No law ranks AI systems by intelligence. Regulators draw the line for "most advanced" with a number instead: under Article 51 of the EU AI Act, a general-purpose model trained with more than 10^25 floating point operations is presumed to have high-impact capabilities, which brings the Act's systemic-risk duties. A 2023 US executive order used 10^26 operations as its reporting line and was rescinded on January 20, 2025.
Key takeaways
- EU law presumes that a general-purpose model trained with more than 10^25 floating point operations has high-impact capabilities, and the European Commission can amend that figure.
- The Commission also uses an indicative scope test of 10^23 operations for what counts as a general-purpose model at all, and calls it "not an absolute rule."
- Epoch AI counted 33 publicly announced models estimated to exceed 10^25 operations as of June 2025, with many figures flagged as low-confidence.
- Compute is a proxy: Annex XIII lists seven criteria for designating a model as systemic-risk, including parameters, data size, modalities, benchmarks, and reach.
Why a count of operations stands in for capability
Measuring how capable a model is requires testing it, and testing happens after training. A law that must apply before a model is released needs something known earlier. The amount of computation used in training is known to the developer and can be stated as one number. That is the logic of a threshold, and it explains why the rules below quote floating point operations (FLOP), which are individual arithmetic steps such as one addition or multiplication on decimal numbers.
Three thresholds side by side
| Rule | Number | What it does | Status on October 24, 2026 |
|---|---|---|---|
| EU AI Act, Article 51(2) | More than 10^25 FLOP of cumulative training compute | Presumes high-impact capabilities, so the model is treated as a general-purpose model with systemic risk | In force; the Commission may amend the figure by delegated act |
| Commission guidelines on general-purpose models | More than 10^23 FLOP, plus ability to generate language, images from text, or video from text | Indicative test for whether a model is general-purpose at all | Guidance; the Commission says it is not an absolute rule |
| US Executive Order 14110, section 4.2(b)(i) | More than 10^26 integer or floating-point operations | Interim condition for reporting by developers of dual-use foundation models | Rescinded January 20, 2025 |
The EU duties for general-purpose model providers began to apply on August 2, 2025, and the Commission's enforcement powers, including fines, take effect on August 2, 2026, according to the Commission's guidelines page.
How large 10^25 is
An illustrative calculation helps. Suppose one accelerator chip sustains 10^15 operations per second (one petaFLOP per second), a round figure chosen for arithmetic, not a measured chip. Reaching 10^25 operations would take 10^10 chip-seconds, which is about 317 years on a single chip. Spread across 10,000 such chips running without pause, the same work takes about 10^6 seconds, or roughly 11.6 days. Real training runs fall short of peak speed, so actual durations are longer.
The point of the exercise is scale. Epoch AI estimates that training a model of this size costs tens of millions of dollars with current hardware, which is why the line captures only the largest developers.
How many models cross the line
Epoch AI, a research group that tracks training compute, reviewed recent releases, benchmark leaders, and popular downloads, and listed 33 publicly announced models estimated to exceed 10^25 operations as of June 2025, from 12 developers. It describes GPT-4, released in March 2023, as the first model trained at this scale, and it identified no image or video model above the threshold. The group warns that many figures are speculative. Benchmark-based estimates averaged within a factor of 2.2 of known values in its cross-checks, and models whose developers disclose little cannot be placed with confidence. Undisclosed training runs are not counted at all.
Why the number cannot settle the question
The Commission itself treats compute as an imperfect stand-in. Its question-and-answer page says the 10^23 criterion is not absolute: a model above it may fall outside the general-purpose definition if it lacks significant generality, and a model below it may fall inside if it shows that generality. For the higher line, Article 51(1) gives a second route. The Commission can designate a model as systemic-risk on its own initiative or after a qualified alert from the scientific panel, using Annex XIII.
Annex XIII names seven factors: parameter count, data quality or size, training computation, input and output modalities, benchmark results, impact on the internal market (presumed when at least 10,000 registered business users in the Union can use the model), and the number of registered end users. A model with modest compute but wide reach, or strong benchmark results, can therefore be designated even without crossing 10^25. Better algorithms also mean the same capability may need less compute over time, which is why the Commission may adjust the figure.
FAQ
Is there an official most advanced AI?
No source checked for this article names one. The legal texts above define a class of models by size and impact, and they do not rank individual products.
Does more training compute always mean a smarter model?
No. Compute correlates with capability across large models, but it is an input, not a result. Data quality, architecture, and post-training all matter, which is why Annex XIII lists several criteria beside compute.
Can a model below 10^25 operations still be regulated as systemic-risk?
Yes. Article 51(1) allows designation by the Commission using the Annex XIII criteria, regardless of the presumption in paragraph 2.
What happened to the US threshold?
Executive Order 14110 set an interim reporting condition at 10^26 operations. The White House order Initial Rescissions of Harmful Executive Orders and Actions lists that order among those rescinded. This article did not verify a replacement federal compute threshold.
This article explains public rules for general readers and is not legal advice. Anyone who needs to know whether a specific model is covered should consult the legal texts or counsel.
Related reading
- ChatGPT as a conversational assistant
- Claims about the most advanced robot
- AI decision support systems
- AI in neurosurgery
More in AI Basics.
Last verified: October 2026
Sources
- Article 51, Classification of general-purpose AI models as general-purpose AI models with systemic risk, Regulation (EU) 2024/1689, European Commission AI Act Service Desk
- Guidelines on obligations for general-purpose AI providers, questions and answers, European Commission, page updated November 11, 2025
- Annex XIII, Criteria for the designation of general-purpose AI models with systemic risk, Regulation (EU) 2024/1689, European Commission AI Act Service Desk
- Guidelines for providers of general-purpose AI models, European Commission, policy page updated April 28, 2026
- Executive Order 14110, Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, October 30, 2023, 88 Federal Register 75191
- Initial Rescissions of Harmful Executive Orders and Actions, The White House, January 2025
- Models over 10^25 FLOP, Epoch AI, data insight updated June 6, 2025