How this list is made

All data retrieved 2026-09-19. Every page on this site stamps its data date.

The two sources

1. LMArena — human preference

LMArena shows two anonymous models to a human, both answer, and the human picks the better one. Millions of those head-to-head votes become an Elo-style Arena Score. We use these boards:

• Text, overall (style control on) — our top-10 backbone, 402 entries
• Text, coding — the "best code" winner
• Text-to-image — the "best image" winner
• Text-to-video — the "best video" winner

2. Artificial Analysis — measurement

Artificial Analysis benchmarks models on standardized tests and publishes their Intelligence Index, API prices, measured output speed, context window and open-weights status. We use it for reasoning, speed, price and value.

Our rules

• Variant collapsing. Versions that differ only by effort setting ("high", "max", "xhigh") share one entry, using the best variant's score. Separate model versions (5 vs 5.1, 4.6 vs 4.7) rank separately.
• Best image / video / code = #1 on the respective LMArena board on the retrieval date.
• Best reasoning = highest Artificial Analysis Intelligence Index among the models we track.
• Best open = first non-proprietary entry on LMArena's overall board (open weights, license stated).
• Best value = Intelligence Index ÷ blended price (3 parts input : 1 part output), per $1M tokens. Today's winner: GLM 5.3 Flash — 41.9 ÷ $0.2375 ≈ 176 index points per $1M.
• Ties keep the source board's own rank order.
• Editorial text ("best at", "not the pick for") is written by us from these numbers and is clearly separated from the data. It is never the basis of a rank.

Honest limits

• Arena Scores move daily; a board leader can change within weeks. We date everything and refresh weekly.
• Image and video boards are young — thousands of votes, not millions. Expect movement.
• The Intelligence Index covers the models Artificial Analysis tracks; a model we did not record could top it.
• Speed and price vary by provider and effort setting; we use the medians the sources publish.

What will never happen here

• No paid placement. Not now, not ever. Nobody buys a rank.
• No affiliate links. "Try" buttons go straight to the maker's own page.
• No self-insertion. If a model made by us ever gets measured on these public boards, it will appear exactly where the votes put it — and nowhere else.

Reproduce it

Open the four LMArena boards and the Artificial Analysis model page on any day and compare with our data file, which stores every number this site showed and the date it was retrieved. Changes go in the changelog.