epoch AI is used under the CC BY 4.0 licence and adapted here: one best result is kept per model family, some organisation names are shortened, and site power totals are summed from Epoch's timelines. The Anthropic Economic Index is used under the MIT License. METR publishes no licence for its results file; it is used with credit. Measured figures are shown as published. Anything fitted or carried forward is marked "Projection" and listed under the hood. Benchmark scores are for the best configuration of each model that was tested. Company revenue and usage figures are disclosures or press reports collected by Epoch AI and carry the confidence Epoch gives them.