Every record, year by year
The AI race
274 models on one scale since Feb 2023. Follow the records, the makers who set them, open models closing in, and what a score costs.
Scroll to beginChapter 1 of 6
Every dot is a model
One scale for every model: Epoch AI’s Capabilities Index.
The index combines results from more than fifty benchmarks into one scale. Each model on it becomes a dot: when it came out, across; how it scored, up.
The field fills in
274 models, from Feb 2023 to Sept 2026. The scale has no top: compare the dots, and don’t read them as percentages.
Who made them
Six makers have the strongest models: Anthropic, OpenAI, Moonshot AI, Google DeepMind, Meta AI and xAI. Every other maker’s models are grey.
The leaders today
- Claude Opus 5.5, Anthropic: 167.3
- GPT-6 Astra, OpenAI: 166.4
- GPT-6.1 Sol, OpenAI: 166.1
- Claude Sonnet 5.5, Anthropic: 165.0
- Claude Fable 5.1, Anthropic: 164.7
Chapter 2 of 6
The record changes hands
Each time a model beat every earlier one, the record moved up.
The first record on the index is Llama 65B, from Meta AI, in Feb 2023, at 110.2.
Up and up
It has been broken 22 times since, and passed from one maker to another 12 times.
Who held it
23 records by 4 makers: OpenAI 15, Anthropic 5, Google 2 and Meta AI 1. The latest is Claude Opus 5.5, at 167.3.
Chapter 3 of 6
Each maker’s climb
The six strongest makers, each against its own best.
Each line follows one maker’s best model to date.
Everyone climbs
Every line only steps up: each step is a model better than all that maker’s earlier ones.
A close finish
Today the six best models are within 10.9 points: from Anthropic at 167.3 to xAI at 156.4.
Chapter 4 of 6
Open weights close in
Models anyone can download, against those only their makers run.
Open-weights models can be downloaded and run by anyone. The best of them today scores 157.4; the best closed model, 167.3.
How far behind
Kimi K3 reached 157.4 in Jul 2026. A closed model first scored that much 4 months earlier: GPT-5.4 Pro.
Part two: price
Getting better is half the story.
Chapter 5 of 6
The price of a score
What the cheapest model at a given level costs, and how fast that falls.
Each model is placed at its release date with today’s list price per million tokens, blended three input tokens to one output token: not its launch price. The line follows the cheapest model at or above one level of the index; the scale falls by ten times at each gridline.
Prices: models.dev (MIT)
477 times cheaper
Of the models released by Dec 2024, the cheapest scoring 140 or more is o1, listed at $26.25. Of those released by Jul 2026, it is Qwen3.7 Flash, at $0.055.
At every level
Each line is one level of the index, from 120 up. Every one only steps down: each point is a new cheapest model at that level.
Chapter 6 of 6
Where the race stands
The highest scores today, and how sure the index is of each.
- Claude Opus 5.5, Anthropic: 167.3
- GPT-6 Astra, OpenAI: 166.4
- GPT-6.1 Sol, OpenAI: 166.1
- Claude Sonnet 5.5, Anthropic: 165.0
- Claude Fable 5.1, Anthropic: 164.7
- Claude Opus 5, Anthropic: 162.8
- GPT-6 Sol, OpenAI: 162.7
- GPT-5.5 Pro, OpenAI: 162.1
The race goes on
Who leads next?
Claude Opus 5.5, from Anthropic, holds the record at 167.3.
This page is drawn from the same data as the rest of the site, so new models appear here as they join the index.
Sources and method
Scores are the Epoch Capabilities Index, with release dates and weight access as Epoch AI records them (CC BY 4.0). Prices are today’s list prices per million tokens from the models.dev catalog (MIT), placed at each model’s release date, not launch prices: the maker’s own listing, or the middle listing across hosts, blended three input tokens to one output token.
A record is a model that scored higher than every model released before it. Artificials draws these charts from the published data; it does not test models itself.