Language model rankings
Quality, speed, price and usage — kept as four separate axes. Usage is popularity, not quality: a cheap fast model wired into one busy application outranks a more capable one used sparingly.
Last synced 2026-10-13 · newest measurement 2026-10-12 · 3 sources(1 stale)
Where these numbers come from
Every figure on this board is a measurement published by one of the sites below and ingested with its source, its measurement date and — where the site publishes one — a link to its own page for that value. Marsel measures nothing itself. The one number Marsel computes is the Marsel score, and its formula is published.
Newest measurement on the board: 2026-10-12
benchmark and evaluation results
Feeds: Coding · Reasoning
- Refreshes:
- daily
- Last fetch:
- 2026-10-13 (19 hours ago)
Licence: Attribution required · terms
Free to use, distribute and reproduce provided the source and authors are credited, under the Creative Commons Attribution licence. Some included external datasets retain their own licences (Apache-2.0 for Aider Polyglot and Terminal-Bench leaderboard data) — check per dataset.
Epoch AI, "AI Benchmarking Hub". Published online at epoch.ai.
- LMArenastale
human-preference Elo from blind head-to-head votes
Feeds: Human preference (Elo)
- Refreshes:
- daily
- Last fetch:
- 2026-10-10 (4 days ago)
Licence: Restricted — licensed use · terms
The site's Terms of Use prohibit reproducing, mirroring, distributing or commercially exploiting the Output. Their published leaderboard dataset release (lmarena-ai/leaderboard-dataset on Hugging Face — latest + full historical splits; Arena-Rank code is Apache-2.0) is the channel they sanction. da.marsel.app is responsible for collecting through a permitted channel and holding any licence required.
usage/token volume routed through providers
Feeds: Usage (tokens)
- Refreshes:
- daily
- Last fetch:
- 2026-10-13 (18 hours ago)
Licence: Attribution required · terms
Rankings data is licensed CC BY 4.0 and may be reused and republished with attribution. Note the site's own caveat: token counts come from each upstream provider's tokenizer, so they are not strictly comparable between providers.
Source: OpenRouter (openrouter.ai/apps)
Sources · Methodology · Marsel measures nothing itself — every figure is a published measurement, ingested with its source and date.