Claude Opus 5.5 costs 20% less: when it's really worth it and why the max level inflates your bill

Claude Opus 5.5 spent almost 20 minutes reasoning about a drawing of a pelican riding a bicycle, twice in a row, and in the end it delivered nothing ¹. Each attempt cost 2.56 dollars ¹. Keep that in mind, because it is the part of the launch that gets talked about least.

The main news is something else. From 22 September 2026, Opus 5.5 costs 4 dollars per million input tokens and 20 per million output tokens, against the 5 and 25 that Anthropic charged for every Opus from version 4.5 to 5 ¹. That is a 20% cut ¹.

The short answer to your question: if you use Opus 5 today, switch to the new one without a second thought. If you use Sonnet or a cheaper model, stay where you are until you find a task that Sonnet gets wrong. And whatever you do, never turn on the max level.

The price that matters most is not the one on the front page

For Opus 5.5, reading from the cache, meaning the text the model has already seen and that you pass back to it on every turn, costs 0.20 dollars per million tokens, a drop of 60% ¹.

It sounds like a detail. It is not. In the long conversations of agents, where the model works over many steps on files and documents, more than 90% of input tokens are paid at the cache price ¹. If you use Claude this way, the cut on the cache affects you more than the cut on the list price.

Above Opus there is Claude Fable 5.1, at 10 dollars for input and 50 for output ¹. Below it there is Sonnet 5.5, released on 28 September 2026 at 2 dollars for input and 10 for output, with cache reads at 0.20 dollars ².

Look at that last figure: on the cache, Opus 5.5 and Sonnet 5.5 cost the same ¹². The real difference comes from new input tokens and, above all, from the tokens the model writes.

Sonnet is enough more often than you think

Anthropic says that Sonnet 5.5 is more than 30% faster and costs up to 30% less for most work ³. These are the vendor's figures, so take them as such. The price list is the same as Sonnet 5, but according to Simon Willison the new model beats it on every benchmark ³.

The figure that matters to you is a different one: on some programming tasks Sonnet 5.5 seems almost as good as Opus 5.5 ³. "Some" and "almost" are the words to read carefully. There is work where Opus stays ahead, but it is not the rule.

Anthropic has also chosen Sonnet 5.5 as the model for the free plan on claude.ai ³. A vendor does not hand out to everyone a model that stumbles on everyday work.

For office tasks, such as summarising emails, writing drafts or extracting data from a PDF, Opus is overkill. You pay for a long-reasoning model for tasks that do not need it.

When Opus is worth the difference

Opus makes sense when a mistake costs more than the model. A contract to read line by line, an analysis you will base a decision on, code that goes into production. There, the difference in the price list is small next to an error that slips past you.

According to Anthropic, Opus 5.5 has the intelligence of Fable 5.1 at a lower price per token than Opus 5.0, and it is very sparing with tokens ¹. If that is true, anyone who uses Fable today for hard jobs has a concrete reason to try stepping down a level.

It is not just marketing. Willison, one of the most widely followed observers in the field, has chosen Opus as his default model for his programming work ¹. That is a signal, not proof: he writes code, while you may be managing quotes.

The max level is a trap, and that is not just a figure of speech

Claude models let you choose how long they should think before answering. It is natural to believe that more is always better. With this generation, it is not.

At the max level, Opus 5.5 kept reasoning until it used up the 128,000-token output limit, without ever writing the answer ¹. Two attempts, same result, 2.56 dollars each ¹. Willison concludes that max is effectively unusable ¹.

Sonnet 5.5 has the same flaw: at maximum it burned through 128,000 tokens, equal to 1.28 dollars, and delivered nothing ³. One level down, at xhigh, the same task cost 5.74 cents and took 41 seconds ³.

Read those figures again. Same model, same request: with max you pay more and end up empty-handed ³.

It is a single case and the task is a game, so do not turn it into a law. But if a model gets stuck on a drawing, you have no reason to think it will behave better on your accounts.

What to do now

If you use Opus 5 via the API, switch models: Opus 5.5 costs less per token and the cache has dropped by 60% ¹.

If you use Sonnet and you are happy with it, do nothing. Moving to Opus is only justified by a specific case where Sonnet gets it wrong and the error costs you.

Check the settings of your tools and remove max everywhere. Stop at xhigh or lower, and only raise it if you see answers that are truly insufficient.

Put a cap on spending. Claude Code now shows how much you have spent in the month in dollars, next to the limit you have set ². It helps you spot unusual consumption before the invoice arrives.

If you run large volumes of simple tasks, look outside Anthropic too: GPT-6 Luna costs 0.10 dollars for input and 0.50 for output, while Haiku 4.5 costs 1 and 5 ¹. Haiku 5.5 came out on 7 October 2026 ¹, but its price is not in the sources for this article: check it on the price list before deciding.

If you do not have time to run the test yourself, AIR can help you set it up on your real documents, which is the only test that counts.

The price per token has come down. The bill, though, is still decided by how long you let the model think.