Claude Opus 5 completed one of my coding tasks considerably faster than Opus 4.8.

There was just one problem: it confidently reported that the issue was fixed when it wasn’t.

That experience captures the trade-off with Anthropic’s latest Opus model. It is faster and more capable on difficult, multi-step work, but polished output can make its mistakes harder to notice.

After testing it on coding and agent tasks, I changed three parts of my workflow.

1. Start with medium reasoning effort

More reasoning is not automatically better. For routine coding tasks, begin with medium effort and increase it only when the problem genuinely requires deeper investigation.

Higher effort can consume more tokens, expand the scope of the task, and produce a solution far more elaborate than the one you requested.

2. Verify outcomes, not explanations

A convincing explanation is not evidence that the task was completed correctly.

Ask for or independently run the relevant tests. Review the files that changed. Confirm the original bug no longer exists.

The dangerous failure mode is not nonsense. It is an incorrect result presented like finished work.

3. Control the scope

Define what the model may change before it begins:

  • Which files can be modified?
  • What behaviour must remain unchanged?
  • Which tests must pass?
  • Can it create subagents or expand the task?

Opus 5 is strongest when the job requires investigation across multiple steps. For a small, clearly defined change, that same initiative can become unnecessary complexity.

The quick version

I condensed my findings, the confidently wrong problem, and the three changes I recommend into this 90-second video:

My broader verdict is simple: Opus 5 is a meaningful upgrade for difficult coding and agent work, but only when verification is part of the workflow.

I published the complete review, including pricing, benchmark comparisons, use cases, and switching advice, on Hashnode:

Have you tested Opus 5? Did it improve your workflow, or merely become more articulate while being wrong?