This week's AI ranking

How to read the ranking

Each row in the table is an AI model, the “engine” that powers an assistant such as ChatGPT, Claude or Gemini. The main column is the composite score: a mark out of 100 that sums up how the model performed across several public tests. The higher it is, the better the model did compared with the other models evaluated.

A few pointers to get the most out of the table:

• Arrows show whether a model has gained or lost places since the previous ranking. A model appearing for the first time is marked “new”. In the very first edition there is no previous ranking yet, so the movement column only appears from the following update.
• Filters let you show only open models (whose “weights”, the parameters learned during training, can be downloaded) or closed ones (available only through their maker’s service), a specific lab, or a given country.
• Sorting works by clicking a column header: you can, for example, rank models on a single benchmark instead of the overall score.
• Hovering over a benchmark name shows a short explanation of what it measures.

Below the table, the “best score by month” chart shows how the top model’s level changes over time, and the history of number ones lists which models have held first place. Both views help you keep perspective: the top spot is often won by a narrow margin and changes hands regularly.

What is a benchmark?

A benchmark is a standardized exam for AI models. Every model gets the same set of questions or tasks, the answers are graded automatically, and the result is a score you can compare from one model to the next. Think of it as a standardized test: every candidate sits the same paper under the same conditions.

The ranking relies on a fixed set of benchmarks covering several skill areas:

• scientific reasoning, with advanced science questions;
• mathematics, with problems that require several reasoning steps;
• coding, where the model has to write or fix programs;
• agentic tasks in a terminal, where the model acts as an agent: it chains commands on a computer by itself to reach a goal, without step-by-step guidance.

Our method in brief

Raw scores come from Epoch AI’s Benchmarking Hub, an independent research organization that publishes its evaluations openly. At each weekly rebuild of the site, on Monday, the calculation follows three steps:

1. Normalization per benchmark: on each test, the best model gets 100 and the weakest gets 0, with the others placed proportionally in between. This “min-max” normalization makes tests with very different scoring scales comparable.
2. Average: a model’s composite score is the average of its normalized scores on the benchmarks it was evaluated on.
3. Minimum threshold: a model only enters the ranking if it has results on at least 4 benchmarks in the set, so that a single strong result cannot push it to the top.

The table then shows the top 30. The full details, including how week-to-week changes are calculated, are on our methodology page.

What the ranking does not tell you

A composite score is a useful signal, not a verdict. Keep its limits in mind:

• A score is not your real-world experience. Benchmarks measure specific skills under controlled conditions. To draft an email, summarize a report or chat day to day, a lower-ranked model may suit you perfectly and cost far less.
• Test contamination. If a benchmark’s questions ended up in a model’s training data, the model may “know the answers” without truly reasoning. Evaluators work to reduce this risk, but nobody can rule it out entirely.
• Models not evaluated. A model missing from the table is not necessarily weaker: it may not have been tested yet, or not on enough benchmarks. Very recent models often show up with a delay.
• Narrow gaps. When two models are a few points apart, you may not notice any difference in practice. Scores also depend on which models are included: a new leader can shift everyone else’s scores even though they have not changed.
• Price, speed and privacy are not counted. The ranking only reflects test results.

What to do next

The ranking gives you a first idea of the strongest models. To go further, depending on your needs:

• Torn between a few models for a project? Put them side by side in the model comparison tool, which lines up price, context window, features and scores.
• Want to browse every model page and filter by lab or price? Explore the AI models index.
• Use AI every day without caring about benchmarks? The which AI is right for me quiz points you to an assistant that fits your use, budget and device.
• Stuck on a term? The AI glossary explains it in plain English.