Anthropic released Claude Opus 5.5 on 22 September. Since then, the number that has travelled furthest is that it costs 40% less than Opus 5. It's accurate, though it lumps together several things that make more sense looked at one by one.
At Liquid we use these models every day, on client projects and internal tools, so we read the announcement and the technical documentation with our own work in mind.
What it costs
The price per token drops 20%. A million input tokens goes from 5 to 4 dollars, and a million output tokens from 25 to 20. Cache reads drop much further, from 0.50 to 0.20 dollars, a 60% cut, and you feel that in products that reuse the same context on every call, like an assistant that always looks up the same documentation or an agent working on the same repository.
The 40% comes from a different calculation. Anthropic measures it with default settings on typical workloads, and it adds up the lower price, the model needing fewer steps to finish a task, and a default reasoning level that is now medium. If your product looks like that typical workload, you'll land close to it. If it doesn't, you could end up above or below, and the only way to know is to test it with your own cases.
What it does better than Opus 5
According to Anthropic, Opus 5.5 performs at the level of Fable 5.1, its most capable model, on most tasks. It pulls furthest ahead of Opus 5 in agentic coding, in long jobs like code migrations or audits, and in analysis. The benchmarks in the announcement put it ahead on almost everything, though the announcement itself admits that at this level the gaps in the tables say less and less about what happens in real use.
It generates responses more than 30% faster. In an internal tool that may not matter, but in a product where the user is waiting and staring at the screen, those seconds count.
It also writes better, and that's the improvement we found most interesting. It puts the important information first, uses less jargon and follows the style rules you give it more closely. In a support chat or an assistant inside an app, where the model talks straight to your users, people notice that even if they couldn't tell you why.
What to check if you already use Opus 5
If you already have Opus 5 in production, the migration has some small print. Reasoning can no longer be switched off, the way you force the model to use a specific tool has changed, and on accounts created from 31 August onwards you can't edit earlier messages in a conversation and send them again. Some of these changes throw errors if the code isn't ready for them, so check them before touching anything. Opus 5 is still available, so there's no rush to switch.
On security and compliance there are two details we care about in particular, because we work with companies here in Europe. Opus 5.5 can be used with zero data retention, which is the first thing many clients ask when we talk about sending their users' data to a third party. And it watermarks the text it generates to comply with the EU AI Act. It also ships with stricter safeguards in cybersecurity and biology, and some requests in those areas get routed to another model automatically. If your product works in those fields, keep that in mind.
What we'd do
What matters most to us in this release is that the price drops and the writing improves at the same time. Some AI features have been hard to justify for a small business because of what each use costs, and at these prices it's worth running the numbers again. We give less weight to the decimal points in the benchmarks. After a few launches, we trust what we see testing on our own projects more.
To switch models on a project in production, we'd first pin the current settings so the comparison is fair. Then we'd test the new model with real cases from the product, work out what each successfully completed task costs, and move traffic over gradually. If you're still deciding whether to put AI into your product, we recently wrote about the questions worth asking before you start.
If you're thinking it over
New models will keep coming out every few months, and which one to use, and how, depends a lot on each product. At Liquid we offer AI advisory to help make those decisions with data from your own case, whether you already have something running or you're just starting. If you'd like, let's talk.
