The numbers presented at the period's coding model launch came from benchmarks based on real repository problems, a format that consolidated as the category's reference.
The difference from synthetic exercises is large. Solving an isolated problem requires generating correct code; solving a repository problem requires understanding an existing base, finding where to intervene, respecting project conventions and passing tests that already existed.
That second skill is what decides practical usefulness, because almost all professional programming happens in code someone else wrote, with history and constraints nobody documented.
The usual reservation applies to numbers published by whoever launches: the choice of which benchmark to show belongs to the party with an interest in the result, and the configuration used is rarely detailed.
The test that informs a purchasing decision remains internal: run the task the operation actually performs, on its own codebase, and count total spend until the result is finished.