Methodology · As of October 12, 2026
How we compute it
Comparison sites run on trust. So we disclose everything here: the data sources, the formula, the weightings and their limits. No paid placements, no affiliate links, and every recommendation is decided by an open formula, not a language model.
The data sources
Benchmarks
Intelligence Index, speed and model prices come from the Data API of Artificial Analysis, an independent benchmark institute that measures all major models with the same tests. One automatic fetch per day, around 6:30 in the morning.
Product facts
We maintain euro prices (Perplexity lists US dollars only), free tiers, Team plans, and data-protection details by hand, each backed by a source at the provider. Benchmarks measure models; which model sits inside ChatGPT, Claude or Copilot and what the subscription costs is product knowledge.
From the sources
The data today
This is what the raw data looks like that every recommendation builds on: intelligence, speed and cost per task of the strongest models.
Each bar is one AI model, coloured by provider. “Self-hostable” means models with open weights you can run on your own infrastructure. Prices in euro (ECB reference rate), cost per task based on a typical 2,000-token task. See the methodology.
The formula
Deterministic and reproducible. No language model decides the recommendation.
Score = (Intelligence Index of the strongest model ÷ 100) × use-case factor
Before that, your answers filter the candidates: with a data-protection requirement only systems remain whose Team plans do not use inputs for training (per the provider). For the image use case, systems without image generation drop out. The system with the highest score wins, the runner-up is named as the alternative.
Systems without a benchmarked model of their own inherit a base: Copilot the index of the best OpenAI model (it uses OpenAI technology), Perplexity the best value in the field (it taps into several top models). Both are written exactly like this in the source code of the calculation.
Prices in euro and cost per task
We work in Europe, so we convert all model prices into euro. Artificial Analysis states prices in US dollars per 1M tokens; we convert them with the daily ECB reference rate (currently €1 = $1.1206 US dollars).
“Cost per task” makes the prices tangible: we assume a typical task of about 2,000 tokens (roughly 1,500 input and 500 output, the 3:1 ratio matches the blended price from Artificial Analysis) and show the cost in euro cents. So you can see at a glance that a task costs between fractions of a cent and a little over one cent, depending on the model.
The model charts and the top 10 collapse multiple reasoning tiers of the same model to its strongest value. “Self-hostable” means models with open weights you can run on your own infrastructure; the prices shown there refer to hosted offerings of these models.
The use-case factors
Benchmarks measure intelligence, but not product fit. The factors capture how well a system is built for a task: Excel integration helps with spreadsheets, source search with research, missing image generation is a knockout for images (factor 0). We review the factors editorially; here they are in full.
| System | Writing text | Coding | Research | Creating images | Spreadsheets & data analysis | Presentations |
|---|---|---|---|---|---|---|
| ChatGPT | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.90 |
| Claude | 1.10 | 1.10 | 1.00 | 0.00 | 1.00 | 0.90 |
| Gemini | 0.90 | 0.85 | 1.00 | 1.10 | 1.00 | 1.00 |
| Microsoft Copilot | 0.85 | 0.80 | 0.85 | 0.80 | 1.05 | 1.10 |
| Perplexity | 0.70 | 0.50 | 1.20 | 0.30 | 0.60 | 0.40 |
| Le Chat | 0.85 | 0.80 | 0.80 | 0.60 | 0.80 | 0.60 |
Limits and principles
- Independence: No paid placements, no affiliate links, no commissions from providers. The AI Leaderboard is a project of the digital agency Collective Brain; we earn from consulting and implementation, not from your tool choice.
- Resilience: If the benchmark API delivers no data on a given day, the previous day's values stay, and the date stamp transparently shows the older date. This last happened between August 28 and October 1, 2026: our daily data job did not complete, and the pages showed the data as of August 27. We do not backfill missing days or months.
- Benchmark limits: An index measures test tasks, not your everyday work. It is the best available basis for comparison, but no substitute for trying it in your own use case.
- Index versions: Artificial Analysis revises the Intelligence Index from time to time, most recently in September 2026 with new tests and new weights (versions 4.2 to 4.3.2, see the version history). The same models score noticeably lower afterwards: Kimi K3 stood at 60 on August 27 and at 44 on October 1. Scores from different versions are therefore not directly comparable.
- Prices and legal: Price figures come from the providers' pages, can change and are without guarantee. Data-protection details are based on provider information and are not legal advice.
- Brand names (ChatGPT, Claude, Gemini, Copilot, Perplexity, Le Chat) belong to their respective owners.
Behind the scenes
We build things like this for you, too.
The AI Leaderboard is an automated data product: a daily data pipeline, deterministic logic, surgical deployment. Exactly these kinds of systems, from content machines to data dashboards, Collective Brain builds for companies.