Claude Opus 4.5 — CVE-Agent-Bench profile

Claude Opus 4.5, deployed via the Claude CLI, achieves a 45.7% pass rate across 136 CVE patch evaluations. The agent produced 58 successful patches, 43 failures, 26 build errors, and 9 infrastructure timeouts. At $2.64 per successful patch, it is one of the most cost-effective agents in the benchmark.

This cost-per-patch metric places Claude Opus 4.5 in the lower quartile of expense. The agent generates functional patches at scale without proportional cost growth.

Behavioral profile

The pattern shows an agent that prioritizes efficiency and precision over raw speed. It takes more conversation turns to reach decisions but produces patches that rarely contain false content. When Opus 4.5 generates a patch, the patch is substantive and intentional. Never empty or padding the output.

Comparison with Opus 4.6

The newer Opus 4.6 model, running through the same Claude CLI, achieves 61.6% pass rate. A 15.9 percentage point improvement. Cost rises modestly from $2.64 to $2.93 per patch. Build failures drop from 26 to 20.

The accuracy dimension jumps from 34 to 96, indicating the model upgrade primarily strengthens analytical ability. Speed remains low, and precision stays maxed out. The 4.6 upgrade preserves cost-efficiency while dramatically improving pass rate.

CLI wrapper effect

The Claude CLI wrapper produces vastly different results than alternative deployment methods for the same Opus 4.5 model. When the same model runs through the OpenCode wrapper, it achieves only 36.8% pass rate at $40.13 per patch. That is 8.9 percentage points lower accuracy at 15.2 times higher cost.

This gap shows how much the wrapper matters. The Claude CLI has optimized tool calling, environment setup, and error handling that the OpenCode wrapper does not replicate. Same model, different infrastructure, dramatically different outcomes.

Build failure analysis

26 build failures from 136 evaluations represents a 19% build-failure rate. These failures occur when the agent generates syntactically correct patches that the repository's build system rejects. Root causes include environment setup issues, missing dependencies, or repository-specific build constraints.

The build failures are not agent hallucinations. They are legitimate environmental mismatches. In an integration scenario where you control the build environment, these failures might be preventable.

Precision dimension at maximum

The 100 precision score means every successful patch output contains actual code changes. The agent does not pad with explanations or generic fixes. This precision is valuable in automated patch selection. You can trust the patch output to contain meaningful changes.

When Opus 4.5 is appropriate

Choose Claude Opus 4.5 via CLI when cost per successful patch is a primary constraint. At $2.64 per patch, it enables high-volume evaluations. The 45.7% pass rate is competitive for many CVE types, particularly memory safety and logic issues where accuracy 34 gives sufficient problem-solving.

If your workload requires higher accuracy or faster turn times, Opus 4.6 offers better performance at modest cost increase.

Learn more

Explore the full benchmark results and leaderboard to see how Opus 4.5 ranks among 15 evaluated agents. Read the Anthropic lab profile for context on all Anthropic models in the benchmark. Compare economics across all agents and cost models.