What Anthropic Actually Shipped
Anthropic released Claude Opus 5.5 on September 22, cutting input and output token prices by 20 percent and raising the cache-read discount to 60 percent off Opus 5's rate. Anthropic says the changes add up to a 40 percent cost reduction on typical workloads and roughly 30 percent faster output generation.
| Per 1M tokens | Opus 5 | Opus 5.5 |
|---|---|---|
| Input | $5.00 | $4.00 |
| Output | $25.00 | $20.00 |
| Cache read | $0.50 | $0.20 |
| Cache write | $6.25 | $5.00 |
On Anthropic's own benchmark suite, Opus 5.5 scores 66.4 percent on Terminal-Bench 4.0 against Opus 5's 52.3 percent and OpenAI's GPT-6 Astra at 57.9 percent, and 57.8 percent on CursorBench 4.0 against GPT-5.6 Sol's 41.7 percent. It is available now through the Claude Platform API as claude-opus-5-5, through Amazon Web Services, Google Cloud and Microsoft Azure, and in the Claude.ai web and app clients.
The Reroute Nobody Else Put In A Headline
Anthropic's own release notes say most cybersecurity-related requests sent to Opus 5.5 are automatically re-routed to the older Opus 4.8 model instead, unless the requester is verified through Anthropic's expanded Cyber Verification Program. A near-identical gate applies to biology-adjacent research queries, which need approval through the Life Sciences Verification Program before Opus 5.5 will answer them directly, the same setup Anthropic uses for Fable 5.1.
None of the coverage that repeated Anthropic's price and benchmark numbers this week mentioned the reroute. For a security team that budgeted for Opus 5.5 expecting one model's behavior throughout a workflow, the practical result is that some requests silently come back from Opus 4.8 instead, with no visible warning in the response, unless the account has already gone through verification. Anthropic also added a preserved-thinking feature that blocks API users from editing context to extract the model's reasoning, aimed at stopping rivals from distilling Opus 5.5 into their own models.
What Changes For EU and UK Teams Specifically
Anthropic says Opus 5.5 output now carries watermarking to meet European Union AI Act transparency requirements, on top of the zero-data-retention option it already offered regulated customers. At today's exchange rate, the new $20-per-million-token output price works out to roughly EUR 17.50 or GBP 15, cheap enough that AI Act-compliant, zero-retention usage is no longer priced only for the largest EU customers.
That combination matters for any EU or UK business that has to show which parts of a document were AI-assisted, whether it is a law firm drafting with Claude or a bank running compliance checks. The watermarking is a technical trace built into the model's output at the provider level, not a setting a customer has to remember to switch on, which is worth knowing before an audit rather than during one.
What This Means If You're Already Paying For Opus 5
A team currently spending on Opus 5 API calls gets the same task done for roughly 40 percent less by switching to Opus 5.5, by Anthropic's own benchmark math, and migration is a one-line change to the model ID, claude-opus-5-5 instead of claude-opus-5.
It is not a silent drop-in, though. An early tester quoted by Anthropic said the new model writes the way they do, meaning its communication style shifted along with its benchmark scores. Anyone with an existing prompt library tuned to Opus 5's phrasing should re-test before flipping every workload over, rather than assuming identical output with a lower invoice.
Read next: Your AI Safety Promise Can Now Cost You the Contract | A Safety Essay Became Evidence in an Antitrust Suit
