Specific releases Real-SWE benchmark for private enterprise codebases

Specific released Real-SWE, a benchmark for frontier AI coding agents using tasks from licensed private production codebases. The company says the tasks cover workflows such as billing, tax calculation and customer migrations, and are evaluated with native harnesses and grading-time verifiers. In a small sample spanning three tasks and six agent runs, the source says models often missed requirements and that independent mesh parsing exposed defects tool self-reports missed.