The levels of reasoning of Claude Opus 5.5: what they are and when to raise them

Claude Opus 5.5 has five levels of reasoning — low, medium, high, xhigh and max — and a rule that governs them all: the model always thinks before answering and decides for himself how much, inside the level you have set. Anthropic calls it adaptive reasoning, and changes two habits: you no longer need to ask to reason in the prompt, and the level must be chosen according to the task, not held to the maximum for prudence.

The choice has a cost in two currencies — token and wait. and a benefit that is not linear: on many tasks the low level is already sufficient, and on some it is even better than an earlier model at the high level. Data verified on 23 September 2026.

What does "adaptive reasoning" mean?

In the previous models the reasoning was largely something you asked for, in the prompt, with formulas like "think step by step". With Opus 5.5 the reasoning is always present and dosed by the model: in front of a simple question think little, in front of a problem with many conditions think more, and the level you set is the roof in which it can move.

The practical consequence is double. First: the lines on the reasoning must be removed from the prompts, because they do not raise the level and lengthen the wait. Second, if you want a dry answer, the right formula is "respond directly," which reduces the reasoning for that answer. The rest of the prompt review is in prompt for Claude Opus 5.5.

The five levels, basically

Anthropic does not publish a formal description of each level, so the following is a user guide, not a specification. Always check the official platform documentation for exact parameters.

The point that escapes most often: the numbers published in the tables are at the highest level. If you use Opus 5.5 to medium and compare the results with the table, you are comparing two different things.

What it costs to raise the level

Two voices, and they must be measured separately.

The reasoning tokens. The thought of the model is paid as output —at Opus 5.5, $20 per million tokens.and is not seen in the answer. At the high level, on a complex task, they can overcome the tokens of the answer itself. As we read the account is in What Are Tokens.

Waiting before the first word. The reasoning happens all before the answer appears: the longer it is, the more the user looks at a still screen. For an interface used by people it is the number that matters the most, as explained in time to first token. For a night lot it doesn't matter.

The data that overturns intuition

For code revision Anthropic reports that a first user has seen Opus 5.5 at the lowest level find more bugs than Opus 5 at the highest level, with fewer false alarms. It is a reported figure, not a benchmark, but says one important thing: the low level of Opus 5.5 is not a reduced version of the model è is the same model with less reflection time, and on attention tasks (read each line well) can suffice.

The General Lesson: try the low level first, measures, and raises only where the results ask for it. The method for review is in Code revision with Opus 5.5.

How to choose the level per task

1. Take fifty real cases of the task, with the criterion of correctness written before.
2. Run low. Count the correct ones. If the percentage is acceptable, you are done: that task turns to low.
3. Climb one level only if the correct ones are not enough, and run again on the same sample.
4. Stop at the first level that crosses the threshold. The next level costs more and you don't need it.
5. Mark the chosen layer next to the prompt, so when you change the model you will know where to start again.

Where you set the layer

The layer is not written in the prompt: it is set in the tool you use.

Always check the platform documentation before putting a level into production: parameter names change, and a level mis-set on an automatic system pays off at every call for months without anyone noticing. The way to notice it is a row in the call log with the level used and the reasoning tokens consumed: if that column grows without the results improving, the level is too high.

Level and Speed: Two Different Things

The level of reasoning acts before of the first word: the higher it is, the more the model thinks before starting to write. The fast mode of Opus 5.5 acts after: brings out the text faster, at the same level of reasoning, at a double cost. They are two independent knobs, and they combine: a low level in fast mode is the most responsive configuration possible, a maximum level in standard mode the most reflective. When it is convenient to pay for the speed is in Fast mode of Opus 5.5.

Practical examples

Logistics company, 30 employees, Piacenza. Sorting incoming mail into six categories. A low: 96% correct classifications on a hundred mails. A medium: 97%. The difference was not worth the cost and the wait: the system runs at low, and the saving on the month is the biggest entry in the project.

Engineering study, 5 people, Trento. Checking specifications with dozens of cross conditions. At low and medium the model lost the constraints between distant sections; at high it held them. There the high level is justified, and as the task is rare —some specifications per month —the cost is negligible. Same model, two opposite settings, both right.

Mistakes to Avoid

• Asking to "think more" in the prompt. It doesn't raise the level, it confuses the pattern.
• Keeping max anywhere. Double cost, long wait, often identical results.
• Comparing your results to medium with benchmarks, which are at max.
• Choosing the level to feel. It's measured on a sample, in one afternoon.
• Using the same layer for the interface and for the night lot. They have opposite constraints.

How to apply it in the company

For those with systems in production, the level of reasoning is a cost item to govern as the model itself:

1. List Automatic Tasks that turn on Opus 5.5 and the current level of each.
2. Measure to low all those of extraction, classification and revision: most are likely to hold.
3. Leave high only where you have a measured pattern, and write it next to the prompt.
4. Resume when changing model: levels are not equivalent between versions.

For the overall picture of the model — prices, benchmarks, where to use — Claude Opus 5.5: what is it.

Conclusion

The reasoning levels of Opus 5.5 are a knob, not a setting to be put to the maximum and forgotten. The model thinks for itself; you decide how much time you leave him, and you pay for that time in tokens and waiting. The rule is simple: start low, measure, and increase only where needed. That is where much of the −40% cost reduction reported by Anthropic is found.

AI becomes useful when it enters processes. Giallo Studio helps SMEs and teams to build concrete, measurable and sustainable automations.