How does Claude Sonnet 5's agentic performance on coding, tool use, and knowledge work benchmarks compare to its predece

What the evidence tells us

Your question asks about Claude Sonnet 5’s agentic performance on coding, tool use, and knowledge work benchmarks, compared with Sonnet 4.6 and Opus 4.8. The available evidence gives us a partial picture — enough to see a couple of bright spots, but plenty of gaps too.

Coding benchmarks (SWE‑bench)

The only hard number we have is that Sonnet 5 is reported to hit 82 % on the SWE‑bench coding benchmark [2]. That’s a specific agentic coding score, but the evidence does not tell us how Sonnet 4.6 or Opus 4.8 perform on the same benchmark, so we can’t make a direct numerical comparison.

Tool use benchmarks

None of the supplied sources mention tool‑use benchmarks for any model. This part of the question can’t be answered from the evidence.

Knowledge work benchmarks

Similarly, there is no information about knowledge‑work benchmarks in the sources. We can’t compare Sonnet 5 to its predecessor or Opus on that front.

Overall comparison with Opus 4.8

The evidence does give a general performance comparison: Sonnet 5 is described as achieving Opus‑level performance at roughly 50 % of the cost [1], and one source flatly states that “Sonnet 5 is about as good as Opus 4.8” [3]. While these are broad statements rather than a benchmark‑by‑benchmark breakdown, they suggest Sonnet 5 matches the larger model’s capability in typical agentic tasks, including coding scenarios, while being cheaper to run.

Comparison with Sonnet 4.6

Unfortunately, the evidence contains no mentions of Sonnet 4.6 at all. We can’t say how Sonnet 5 stacks up against its direct predecessor on any of the requested dimensions.

In short: the evidence shows Sonnet 5 rivaling Opus 4.8 on overall agentic performance and hitting 82 % on SWE‑bench, but it is insufficient for comparing tool use, knowledge work, or Sonnet 4.6.