Open weights against closed

The best closed model scores 35.8 points above the best open-weight one. That is more than the two half-widths together (13.81), so the leaderboard separates them.

Half-widths are LMArena's published intervals. Two scores are separable when they differ by more than the two half-widths together, the same rule as the leaderboard's significance rank.

Of the 12 models on the leaderboard's value frontier, 2 are open-weight, 2 closed, and 8 have no recorded weights status. The frontier is the whole board's, at the balanced workload, not one drawn inside each group.

The weights status is not read from the price feed. Open is set model by model, by hand. Closed is set the same way for 33 of the 58 closed models, and for 25 it follows from the maker’s own pages, for the lines of models it publishes no weights for: Anthropic’s models, OpenAI’s except its open gpt-oss family, and Google’s Gemini line but not Gemma. The pages are quoted in data/maker-weights.json. Anthropic says so in a statement; OpenAI’s and Google’s are terms that forbid extracting the model, weaker evidence than a statement. A model whose record does not set it has no status, which is why the third group is large. The price feed names a Hugging Face repository for 61 of those 76. A named repository is evidence, not a reviewed status, so they stay unknown here rather than being guessed into either column. The open-weights list has every flagged model, rated or not.