Benchmark scores, prices and release dates for today’s language models, fetched again every night. It answers one question: which model should get the work.
Open weights means OpenRouter lists a Hugging Face id for the model, closed means it lists the model without one. A model it does not list takes Epoch AI's answer where Epoch rates it, and is unknown otherwise. Cards are picked from one row per model, batch tiers and announced rows left out, and the filters do not move them.
≈ and italics mark an estimated score. The model has no result on one or more of that score's benchmarks, and each gap is filled in from the eight fully measured models that score closest to it. Hover or tap a score for what was measured and what was filled in.
Every score runs from 0 to 100. A benchmark scored in percent counts as its percent. An arena rating counts as the share of votes the model would be expected to win against that arena's median model. A category is the mean of its benchmarks, and Overall is the mean of the categories. A score needs at least half of its parts. A missing part is filled in from the eight fully measured models that score closest to this one on the parts it has: the model's mean over those parts, moved by how far the eight score on the missing part above or below theirs, and never past the best or worst score measured on that part. A score with a part filled in is an estimate: it carries ≈, it is in italics, and its tooltip says which part was filled in and with what.
Click a tile to pin that model.