Know what the numbers mean.
We collect official Arena leaderboard history and the Models.dev catalogue daily at approximately 05:17 UTC. The visualizations use published measurements, with missing values left missing. Their public snapshot updates independently of frontend releases.
Human preference is one measure.
Arena ratings summarize human comparisons within a leaderboard. They are not a universal intelligence index, an IQ score or a probability of task success. Compare the same category and methodology. We show confidence intervals and vote counts; overlapping intervals do not establish a significant winner. A changing model pool can change relative ratings.
The formulas
| View | Meaning |
|---|---|
| Arena · preference | Normalize ratings within one board. Bubble radius = 26 + 40 × √normalized rating. A common fit factor reduces radii for dense boards. Larger means more preferred; area is not claimed to be proportional to raw rating. The default focuses on the top 18 and up to four open-weight challengers; Show all includes the whole curated board. |
| Arena · price | Normalize −ln(workload token price). Larger means lower token price. Missing, negative and zero prices are excluded from this transform. |
| Field / value frontier | A model is on the Pareto frontier if no priced model has both an equal-or-higher preference rating and an equal-or-lower price, with one strict advantage. We never divide Arena ratings by dollars. |
| Graveyard | Gap = board leader rating − model rating. Orbit radius = max(76, 236 − 0.55 × gap). The horizon is a visual metaphor, not a claim of obsolescence. |
| Rising | Each lab’s best closed-weight model minus the strongest open-weight model on the same board. Lines are unsmoothed, use publication dates and break across methodology versions. The six featured labs are selected from the board being viewed. |
| Waterline | Δ = model rating − strongest open-weight model rating. Height = clamp(50, 530, 290 − sign(Δ) × √|Δ| × 10). Collision avoidance changes horizontal position only. |
Token price is a workload choice.
For input share w, price per million tokens = w × input price + (1 − w) × output price. Default w = 50%. Prices are USD, uncached standard tiers from the named serving provider. Cached, audio, reasoning-token usage, taxes, bulk tiers and actual task success can change the bill. These are listed token prices, not measured cost to complete a task. Exact model identity is required before joining a rating to a price.
History has two meanings.
Published archive reconstructs publisher-labelled boards using the latest collected revisions. Earlier rows may have been restated by the publisher. This is descriptive history, not backtest-safe history.
As known at the time requires a board to have been collected and available by the selected cutoff. Our collection began on 12 October 2026, so earlier strict history is empty. Prices and identity metadata cannot travel backward from today into an earlier replay. Unknown measurement windows and exact publication timestamps remain unknown. We show the publication date separately from the collection date.
Open weights ≠ verified open source.
Weight availability is taken from a model-specific Models.dev claim or a named weight-licence claim on an Arena board. Unknown and conflicting claims remain unknown. An open-source label requires separate licence verification. We do not classify a model from its country or its lab, and dataset licences do not grant rights to model weights.
Sources and attribution
- Official Arena leaderboard dataset — CC BY 4.0, attribution to LMArena. This site extracts a focused subset and derives visual layouts and same-board differences. Licence · Methodology changes.
- Models.dev by Anomaly — MIT-licensed catalogue. Copyright and permission notice. Prices, specifications and release-date claims can have missing fields.
What is still missing?
Some rated model variants have no exact catalogue match; we show them without prices. We have not connected a redistributable output-speed measurement source or a broader capability benchmark yet. Neither field is filled with estimates. The pipeline checks freshness, provenance, source rights, privacy, identity ambiguity and historical integrity before publishing a new snapshot. If a run fails, the last successful snapshot stays available and its age is visible.
Capability evaluations such as LiveBench will be added as a separate benchmark with its own methodology and release version. We will not silently combine it with human preference.