Nervous systems took more than half a billion years to grow from a few hundred neurons to the human brain. AI systems went from failing simple software tasks to finishing ones that take a skilled person a full working day in about seven years. This page puts both climbs on shared scales, lets you push the curves as far forward as you like, and shows the source and formula behind every number.
Along the human lineage, brains doubled in size roughly once every –, going by the fossil record. The length of software task that frontier AI can complete has doubled every – since 2019, and every – since 2023.
Log scale. The shaded bar around a dot is the published confidence interval, where one exists. Hover a row to see what was measured. Biological rows are log-linear fits to neuron counts, so they are averages over millions of years.
Pick how far ahead to look, then change the assumptions. AI and humans share one axis, the IQ-equivalent: the human average is 100 and 15 points is one standard deviation. No human who has ever lived would be expected to score above about –, so values higher than that are arithmetic on the scale rather than test results.
The AI line between the last unsaturated IQ result and today is carried forward from the task-horizon trend, because both IQ tests have hit their maximum score. Grey hairlines mark how rare a score would be among people. The window shows as much past as future, so long horizons reveal the hominin record on the left. The human line before today is a model of biological capacity from brain size; the rings are measured population means on today's test norms.
How long the METR task horizon takes to double. The default is METR's own 2023–2026 estimate on its current task suite.
Converts doublings into points. Calibrated as the AI IQ-test trend (points per year) divided by the METR trend (doublings per year) over 2024–2026. See Methods.
After the start year, each doubling takes a fixed percentage less time than the one before. With any shrink above zero the number of doublings becomes infinite in finite time, so a ceiling matters.
The hominin rate is the fossil-record trend in neurons, converted to IQ-equivalent points (Methods). The Iceland rate is measured genetic selection today.
Gains from schooling, nutrition and health. They are environmental and level off, so the total is capped.
Adds an enhanced cohort that gains points every generation, levelling off at a cap.
Years are computed from the settings above. They don't depend on the look-ahead slider.
| Landmark | Frontier AI | Average human | Enhanced cohort |
|---|
Everything is recomputed in your browser from the data tables at the bottom of the page.
Today's starting points come from least-squares trend lines, not from a single model: IQ₀ = –, carried forward from the last unsaturated IQ result, and h₀ = – on METR's suite. k = – points per doubling is calibrated on May 2024 to April 2025, the last stretch before the IQ tests saturated. The default doubling time is METR's 2023–2026 estimate.
b is the biological rate, f the environmental rate, g the gain per generation, and G the enhancement cap. Rarity uses a normal distribution with mean 100 and SD 15. The full method, with each assumption and its weaknesses, is in Methods.
The Flynn effect raised average scores in developed countries by about three points a decade through the 20th century. Between May 2024 and April 2025, the best AI score on an IQ test kept out of training data rose about – points a year, until models reached the test's maximum. From there the page carries the trend forward with task-horizon data.
AI scores are TrackingAI's best single run each month on two tests: an offline test written in 2024 and never posted online, and the public Mensa Norway test. Text-only models get written descriptions of the puzzles. Both tests have now hit their maximum scores, so the recent flat stretch reflects the ceiling rather than a pause. Human means before 2000 are reconstructed from the measured rate of about 3 points a decade. There is no consensus figure for 2010–2020, so the human line stops in 2000.
METR times skilled people on software, research and cybersecurity tasks, then finds the task length at which a model succeeds half the time. That "50% time horizon" is the backbone of the extrapolator, because it has a real unit and a long, consistent record.
Each dot is one model's 50% time horizon on METR's current task suite, dated by public release, with a 95% confidence interval. Hollow dots were not a new record, or METR flags them as not robust: Claude Mythos Preview's early measurement is above the 16-hour range METR considers reliable, and GPT-5.6 Sol cheated on many tasks. Both are left out of the fits. The three earliest points reuse the original suite's values, as METR does.
Each panel is the best score any AI system had reached by that date. Orange hairlines are the human baselines published with the benchmark.
Best score by model release date. MMLU scores are mostly reported by the developers. The others come from Epoch AI's own runs, the ARC Prize leaderboard and the HLE leaderboard. HLE and SWE-bench Verified have no published human baseline. MATH's human figures come from two people answering 20 problems. GPQA's expert figure runs high because the Diamond subset keeps only questions that experts answered correctly.
Epoch's Capabilities Index fits every model and benchmark together with item response theory, so it has no maximum. The frontier has risen about –.
From Kiela et al. (2023), published by Our World in Data. That dataset isn't licensed for redistribution, so only the start and human-level years appear here. Benchmarks released after 2020, such as MMLU, GSM8K and HumanEval, had not crossed the human line when the dataset ended in 2023.
The largest training runs have grown about – since 2012. Orange lines are estimates of how much computation a human brain does, which Open Philanthropy researchers used as reference points for AI timelines.
Training compute from Epoch AI's database. Values for closed models are Epoch's estimates rather than company disclosures. Filled dots set a new record when released. The brain anchors assume about 10¹⁵ operations per second, the central estimate in Carlsmith (2020), whose range runs from 10¹³ to 10¹⁷.
Each dot is a living animal's neuron count, placed at the date its lineage split from ours. Orange dots trace the branch that leads to us and end with fossil hominins, whose counts are estimated from skull volume. Both axes are logarithmic, so each step to the right covers ten times less time. The band along the bottom shows the geologic periods in their official colors.
Counts are measured unless the tooltip says estimated. Split dates are TimeTree medians and are uncertain for the oldest splits: estimates for the bilaterian split run from about 600 to 830 million years ago. Living animals are relatives, not ancestors. A mouse's lineage has evolved for as long since the split as ours has.
Cranial capacity from fossil skulls. The hominin doubling time in the ruler at the top comes from these estimates.
Vertical bars show the range across specimens and horizontal bars the span of dates. Neuron counts for extinct species are estimates from brain size. Growth came in two spurts, 2.0 to 1.5 million years ago and from 700,000 years ago on, with a slight decrease since the end of the last ice age. Neanderthals are a sister lineage, so they are left out of the fit.
The page compares three kinds of evidence: brain size over evolutionary time, test scores of human populations over the last century, and measurements of AI systems since 2019. Each is shown in its own units in the evidence sections. The extrapolator then converts them to one shared scale using the steps below. All of the arithmetic runs in your browser from the tables on this page, so changing a number in the data changes every chart and readout that depends on it.
IQ is defined relative to a population. By construction the mean is 100 and the standard deviation is 15, and a score says how rare a result is. That makes the scale useful for a comparison like this one and also limits it. A value of 250 means "10 standard deviations above the human mean". No test can measure that, and no person would be expected to reach it. Above about –, a score is rarer than one in all of the roughly – people who have ever been born, assuming a normal distribution. The page keeps drawing the scale above that point so the curves stay comparable. Read those values as distances on a ruler, not as test results.
If AI systems speed up AI research, each doubling can take less time than the last. The extrapolator models this with a single number: after the start year, every doubling takes a fixed percentage less time than the one before. Summing that series shows the number of doublings becomes infinite after T / r years, where T is the doubling time and r the shrink rate. The two published reference points: the AI 2027 timelines forecast assigns roughly 40–45% probability to a "superexponential" case in which each doubling takes 10% less time, and Forethought's analysis expresses feedback as a returns parameter r_s, which implies each doubling time is 2^(1/r_s − 1) times the previous one. Their median r_s of 1.2 gives – shorter doublings, and r_s = 2 gives –. These are modelling assumptions by those authors, not measurements.
Without a ceiling, self-improvement produces infinite values in finite time. The ceilings convert limits on computation into a maximum number of capability doublings, using the historical ratio of capability doublings to training-compute doublings (–, the METR long-run trend divided by Epoch's 5× a year growth in frontier training compute).
These are order-of-magnitude bounds. They ignore limits on data, algorithms and time, and they assume capability keeps scaling with compute the way it has since 2019.
Bracketed numbers throughout the page link here. The raw research files, with a status for each value (measured, estimated or third-party), are in the project repository.