Cursorbench only runs benchmarks for a handful of (mostly closed) models. If it is true that it is really the best benchmark for agentic coding, it should be evaluated for many more models. I would ideally like to see scores for ~50 or so models, including many open weight ones, and including models of the same family but different parameter size or quantization.
As a person who is evaluating self hosting, I am very interested to see how many of the open weight, medium sized or dense models perform.
One thing to add, Cursor · CursorBench isn’t all closed models. It currently includes GLM 5.2, Kimi K3, and Kimi K2.7 Code, and most models are broken out by effort level rather than reduced to a single number. The page also reports cost, tokens, and steps per task next to each score.
On expanding coverage: the models listed are the ones we serve in Cursor, and each entry is a full agentic run across the whole task set, scored by agentic graders, then repeated at each effort level. That’s why the list grows deliberately instead of covering the field. There is a lot that goes into each model added to Cursor bench so it isn’t trivial to add more, and we lean towards gradually expanding the list when we can. There’s more on how the suite is built and graded here: How we compare model quality in Cursor · Cursor