OpenCode Claude Opus 4.5 — CVE-Agent-Bench profile

Claude Opus 4.5, deployed via the OpenCode wrapper, achieves 36.8% pass rate across 136 CVE patch evaluations. The agent produced 46 successful patches, 29 failures, 50 build errors, and 11 infrastructure timeouts. At $40.13 per successful patch, it is among the most expensive agents in the benchmark.

The 50 build failures represent a 36.8% build-failure rate. Highest in the benchmark. The cost per patch is 15.2 times higher than the same model deployed through Claude CLI.

Behavioral profile

This profile indicates a speed-runner agent that reaches conclusions quickly but with minimal exploration. Speed 100 means fewest conversation turns. Breadth 0 means minimal tool invocation. The agent attempts to generate patches in the shortest path, which often leads to build failures.

Despite the low accuracy score, precision is maxed at 100. When patches succeed, they contain actual substantive code changes, not padding or hallucinations.

Wrapper penalty analysis

The same Opus 4.5 model achieves 45.7% pass rate at $2.64 per patch when deployed through Claude CLI. Through the OpenCode wrapper, it achieves only 36.8% at $40.13 per patch.

This is a 15.2x cost multiplier for 8.9 percentage point accuracy loss. The OpenCode wrapper does not replicate the Claude CLI's tool calling optimization, environment setup, or error recovery.

The wrapper penalty is the dominant factor in this agent's poor performance. The model itself is competent. The deployment infrastructure is the bottleneck.

Build failure root causes

50 out of 136 evaluations end in build failures. This is the highest rate in the benchmark. Possible causes include:

• Repository environment setup mismatches (missing dependencies, version conflicts)
• Incomplete tool support in the OpenCode wrapper (fewer system commands available)
• Aggressive time limits that cut patch generation short
• Weaker error handling for compilation errors

The OpenCode wrapper's speed-first design may be terminating patch attempts before the agent completes dependency resolution or environment configuration.

Accuracy dimension at zero

The accuracy 0 score is the lowest in the benchmark. This reflects the combination of 29 actual failures (incorrect patches) plus 50 build failures that block validation. When the agent attempts to work quickly, it misses context that would prevent wrong patches.

Precision preserved

Despite accuracy 0, precision remains at 100. Every patch that compiles contains real code changes. The agent does not produce padding, explanations, or empty diffs. This precision is valuable in mixed portfolios. If you filter by compilation success, the remaining patches are substantive.

Multi-model advantage

The primary reason to use OpenCode despite benchmark disadvantage is multi-model support. OpenCode is a single CLI that works with multiple models (Claude, GPT, Gemini, Llama, etc.). If your workflow requires flexibility across model providers, OpenCode offers that integration.

For single-model use cases focused on Claude, the CLI delivers better performance at lower cost.

When OpenCode Opus 4.5 is appropriate

Use this agent only if:

1. Multi-model flexibility is non-negotiable, and
2. Build failures are acceptable or handleable in your environment, and
3. You are not cost-sensitive

For most single-model workloads, Claude CLI Opus 4.5 is superior.

Upgrade path

If you are using OpenCode Opus 4.5 and want to improve, switch to Claude CLI Opus 4.5 (8.9pp improvement, 15x cost reduction) or switch to OpenCode Opus 4.6 (10.7pp improvement, same cost).

Learn more

Compare this agent against all evaluated agents in the benchmark. Read the Anthropic lab profile to understand Anthropic's model positioning. Explore wrapper impacts on performance across all three Claude deployment methods.