Our methodology

Our principles

Info Agent IA has a simple goal: to help you understand, compare and choose AI tools based on information you can check. Four principles guide our work.

• Public data, dated and sourced. Every price, score or spec comes from an identified public source. The date of the data is shown, so you always know how current it is.
• Separating facts from analysis. A price table is a fact; saying a model suits a given use better is analysis. We make the difference clear and flag uncertainty where it exists.
• No made-up tests. We never claim to have tested a model or assistant unless that test is documented. The scores we show come from public evaluations, not impressions. When a text suggests that a type of model suits a given use, it is general guidance: it does not replace trying the tool yourself with your own documents and questions, which remains the best test.
• Prompt corrections. A reported and confirmed error gets fixed. Automated data is corrected at the next site rebuild; editorial text, as soon as possible.

Our sources

The site relies on the following sources. Full attributions are shown below this text.

Source	What we use it for
Models.dev (open source project)	Prices, context window, max output, features and release dates of models
OpenRouter (public model list)	Cross-checking prices
Epoch AI — AI Benchmarking Hub	Benchmark scores, the basis of the ranking
Epoch AI — Data on AI Models	Notable models in AI history, for the timeline
Hugging Face Hub	Most downloaded open source models: license, size, downloads
GitHub	Stars, license and latest release of agent frameworks
arXiv	Latest research papers on AI agents
Hacker News	Recent technical discussions
Our World in Data and Eurostat	AI statistics, including business use of AI in the European Union
Official assistant websites	Prices and features of consumer AI assistants

Update schedule

The site is rebuilt every Monday. Each rebuild fetches all automated data again: prices, features, scores, the ranking, and lists of models and frameworks.

If a source is unavailable at rebuild time, we do not show an empty page or partial data: the last valid version is used, and the actual date of the data is displayed alongside it. You can always tell whether a piece of information is from this week or a little older.

How the ranking works

The AI ranking is recalculated at every weekly rebuild, using Epoch AI’s Benchmarking Hub.

1. A fixed set of benchmarks. A benchmark is a standardized test given in exactly the same way to every model. The set covers scientific reasoning, mathematics, coding and agentic tasks in a terminal (the model chains commands on a computer by itself to reach a goal). The set is fixed so results stay comparable from week to week.
2. Min-max normalization per benchmark. On each benchmark, the best evaluated model gets 100 and the weakest gets 0, with the others placed proportionally in between. This lets us combine tests that use different scoring scales.
3. Average of normalized scores. A model’s composite score is the average of its normalized scores on the benchmarks available for it.
4. A 4-benchmark threshold. A model needs results on at least 4 benchmarks in the set to be ranked. This prevents a model tested on a single benchmark from landing at the top by chance.
5. Display. The table shows the top 30, sortable and filterable (open or closed, lab, country).
6. Changes. Arrows are calculated against the ranking published the previous week. If that ranking does not exist, or the model was not in it, no arrow is shown.

The composite score is relative: it compares evaluated models with one another. A strong new arrival can lower other models’ scores even though their results have not changed. It does not measure price, speed or real-world experience, and it can be affected by test contamination (benchmark questions finding their way into a model’s training data).

Model selection

The AI models index cannot list every variant in existence. A model page is created for:

• models from the 33 labs we follow, as listed on models.dev;
• released within the last 24 months, or still offered by their maker;
• one page per underlying model: a model offered by several hosting providers is not duplicated;
• with priority given to models evaluated by Epoch AI.

Models that do not meet these criteria remain available in the model comparison tool and the cost calculator.

Model comparisons

Our comparison pages pit models from the same tier against each other, so the comparison makes sense:

• open weights: models whose weights can be downloaded;
• frontier: other models whose composite score is in the top 30% of scored models or, without a score, whose output price is at least $10 per million tokens;
• budget: remaining models whose output price is at most $2 per million tokens;
• mid-range: all the others.

We also compare pairs of consumer assistants, for readers choosing an app rather than a model.

Quiz recommendations

The which AI is right for me quiz calculates a weighted score for each assistant based on your answers: budget, use, device, privacy, skill level and language. The calculation runs in your browser and your answers are not sent anywhere. No company is ever paid to be recommended.

Here are the points awarded:

• Main use: +4 if it is one of the assistant’s stated strengths, +2 if it offers the feature that matters for that use (web search, file upload, image generation, agent mode or browser, code execution; nothing for writing, which no feature measures).
• “Free only” budget: +3 if the free version is complete, +2 if it is limited, −5 if the information is missing; an assistant with no free version is left out.
• “Under $10” budget: +2 if a paid plan costs less than 10 in its own currency (no conversion), otherwise +2 for a complete free version or +1 for a limited one, otherwise −3.
• “$20 or more” budget: +1 if the assistant has an agent mode.
• Device: on a phone, +2 with both an iPhone and an Android app, +1 with one, −2 with none; on a computer, +1 if it works on the web, Windows or macOS.
• Privacy: for each setting offered (training opt-out, temporary chat, data stored in Europe), +2 if privacy matters a lot to you, +1 if it matters a little.
• Skill level: for beginners, +1 for a complete free version and −2 for a coding-only tool; for advanced users, +1 for code execution and +1 for agent mode.
• Language: in French or Spanish, +2 if the interface exists in your language, −3 if it does not, 0 if the information is missing; no adjustment for English.

Ties follow the most widely used assistants (ChatGPT, Claude, Gemini, Copilot, Le Chat, Perplexity, Meta AI, Grok, DeepSeek, Qwen Chat, Kimi, Duck.ai, Lumo), then alphabetical order. The three reasons shown are the criteria that earned the most points.

Generated profiles and comparisons

Model profiles and comparisons are written automatically from the data above: every sentence relies on a recorded value (price, context, benchmarks, release date), and missing data is flagged as such rather than guessed. Profiles of assistants, labs, frameworks and glossary terms are written by our editors from official pages, then completed with automated data (GitHub stars, a lab’s models). Every page shows its sources and the date of its data. A profile may show a composite score based on only 3 benchmarks, but a model needs at least 4 to enter the public ranking.

Token estimates

A token is a small chunk of text, the unit models use to read, write and bill. The token estimator applies a language-specific heuristic based on the number of characters and words in your text. Each model splits text in its own way, so the result is approximate: good enough to gauge an order of magnitude or a cost, not for exact billing.

The coefficients used: about 4 characters or 1.3 tokens per word in English, 3.8 characters or 1.5 tokens per word in French, 3.9 characters or 1.5 tokens per word in Spanish, 3.5 characters or 1.6 tokens per word for other languages, and 3 characters per token for code. The estimate is the average of the two methods, with a ±15% range. Each Chinese, Japanese or Korean character counts as one token. A text is treated as code when at least 40% of its lines look like code.

Context window and agent cost

The context window tool converts documents into tokens with these benchmarks: 0.75 words per token, 500 words per page, a 300-page novel, 9,000 words per hour of transcribed meeting (150 words a minute) and 10 tokens per line of code. The agent cost calculator multiplies input and output tokens by the prices providers publish; the share of cached calls is billed at the cache read price when one is published, otherwise at the normal input price.

Running models locally

On the open source models page, the memory needed is estimated as: billions of parameters × bytes per parameter (2 at 16-bit, 1 at 8-bit, 0.5 at 4-bit) × 1.2 for the rest of the software. The verdict is green if this estimate stays within 80% of the memory you enter, orange between 80 and 100%, red above that. It is an order of magnitude: a long context needs more.

Assistant profiles

Our AI assistant profiles (subscription prices, features, availability) are collected from each company’s official pages. Every profile shows its verification date, and all profiles are checked every month. Plans change often, so always confirm the price on the official site before subscribing.

Independence

There are no sponsored links in our rankings. No assistant or lab pays to be recommended, ranked higher or featured. Rankings come from the same automated calculations for everyone, using the sources listed above.

Report an error

Does a figure look wrong or out of date? There is a form below this text. Tell us the page, the information in question and, if possible, a link to an official source. Every report is reviewed; a confirmed correction is applied and, for automated data, reflected at the next rebuild.

The Info Agent IA editorial team