How we score

How a score is built

1. Everything becomes a percentile. For each test, we rank every model that took it and turn its place into a number from 0 to 100: "better than X% of the models that took this test". This puts chess-style vote ratings and percent-correct test scores on the same scale.

2. Categories average their tests. A category score is the plain average of whichever of its tests the model has. We never show a number for a test a model didn't take.

3. The overall score is a weighted average of categories, using the weights above. We only show it when the categories with data carry at least 70% of the total weight. Otherwise you'll see "no score yet". Today 406 models pass. If a model is missing a category, that category counts at the model's own weakest measured category, so a gap in the data can never push a model up the ranking.

4. Your weights, your ranking. On the leaderboard you can change the weights. That re-ranks models in your browser only, and the link keeps your settings.

Matching models across sources

Arena lists each model with its reasoning setting (for example "High" or "Max"), and we treat each one as its own entry. An Epoch AI result is attached only when it is clearly the same model and the same setting: same reasoning level, same thinking budget, and the same dated snapshot. If Epoch tested a different setting or snapshot, we leave it out rather than guess. A list price attaches only to the same snapshot. It applies to every reasoning setting of that model, because providers charge the same per token; higher settings just use more tokens.

Price

Prices are list prices in US dollars per 1 million tokens, from the OpenRouter models API. The single "blended" price assumes 3 parts input to 1 part output, a typical chat mix. Free, batch-only and router listings are ignored. Open models are served by several providers on OpenRouter, so their price is OpenRouter's listed price, not the only one available. If a model isn't listed, we say "no price yet".

Speed

Coming soon. We're building our own hourly speed tests. Until they run, we show no speed numbers at all.

What we never do

• Invent, estimate or fill in missing numbers. Missing means "no data yet".
• Take paid placements. SIscope is free forever.
• Use Artificial Analysis data, directly or relayed through another feed. Terms forbid use in a competing ranking product. Never used, including when relayed by other APIs.
• Scrape websites. Every source is an open dataset or a public API.

Sources and licences

Epoch AI tests we use

• GPQA Diamond: PhD-level science multiple choice questions. 333 models tested.
• SimpleQA Verified: Short factual questions; rewards not making things up. 100 models tested.
• OTIS Mock AIME 2024-25: Competition maths problems in the style of AIME. 313 models tested.
• FrontierMath Tiers 1-3: Very hard, unpublished research-style maths problems. 122 models tested.
• SWE-bench Verified: Fixing real GitHub issues in Python repositories. 35 models tested.

Changelog

• v0.2.0 (Nov 3, 2026): missing categories now count at the model's weakest measured category instead of being skipped, and Epoch results attach only to an exact setting match.
• v0.1.0 (3 Nov 2026): first public version. Six categories, Arena + Epoch AI inputs, 70% coverage rule.

Raw data is free to use: /api/v0/models.json. Please keep the source attributions.