No single list names the top artificial intelligence systems, because "top" depends on the yardstick: training compute, citations, user numbers, or test scores each produce a different ranking. The most systematic public record is Epoch AI's database of more than 3,600 machine learning models, which labels a model "frontier" if it ranked among the five largest by training compute when it was released. Epoch AI's trend analysis, updated in February 2026, shows compute for the largest language models growing about fivefold per year since 2020.
Key takeaways
- Epoch AI's public dataset tracks more than 3,600 models from 1950 onward and was last updated on October 26, 2026.
- In that dataset, "frontier" means top five by training compute at release, and "large-scale" means at least 10^23 floating-point operations.
- Epoch AI reports that training compute for frontier language models has grown about 5 times per year since 2020, a doubling roughly every 5.2 months.
- Compute measures size, not quality, safety, or usefulness, so it answers only one version of the "top systems" question.
- Compute figures carry stated error bands of 3 times, 10 times, or 30 times depending on how they were estimated.
Why a single ranking does not exist
Ask which AI system is best and the answer changes with the question. A system can lead on the number of people who use it, on a particular exam, on the cost of running it, or on the scale of the computing power spent building it. These measures rarely agree, and each has a different owner: usage comes from company disclosures and surveys, test scores from benchmark publishers, and size from research groups that collect training records. Without a stated yardstick, a list of "the top systems" is an opinion with a date attached, and it ages quickly.
That is why the more careful sources avoid a podium. They publish a table of records and let readers sort it. The practical question for a reader becomes which column matters for the decision at hand: choosing a tool for a job, understanding who can afford to build frontier systems, or judging how fast capabilities are moving.
What Epoch AI counts and how
Epoch AI, a research organization that studies AI trends, maintains its database of notable AI models under a Creative Commons Attribution license, with the full file refreshed daily. Its documentation sorts models into overlapping groups. "Notable" models meet at least one criterion from a list that includes a state-of-the-art result on a recognized benchmark, a high citation count, historical relevance, or significant use. "Frontier" models were in the top five for training compute at the time of release. "Large-scale" models were trained with at least 10^23 floating-point operations, a unit of arithmetic work.
Because frontier status is defined by compute at release, the list is a snapshot of who could afford the largest training runs at each moment. A system can drop off the frontier without getting worse, simply because newer ones are bigger.
How reliable the numbers are
Developers do not always disclose training compute. Epoch AI therefore takes figures from publications where it can and otherwise estimates them from the model architecture and training data, or from the hardware used and the training time. Each record carries a confidence label. A "Confident" estimate is described as within a factor of 3 of the true value, "Likely" within a factor of 10, and "Speculative" within a factor of 30. A factor of 30 is a wide band, so small differences between two models in the dataset should not be read as a ranking.
The growth trend behind the word "frontier"
On its trends page, last updated February 5, 2026, Epoch AI reports that training compute for frontier language models has grown about 5 times per year since 2020, with a 90 percent confidence interval of 4 to 6 times. That corresponds to a doubling roughly every 5.2 months, and to roughly a 10,000-fold increase among the top five models since 2020. The same page reports training costs rising about 3.5 times per year, a doubling about every seven months.
Two cautions apply. Trend lines describe the past, and the page does not claim they must continue. And compute is an input: more of it does not guarantee a better system, only a more expensive one to build.
Illustrative example: four lenses on "top"
The table below is an illustrative framing, not a ranking. It shows how the same question splits depending on the measure.
| Lens | Typical source | What it captures | What it misses |
|---|---|---|---|
| Size at release | Epoch AI database (training compute) | Who could fund the largest training runs | Quality, safety, usefulness |
| Notability | Epoch AI criteria (citations, benchmarks, use) | Influence on research and practice | Models that are good but little known |
| Test performance | Benchmark publishers and leaderboards | Results on a fixed set of tasks | Real-world fit, test contamination |
| Adoption | Surveys and company disclosures | What people actually use | Capability, since popularity can follow marketing |
A short checklist for reading any "top AI systems" list
- Is the measure named, and does it match the question being asked?
- Is there a date, given how quickly rankings change?
- Is the source a dataset with documented methods, or a list without a method?
- Does the list separate what was measured from what was estimated?
- Does the publisher sell any of the systems on the list?
This article describes public research data and does not recommend or rank any product.
FAQ
Which AI system is the largest right now?
That depends on the date and on the measure. Epoch AI's frontier list is keyed to training compute at release and is updated as new models are added, so the current dataset, not a static article, is the place to check.
Does more training compute mean a better AI system?
Not by itself. Compute describes how much work went into training. Quality, accuracy, and safety are measured separately, and a smaller system can outperform a larger one on a given task.
Can Epoch AI's figures be reused?
The dataset is licensed under Creative Commons Attribution, which allows reuse with credit to the source and authors. Anyone citing a figure should state the update date.
Are all AI systems in the dataset chatbots?
No. The database covers machine learning models back to 1950, including image, speech, game-playing, and scientific models, not only language systems.
Related reading
- Well-known AI technologies and what they do
- Whether AI could dominate in the near future
- Measured benefits of AI
- AI and self-driving cars
More in AI Basics.
Last verified: October 2026
Sources
- Data on AI models (notable AI models), Epoch AI, updated October 26, 2026
- Data on AI models documentation, Epoch AI, 2026
- Key trends and figures, Epoch AI, February 5, 2026