SWE-Marathon: Results as Published

SWE-Marathon tests whether AI agents can autonomously complete ultra-long software work. It has 20 long-horizon tasks across software engineering and adjacent technical fields, each with a unique executable environment, a human-written reference solution and a multi-layer verification suite. A task counts as resolved when the verifier accepts the agent's work, and the paper says frontier coding agents resolve fewer than 20 percent of tasks.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.