Claude Opus 5

1. Four Effort Levels

Opus 5 supports four effort settings: low, medium, high, and max.

Higher settings give the model more room to reason internally, but they also use more tokens and take longer.

This is one of the more useful changes in practice. A support workflow or code-review system does not need to run every request at max. Routine jobs can stay on low or medium, while harder cases get more compute.

Artificial Analysis found that even at lower effort settings, Opus 5 still beat the best scores of many competing models.

2. Stronger Agentic Work

Opus 5 performs well on tasks where the model has to gather information, use tools, interact with a computer, and keep working through several steps.

On OSWorld 2.0, it beat Fable 5’s best score while using roughly one-third of the cost. On AutomationBench, its task completion rate at a similar cost was 1.5× higher than the next-best model.

Anthropic also showed a more unusual example.

The model was given a drawing of a mechanical part but was not allowed to inspect the image directly. Instead, it wrote its own computer-vision pipeline, extracted geometry from the raw pixels, and rebuilt the part as a 3D model.

Other models failed across five attempts.

That example is more interesting than simply saying the model “uses tools well.” It built a tool for itself when the obvious path was blocked.

3. Better at Checking Its Own Work — Sometimes Too Much

Anthropic says Opus 5 is better at verifying its answers and tracing bugs back to their root cause instead of making a quick surface-level fix.

That behavior is useful for coding agents.

The System Card also shows the downside. Opus 5 can sometimes get stuck in what amounts to a self-verification loop.

In one biology research task, the model spent eight hours building and running checks, then produced no final output.

That is the kind of failure I would care about in long autonomous runs. More checking is helpful until the checking becomes the task.

4. What Changed From Opus 4.8

Opus 4.8 arrived two months earlier, and the API price is unchanged.

The jump shows up in a few numbers:

• Frontier-Bench: 18.7% → 43.3%
• CursorBench 3.2: within 0.5% of Fable 5’s best result, at roughly half the cost
• Safety-classifier intervention rate: about 85% lower than Fable 5

There is also a practical migration detail.

Opus 5 has thinking enabled by default, and it can only be disabled when effort is set to high or lower.

If your existing code assumes Opus 4.8-style behavior, that is worth checking before switching models.

Opus 5 makes the most sense to me as a model for teams that want high-end agent and coding performance without paying top-end prices on every call.

The effort setting helps here. Routine work can run cheaper, while harder jobs get more reasoning only when needed.

Its weaker point is long autonomous execution.

The eight-hour self-verification case in the System Card is hard to ignore. A model can be strong at individual steps and still get lost when nobody is watching for a long stretch.

I would trust Opus 5 with difficult work. I would also give it more room to use tools and make decisions than I would with a weaker model.

What I would not do yet is start a multi-hour or multi-day task, walk away, and assume the final result will be ready when I come back.

Opus 5 is already close to being a much cheaper flagship. The remaining question is whether it can stay reliable once the job stops being a single hard prompt and turns into hours of uninterrupted work.