Epoch Capabilities Index Explained: What an ECI of 169 Means (Not an IQ)

The Epoch Capabilities Index (ECI) is a single score for an AI model's general capability, built by Epoch AI from the model's results on dozens of benchmarks. In the data files Epoch updated on 23 October 2026, Claude Opus 5.5 is first at 168.83, GPT-6 Astra second at 168.81 and GPT-6.1 Sol third at 168.37. Those three scores are less than half a point apart, and the confidence interval around each is about nine points wide, so the ranking among them is not settled. The ECI is also not an IQ: its two fixed points, 130 and 150, are AI models rather than a human norm group. This guide sets out how the index is built, who leads, how fast the top has risen, and what the numbers can and cannot be compared with.

What is the Epoch Capabilities Index?

Epoch AI, a research group that tracks AI progress, describes the ECI as a composite that combines scores from over 50 distinct benchmarks into a single general capability scale. Individual benchmarks saturate within months: once the best models score near 100%, the test can no longer tell them apart. The ECI is meant to keep a ruler that still works after that. The method comes from a paper by Anson Ho, Jean-Stanislas Denain, David Atanasov, Samuel Albanie and Rohin Shah, A Rosetta Stone for AI Benchmarks (arXiv, 28 November 2025). The authors are at Epoch AI and Google DeepMind, and the paper states that Google DeepMind financially supported the work.

The model behind it is compact. Each model gets one capability number and each benchmark gets a difficulty and a slope. The predicted score of a model on a benchmark is a sigmoid of the slope times the gap between capability and difficulty: a model far above a benchmark's difficulty scores near 100%, one far below scores near zero, and in between the score rises steadily. Epoch's public code, written as performance = sigmoid(discriminability x (capability – difficulty)), fits all models and benchmarks together. The paper says this is similar in spirit to Item Response Theory, the family of methods used to calibrate human test items, with one difference that matters: IRT works on pass-or-fail results for single questions, while the ECI uses each benchmark's average score as its input.

Raw fitted values have no natural unit, so Epoch rescaled them. The page says the raw scores were scaled so that Claude 3.5 Sonnet equals 130 and GPT-5 equals 150, and the data file confirms it: Claude 3.5 Sonnet (released 20 June 2024) is exactly 130.00 and GPT-5 (7 August 2025) exactly 150.00. Because both are pinned by definition, they are the only two models in the file without a confidence interval. For every other model Epoch publishes an interval; its code describes these as quantiles of bootstrap draws that resample each model's benchmark results.

Which AI model has the highest ECI?

We read the three data files on the page, model scores, benchmark results and benchmark difficulty, on 23 October 2026. The file lists 270 models released between 24 February 2023 and 15 October 2026. The top ten, with the interval Epoch gives for each:

• 1. Claude Opus 5.5: 168.83 (165.25 to 173.95), released 22 September 2026
• 2. GPT-6 Astra: 168.81 (165.45 to 174.51), 3 September 2026
• 3. GPT-6.1 Sol: 168.37 (165.00 to 174.43), 29 September 2026
• 4. Claude Fable 5.1: 166.49 (163.47 to 171.35), 1 September 2026
• 5. Gemini 4 Argon: 166.45 (163.23 to 171.52), 30 September 2026, limited access
• 6. Claude Sonnet 5.5: 166.18 (163.06 to 171.34), 28 September 2026
• 7. Claude Opus 5: 163.94 (161.12 to 168.24), 24 July 2026
• 8. GPT-6 Sol: 163.73 (160.92 to 167.44), 22 September 2026
• 9. Claude Fable 5: 163.42 (160.53 to 168.18), 9 June 2026
• 10. GPT-5.5 Pro: 162.91 (160.04 to 166.60), 23 April 2026

The gap between first and tenth is 5.9 points; between first and twentieth, 11.4. The median interval across the 268 models that have one is 6.7 points wide, and for the top ten it averages 7.9. When two models are 0.02 points apart, as Opus 5.5 and GPT-6 Astra are, a reader should treat them as tied. The intervals of 12 models, the leader included, overlap the leader's interval. The same caution applies to any ranking built on a few hundred benchmark scores, as our guide to why AI benchmark scores disagree explains.

Behind each score is less data than the headline suggests. The benchmark file holds 2,888 results, a median of 9 per model. The five models at the top have between 18 and 24 results each out of the 59 benchmarks, so even the leaders are measured on about a third of the set. The ECI is built to cope with that overlap, and it is the reason the intervals are as wide as they are.

How fast is the ECI rising?

Sorting the models by release date and keeping each new record gives 24 record-setters, from o3 in April 2025 up to the present. The latest are Claude Fable 5 (163.42, 9 June 2026), Claude Opus 5 (163.94, 24 July 2026), Claude Fable 5.1 (166.49, 1 September 2026), GPT-6 Astra (168.81, 3 September 2026) and Claude Opus 5.5 (168.83, 22 September 2026). The two anchors give a clean yardstick: the 20 points from Claude 3.5 Sonnet at 130 to GPT-5 at 150 took 413 days, and the next 18.8 points, from GPT-5 to Opus 5.5, took 411 days.

• Best ECI released by 1 January 2024: 125.74; by 1 January 2025: 141.72; by 1 January 2026: 156.33; by 23 October 2026: 168.83
• Models at 150 or above: 61, of which 54 were released in 2026 and 7 in 2025
• Best open-weights model: Kimi K3 at 157.86 (16 July 2026), 11.0 points below the best overall
• The frontier first passed 157.86 with GPT-5.4 Pro (159.57) on 5 March 2026, 133 days before Kimi K3 appeared
• Next open-weights models: GLM-5.3 155.99, DeepSeek V4.1 Flash 155.93, DeepSeek V4 Pro 0813 155.64, Muse Spark Open 155.11

A straight-line fit of the 22 record-setting models released since 1 January 2024 against their release dates gives a slope of 16.2 ECI points a year, with an R-squared of 0.99; across all 24 record-setters it is 15.1. Those fits are ours, not Epoch's. A running maximum is smooth by construction, which is why the R-squared is so high, and the fit says nothing about what the next year will bring. It does give a rough scale for the headlines: a lead of two or three points between two labs is about two months of frontier progress, and the 11-point gap between open-weights and the overall best is under nine months at that rate.

What do the numbers 130 and 150 mean?

They mean "Claude 3.5 Sonnet" and "GPT-5". That is all. They were chosen, Epoch says, to put values in a convenient range, and they resemble the IQ scale's landmarks only by looks. Across the 270 models the ECI has a mean of 132.5 and a standard deviation of 21.2; for the 85 models released in 2026 it is 153.2 and 7.1. Those figures describe a crowd of AI models that keeps being replaced by better ones. An IQ scale is defined the other way round: a representative sample of people sets the mean at 100 and the standard deviation at 15, and a person's score says where they stand among those people, as the IQ bell curve shows.

It is tempting to ask what 168.83 would be as an IQ. Read as one, with mean 100 and SD 15, it is a z-score of 4.59, a rarity of about 1 in 448,000 on the IQ percentile calculator's logic, and the next frontier model would be rarer still. The calculation is shown only to make the point that it is meaningless. There are no people behind the ECI's mean, its spread depends on which models are in the file, and it has no ceiling. Epoch states the limit itself: absolute ECI values are meaningless by themselves, but meaningful comparisons can be made between models. See our look at why scoring an AI on an IQ-style test is not an IQ for the same problem from the other side.

How hard are the benchmarks behind the ECI?

The same scale carries a difficulty for every benchmark, in a third file. Of the 59 benchmarks, 29 have a difficulty below 150, the GPT-5 anchor, and 14 sit below 130, the Claude 3.5 Sonnet anchor; the easiest are TriviaQA (57.4), LAMBADA (58.5), PIQA (86.3) and HellaSwag (87.0). Only six benchmarks are harder than the current leader's 168.83: CL-bench (184.8), Remote Labor Index (178.7), CL-bench Life (177.9), SpireClimb (176.5), PostTrainBench (173.5) and OSWorld 2.0 (170.2). WeirdML v3 (167.6) and Furniture Assembly (165.2) sit just under the top. A benchmark whose difficulty is near a model's ECI is where that model's score is most informative; far below it, the test is saturated.

The classic exams people know are in the middle. For how the question-set benchmarks behave at the top of the scale, see our guides to Humanity's Last Exam scores and the GPQA Diamond score; for a composite built a different way, the Artificial Analysis Intelligence Index; and for a measure of task length rather than exam accuracy, METR's time horizon.

Is the ECI an IQ test for AI?

Epoch's own page uses the comparison, saying that, similar to an IQ test, the ECI is designed to capture a broad underlying capability that is useful across many tasks, rather than performance on specific skills. That is a statement about intent. An IQ test reports where a person stands in a norm group, is built from questions that were new to that person, and has a published reliability. The ECI reports a fitted position among models, uses public benchmarks that labs can and do tune for, and moves when the benchmark list changes. The paper tests this directly by comparing benchmarks that labs report with those they do not, and its authors list the set of benchmarks as a source of bias.

The page lists further limits: values are not linearly related to benchmark accuracy because benchmarks differ in difficulty and slope; highly specialised models can get low scores despite being very capable in their domain; and developers can optimise for particular benchmarks. Whether AI abilities share one general factor the way human abilities do is a separate research question, covered in is there a g factor for AI models. And an ECI says nothing about the popular versions of the question, such as ChatGPT's IQ score or whether AI is smarter than humans.

How to read an ECI in a headline

• Name the date and the data file: ECI values shift as models and benchmarks are added
• Ignore gaps smaller than the interval width, about 7 to 9 points near the top
• Compare models with each other, never with a human IQ: the scale has no human norm group
• Check which benchmarks a model was run on: leaders have 18 to 24 of the 59
• For pace, use differences over time, such as 16 points a year since 2024, not the level

For the index that tops most "most intelligent AI" headlines, read which AI model the index calls most intelligent, and for the ARC-AGI puzzle tests that set a human baseline, the ARC-AGI-3 score and its two harnesses. To see where a human result lands, take the IQ Metrics test and read it against the IQ bell curve.

Common questions