Claude Opus 4.7 benchmarks: key results and how to read 87.6% on SWE-bench Verified

What 87.6% on SWE-bench Verified means

For teams evaluating models as coding agents, SWE-bench Verified is the cleanest headline figure in the available source set: AWS reports 87.6% for Claude Opus 4.7 . In practical terms, that puts the spotlight on software engineering tasks and code problem-solving, which matches Anthropic’s own positioning of Opus 4.7 as a strong model for complex reasoning and agentic coding .

The important caveat: this is not a general-purpose score for everything the model might do. SWE-bench Verified measures a particular kind of capability. It does not replace benchmarks for terminal work, finance, vision, long-context workflows or research-style multi-step tasks.

So if you are making a technical decision, the safer read is to look at SWE-bench Verified together with SWE-bench Pro and Terminal-Bench 2.0, rather than treating one number as the whole story .

Why different sources may show different numbers

Not every source reports the same SWE-bench Verified score. One secondary source lists 82.4% on SWE-bench Verified, while AWS reports 87.6% for Claude Opus 4.7 . That gap matters. It means the responsible way to cite the result is to include three things: the benchmark name, the score and the source.

It also reinforces a broader benchmarking lesson. AWS says Opus 4.7 may require prompting changes and harness tweaks to get the best performance . In other words, the evaluation setup can affect the observed result, especially for agentic coding systems where prompts, tools and test harnesses are part of the product.

Which benchmark should you care about?

If your main use case is software development, start with SWE-bench Verified — but do not stop there. SWE-bench Pro and Terminal-Bench 2.0 are useful when your system needs to handle more complex software tasks, work through tools or operate in environments that resemble a terminal .

If your use case is finance or research, Anthropic’s internal research-agent data is closer to that style of work. In that internal benchmark, Opus 4.7 posted an overall score of 0.715 and scored 0.813 on General Finance, compared with 0.767 for Opus 4.6 on that module . Those figures are useful signals, but they should still be treated as internal evaluations rather than independent verification.

If you are looking at long-running enterprise workflows, the public information points to improvements in long-duration tasks, instruction following and work under ambiguity, according to AWS citing Anthropic . In that setting, benchmarks are a starting point. The real test is whether the model performs inside your own tools, prompts, data and evaluation harness.

Bottom line

The easiest benchmark to quote for Claude Opus 4.7 is 87.6% on SWE-bench Verified, and it is especially relevant for agentic coding . But the more useful interpretation is broader: the model is also reported at 64.3% on SWE-bench Pro, 69.4% on Terminal-Bench 2.0 and 64.4% on Finance Agent v1.1, while Anthropic highlights internal gains in multi-step research work and finance .

The responsible way to compare Claude Opus 4.7 is not to ask for one benchmark in isolation. Pick the benchmark that most closely resembles your actual workflow, check the source and setup, and then validate the model with your own prompts, tools and harness.