We ran Claude Opus 5 through 260 debugging attempts at Low, Medium, High, and XHigh effort, then compared its reliability, cost, latency, token use, and patch behavior with nine other models.

Anthropic introduced Claude Opus 5 as a thoughtful, proactive model built for complex coding, agentic work, and long-running tasks. It gives developers five effort levels that control how much work the model puts into a response.

That creates a practical question for developers: How much reasoning should they actually buy?

A model that spends more time investigating a bug might find a subtle cause that a faster attempt misses. But deeper reasoning also means more tokens, tool calls, latency, and API cost. If the extra work does not produce more verified patches, the higher setting is simply consuming additional resources.

To measure that tradeoff, we tested Opus 5 at Low, Medium, High, and XHigh effort. Each configuration attempted the same 13 debugging tasks five times, producing 260 scored attempts.

More effort improved reliability, but the gains were small compared with the increase in cost, token use, and latency.

Meanwhile, cost per successful fix increased from $0.083 at Low to $0.477 at XHigh.

The strongest overall balance appeared at Medium. Eleven of the 13 tasks were already solved perfectly at Low.

For this benchmark, selectively increasing effort worked better than using High or XHigh for every bug.

The comparison with other frontier models revealed another important result. Opus 5 at XHigh fixed 64 of 65 attempts, while three other models fixed all 65 at lower cost and latency.

What Anthropic claims about Opus 5

Claude Opus 5 supports a one-million-token context window, up to 128,000 output tokens, and adaptive thinking. It is priced at $5 per million input tokens and $25 per million output tokens, the same standard rates as Opus 4.8.

Anthropic makes High the default effort level. According to its documentation, explicitly setting High and omitting the effort parameter produce the same behavior. Developers can lower the setting to save tokens or raise it to XHigh and Max for more demanding work.

The company positions Opus 5 as a major improvement for software engineering and long-horizon work. Anthropic reports that it more than doubles Opus 4.8’s performance on Frontier-Bench at a lower cost per task. At Max effort on CursorBench, it reportedly comes within 0.5% of Fable 5’s peak score at around half the cost per task.

Independent testing supports the capability story, but it also reveals the cost of additional deliberation.

Artificial Analysis ranked Opus 5 first on its AA-Briefcase agentic knowledge-work benchmark. Its High, XHigh, and Max configurations took the top three positions. However, all three averaged more than 25 minutes per task. Max averaged 36.2 minutes and 103 turns, compared with 24.1 minutes and 55 turns for Opus 4.8 at Max.

CodeRabbit found a similar tradeoff in code review. Opus 5 at XHigh produced a more precise set of actionable comments than its production baseline, but caught fewer known issues and generated roughly four times as many nitpicks. It also used about 50% more input tokens and 65% more output tokens than the GPT-5.6 models in the same evaluation.

Those results suggest that higher effort does not improve every part of the model’s work equally. It can change how long Opus investigates, how often it checks its work, and how broadly it interprets the task. Whether that helps depends on the bug.

Our benchmark measures how Opus 5 handles realistic bug reports in an unfamiliar application and whether higher effort leads to more fixes that pass every test.