How We Score Models
Every number on this site should be traceable, understandable and falsifiable. This page explains exactly how our scores are built — and where they fall short.
The seven dimensions
Each model receives a 0–100 score per dimension. The overall score is a weighted average of these dimensions.
| Dimension | Weight | What it measures |
|---|---|---|
| Intelligence | 30% | General capability and knowledge |
| Coding | 20% | Software engineering and tool use |
| Reasoning | 15% | Math and logical reasoning |
| Chinese | 10% | Chinese-language ability |
| Long context | 8% | Long-document handling |
| Multimodal | 7% | Image / audio / video |
| Value | 10% | Capability per dollar |
Where the data comes from
We aggregate from three types of sources, in descending order of trust:
- Independent leaderboards — e.g. Artificial Analysis Intelligence Index, LMArena, SWE-bench, GPQA, LiveBench, OpenCompass, SuperCLUE. These use standardized harnesses and are our primary input.
- Vendor model cards & pricing pages — for context windows, output limits, prices, modalities and release dates. We note when a figure is vendor-reported rather than independently verified.
- Reputable trackers & roundups — used only to cross-check and fill gaps, and always linked on the relevant model page.
Normalization
Different benchmarks use different scales. We map each underlying signal onto a 0–100 dimension score, then average signals within a dimension. Where evidence is thin or conflicting, we lower our confidence and keep the estimate conservative. Speed is reported in raw tokens/second and only converted to a 0–100 bar for visual comparison.
Known limitations
- Benchmarks rotate. They saturate, leak into training data, and get replaced by harder versions. Scores from different index versions are not comparable.
- Harnesses differ. A vendor's own measurement is usually a ceiling; a standardized leaderboard is closer to a floor. We try to label which is which.
- Prices and availability change frequently. Promotional rates expire and models move between preview and GA.
- Scores are not your workload. The only evaluation that fully matters is the one you run on your own tasks and data.
Keeping data fresh
Every model record carries an asOf date showing when we last reviewed it, and the site footer shows the overall data date (2026-11-20). We aim to update monthly and after major launches. If you spot an error, please open an issue or contact us — corrections are welcome.
Citation & licence
You may cite this site with attribution. Underlying benchmark and price data belong to their respective sources; follow each source's terms.