Methodology
RankingIA does not run its own tests. It combines public leaderboards that measure different things and shows them in one place, next to price and context. Data is collected once a day.
Sources
- LiveBench: objective benchmark with new questions every release to avoid contamination. We use the LiveBench 2.0 overall score and each category.
- LMArena (arena.ai): people compare two anonymous answers and vote; this becomes an Elo rating. We use the text, WebDev and vision arenas. When a model has several modes ("High", "Max"), the best Elo is kept.
- OpenRouter: price per million tokens, context window, release date and whether the weights are open.
- Artificial Analysis: Intelligence Index, Humanity's Last Exam, Terminal-Bench, GPQA and long context, plus price and speed. In use.
- Frankfurter: European Central Bank reference rates, to show prices in other currencies.
RankingIA score
- The overall ranking includes models with at least one objective benchmark (LiveBench or AA Intelligence Index or Humanity's Last Exam). The rest appear under "other models", without a score.
- For each metric we compute the percentile: the share of ranked models this model beats (a tie counts as half). This puts Elo ratings and accuracy on the same 0–100 scale.
- The score is the weighted average of the percentiles the model has. A missing metric counts neither for nor against it.
- So that a model with little evidence cannot jump to the top on a single metric, the average gets 1 extra weight at percentile 50.
| Metric | Weight | Source |
|---|---|---|
| LiveBench | 3 | LiveBench |
| AA Intelligence Index | 3 | Artificial Analysis |
| Arena text (Elo) | 2 | LMArena |
| Arena WebDev (Elo) | 1 | LMArena |
| Arena vision (Elo) | 0.5 | LMArena |
| Humanity's Last Exam | 1 | Artificial Analysis |
| Terminal-Bench | 1 | Artificial Analysis |
Categories
Each category has its own score: the average percentile of its metrics, with the same shrinkage. Categories and their metrics:
| Category | Metrics |
|---|---|
| Agentic | LiveBench long agentic, Terminal-Bench |
| Coding | LiveBench coding, Arena WebDev (Elo) |
| Reasoning | LiveBench reasoning, Humanity's Last Exam, GPQA Diamond |
| Data analysis | LiveBench data analysis |
| Research | LiveBench research, Long-context reasoning |
| Design | LiveBench design, Arena WebDev (Elo) |
| Browser use | LiveBench browser |
| Chat | Arena text (Elo) |
| Vision | Arena vision (Elo) |
History of each model
Models of the same family (for example Claude Opus 3, 4, 4.5 and 5.5) are grouped by name without the version number. Each model page shows the whole family in release order.
Model names
Each source writes names differently ("Claude Opus 5.5 (High)", "anthropic/claude-opus-5.5"). We match them with a key without company, date and effort mode; cases the rule misses go into a hand-reviewed alias list. If a model looks duplicated or mixed up, open an issue.
Limitations
- Percentiles depend on the group: when a strong model arrives, others can drop without getting worse.
- Elo measures preference, not accuracy. Style and answer length have an effect.
- Prices come from OpenRouter (the lowest across providers); the vendor's own price may differ.
Other references
These sites also track AI models but are not part of the score:
- BenchLM: aggregates hundreds of benchmarks. Its terms require a separate agreement to display its data on a public site.
- Stanford HAI — AI Index: annual report on the state of AI; not a per-model leaderboard updated daily.
- SWEN.AI: Portuguese-language ranking built only on Artificial Analysis data, which RankingIA already uses directly.
Source status
| Source | Last collection | Data from |
|---|---|---|
| OpenRouter | ok | 10/30/2026, 11:45 AM |
| LiveBench | ok | 10/30/2026, 11:45 AM |
| LMArena | ok | 10/30/2026, 11:45 AM |
| Artificial Analysis | ok | 10/30/2026, 11:45 AM |
| Câmbio do BCE (Frankfurter) | ok | 10/30/2026, 11:54 AM |
Each source's data belongs to its owners and follows their terms. RankingIA's code is open source under the MIT license at github.com/RicardoBiazin/RankingIA.