JobBench: Results as Published

JobBench tests AI agents on professional work that experts say they would want to hand off to an agent, rather than on tasks chosen for their economic value. It has 130 tasks across 35 occupations, and each task is a workspace of mixed and sometimes conflicting reference files that the agent must sort through. Outputs are graded against chains of rubrics that average about 35.6 binary criteria per task.

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.