Primer · Interactive
Where AI Stands in October 2026, and Where It's Heading
Four charts that explain what today's AI can do, how fast it is improving, and what to expect over the next two years. No background needed.
TL;DR
- The best AI systems now outscore PhD experts on hard science exams and can work on their own for many hours at a time.
- The length of task they can finish on their own has doubled about every four months since 2023.
- Free-to-download models trail the best paid ones by only about four months, at a fraction of the price.
- The raw computing power once thought to match a human brain is now affordable to a business. The software was the missing piece, and it is catching up fast.
Five words you need
- Model
- The AI itself: a very large program trained on text, code and images. ChatGPT, Claude and Gemini are apps built on models.
- Benchmark
- A standard test used to compare models, like an exam with a fixed set of questions or tasks.
- Agent
- A model that works through a task step by step on its own, using tools such as a web browser or a code editor.
- Open model
- A model whose files anyone can download and run on their own computer. Closed models are only available through the company's app or paid service.
- Compute
- Raw computing power, counted as calculations per second. More compute lets you build and run bigger, smarter models.
01AI can work alone for longer and longer
The clearest way to measure progress is to ask how long a task an AI agent can finish by itself. A research group called METR times how long each task takes a skilled person, then checks which tasks the AI completes successfully at least half the time.
In 2019 that was a few seconds of work. By April 2026 it was around 17 hours, for Claude Mythos Preview. That growth is remarkably steady on a log scale, where each gridline is a big jump in length.
Show the numbers
The newest models have not been measured yet. METR's test only goes up to about 16 hours of work and the best model measured is already past that, so everything after April 2026 is a projection.
02Who's ahead right now
Seven new top-tier models arrived between mid-September and mid-October. The independent group Artificial Analysis runs every model through the same ten hard tests and combines them into one score. Switch the chart to cost to see the other half of the story: models close to the top are getting much cheaper.
Show the numbers
Anthropic's Claude holds the top two spots. OpenAI's GPT-6.1 Sol is the bargain: nearly as capable for about an eighth of the cost. The best open model, Xiaomi's MiMo, scores 46 and costs 13 cents a task.
03Free models are only months behind
This chart uses one hard science exam with questions written by PhD researchers. Experts score about 65% in their own field. Both paid and free models passed that line within months of each other in late 2024 and early 2025, and both are now close to the maximum.
Show the numbers
Across many tests, researchers at Epoch AI find open models trail the best closed ones by about four months on average. Most of the leading open models now come from Chinese labs.
04Computing power keeps getting cheaper
Computing power per dollar doubles about every two and a half years. Scientists' estimates of the computing power a human brain uses vary widely, from about ten trillion to a hundred million trillion calculations per second. The most careful recent review lands on about a thousand trillion. Pick a budget and an estimate to see when that budget could buy that much.
Where do the brain numbers come from?
Nobody has measured the brain's computing power directly. Each estimate picks a way to translate biology into computer operations, and the choice of method moves the answer by a factor of a million or more. These are the main approaches, with links so you can read the reasoning yourself.
- Open Philanthropy best guess · 10¹³–10¹⁷, best guess 10¹⁵
- Joe Carlsmith (2020) consulted more than 30 experts and compared four methods. Modelling how neurons signal and learn gives 10¹³–10¹⁷ FLOP/s. He judges it more likely than not that 10¹⁵ is enough, and under 10% that more than 10²¹ is needed. Full report · Author's summary
- Retina scaled up · 10¹²–10¹⁵ (or 10¹⁶–10²⁰)
- Work out what it takes to match the retina, then scale up to the whole brain. Hans Moravec's 1998 version gives about 10¹⁴; the same idea using modern deep-network models of the retina gives 10¹⁶–10²⁰, a sign of how sensitive these numbers are. Moravec, 1998
- Kurzweil · 10¹⁴–10¹⁶
- Ray Kurzweil's The Singularity Is Near (2005) derived about 10¹⁴ calculations per second but used 10¹⁶ as a conservative figure in his forecasts. AI Impacts summary
- Brain traffic · ~10¹⁶–3×10¹⁷
- Count how many signals pass between neurons per second, then convert using the ratio computers show between that kind of traffic and arithmetic. AI Impacts notes the conversion is unchecked. AI Impacts, brain performance in TEPS
- Neuron-by-neuron simulation · 10¹⁸ and up
- Anders Sandberg and Nick Bostrom's 2008 whole brain emulation roadmap: 10¹⁸ FLOP/s to simulate every neuron as a spiking model, 10²² with detailed electrical models, 10²⁵ at the level of molecules. This is the cost of copying the brain exactly, which is likely far more than doing the same job. AI Impacts, brain performance in FLOPS
- Energy limit · ~7×10²¹ at most
- An upper bound, not an estimate: the most computation the brain's roughly 20-watt energy budget could pay for under the physical limit set by Landauer's principle (Carlsmith 2020). Full report
One caution: older estimates count "calculations" or "instructions" per second, newer ones count floating-point operations (FLOP/s). They are close enough for a chart that spans ten orders of magnitude, but not the same thing.
A business could buy brain-scale raw power years ago. That did not produce human-level AI on its own, because the software was the hard part. The first chart shows that software is now improving quickly too. NVIDIA's next chip, Vera Rubin, is shipping this autumn with about five times the AI computing power of the previous generation, and next year's models will be trained on it.
05What to expect next
These are trend projections, not certainties. Each could arrive sooner or later.
A free, downloadable model matches today's best paid model (Claude Opus 5.5). A version small enough for a high-end desktop follows a month or two later.
AI agents finish tasks that take a person a full working week, on their own, about half the time.
Tasks that take a person a month of work become possible. Reliability, not raw ability, becomes the main limit on what people trust AI to do.
The first models trained on the new generation of chips arrive, and the cost of today's top-level AI keeps falling.
Two cautions. Benchmark scores are often reported by the companies themselves, and doing well on a test is not the same as doing a job well. Treat any single number as a signal, not a guarantee.
For the latest numbers, updated every morning, see the AI Frontier Tracker.
Sources