Leaderboard

How the leaderboard is made

Five boards, each with its source, licence and data date printed above it. Three are LMArena’s. Two only WayJet can make, because they use what WayJet charges and how fast models answer on the WayJet gateway.

Overall, coding and vision: LMArena

These three boards are LMArena’s, read every day from its public dataset on Hugging Face and shown under its licence, CC BY 4.0. People put a prompt to two anonymous models and vote for the better answer; the votes become a score.

  • Overall is the text arena with style control, which is what LMArena shows by default: it discounts answers that win on length or formatting alone.
  • Coding is the WebDev arena: models build a working web app from a prompt, and people vote on the result.
  • Vision is the vision arena with style control: prompts that come with an image.

Each score has a 95% confidence interval. When two models’ intervals overlap, LMArena gives them the same rank and the board marks them tied (=) rather than ordering them. The change column compares the rank with LMArena’s previous published day.

Models WayJet does not sell stay on these boards. A board that listed only what we sell would not be a ranking. Models you can call on WayJet have a “Try it” link.

Value: score for the price

Only models you can call on WayJet, ranked by:

value = overall score ÷ (0.75 × input price + 0.25 × output price)

Prices are WayJet’s, in US dollars per million tokens, on the day the page was built. The 3:1 weighting is roughly the mix of a chat workload, where prompts are longer than answers. When LMArena lists a model in several settings (High, Low, Thinking), its best overall score is used. A model with no token price (an image or video model, for example) is not on this board.

Speed: measured on WayJet

Only models you can call on WayJet, measured on real requests through the WayJet gateway over the last seven days:

  • Output speed is output tokens divided by the whole request time, so a slow first token counts against it. The board shows the median.
  • Time to first token is how long a streamed request waited for the first piece of text, reasoning or tool call. The board shows the median.
  • Request time is the median time from request to the end of the answer.

Only successful chat requests with at least 100 output tokens count, so the number is the model’s and not the network’s, and a model needs at least 20 such requests in the window to be listed. The board shows medians only: no request counts, and nothing about any customer.

Matching names

LMArena names models its own way (“Claude Opus 5.5 (High)”). Each name is matched to a WayJet model automatically when exactly one fits, and an editor checks the rest. A model LMArena lists that WayJet does not sell keeps its place on the board without a link.

Updates

LMArena’s boards and the speed figures are refreshed every day at 06:00 (UTC+8). Each board prints the date its numbers are from. If a day’s data cannot be read, the previous day’s stays in place.

What is not on the boards

Benchmark scores that vendors publish about their own models are cited in the model profiles, not ranked here: each vendor tests under its own conditions, so the numbers cannot be compared. Other public benchmarks join once their terms allow it.