Claude Opus 5, as a governed external route

Claude Opus 5 is now in the model catalogue as anthropic/claude-opus-5. It is Anthropic's newest frontier model, positioned by them for complex agentic coding and enterprise work, with a 1,000,000-token context window, up to 128,000 output tokens, adaptive thinking, vision, tool use, and streaming. Enable it for your organization and call it by id against the same base URL as everything else.

Then the part that decides whether you can use it at all, stated before the specifications: Opus 5 is not EU-sovereign.

It is not an EU-sovereign model

Anthropic is a US provider. Routing a request through Akumi does not change where Anthropic serves it. Opus 5 is an external model, and on this platform external means governed rather than casual.

The EU-sovereign default has not moved. Mistral is what answers when residency is the requirement: mistral/mistral-small-2603 and the rest of that family are served in the EU, and a request to one of them goes straight through with no acceptance to record and no guard to clear. If your rule is that data never leaves European infrastructure, that is still the answer, and Opus 5 is not it.

We are saying that plainly because the alternative is worse. A platform that lists a US frontier model beside an EU one without marking the difference has quietly made the difference the customer's problem.

What it takes to reach it

The egress guard fails closed. An EU-resident target is always permitted. Any other target is refused unless two things hold:

• Your organization has established billing. Before that, no non-EU provider is reachable at all, and no firewall or acknowledgement unlocks it.
• The request is either firewalled by the PII firewall, or it carries a valid compliance acknowledgement.

That acknowledgement is not a request flag you can set yourself. It exists only when a workspace admin has accepted non-EU routing on the record, with their identity, the timestamp, and the version of the warning they read. The acceptance is scoped to one workspace, so a sibling workspace's decision never unlocks egress for yours, and it stops counting the moment the warning copy changes, which forces a fresh acceptance rather than letting an old click cover new terms.

Miss any of it and the call is rejected with a compliance warning that names the reason. It is not silently rerouted to a different model, and it is not sent anyway. There is no default-allow branch in the guard, on purpose: the code that permits a transfer is the code most worth keeping boring.

Firewalled is the ordinary path. The PII firewall pseudonymizes personal data before egress and restores it in the response, so what Anthropic receives is tokens rather than people. Worth being exact about what that buys: pseudonymized data is still personal data under GDPR, so the transfer is still a transfer and still needs the paperwork underneath it. The firewall reduces what is exposed. It does not make the transfer disappear, and we will not describe it as if it does.

A million tokens of context

Opus 5 accepts a 1,000,000-token context window and returns up to 128,000 tokens, with a knowledge cutoff of May 2026. That is the size at which long agentic work stops fighting the window: a service and its full test suite in one prompt, a document set instead of an excerpt, a tool-calling session that runs to the end without being compacted halfway through.

It is also, in governance terms, a great deal of data in a single transfer. A larger window does not change the rules, it raises the cost of getting them wrong. The request that used to carry a snippet now carries a repository. That is precisely the case routing policy is for: keep your crown-jewel data classes pinned to EU-only targets and let general work reach Opus 5, rather than making the choice once for everything.

Metered like everything else

Opus 5 is priced at 7 credits per 1,000 input tokens and 33 per 1,000 output tokens. That matches Claude Opus 4.8, whose upstream price it shares, and it sits on its own configuration keys so the two can move apart when Anthropic's pricing does.

Every response carries usage.credits, split into base for the inference itself, modules for the firewall and anything else in the pipeline that ran, and the total. You read what a call cost from the call, not from a reconciliation at the end of the month.

The audit trail records the model you asked for, the model that actually served (they differ when an alias or a fallback chain moves the request), the region, the compliance mode, and the acting key. It never records the content of your prompt.

The point

Adding a frontier model is not itself news. The model exists whether or not we list it, and you could always call Anthropic directly. What is worth something is that a regulated EU team can now reach it without stepping outside the controls: one base URL, an acceptance that is recorded rather than assumed, a firewall that runs by default, a guard that refuses instead of guessing, and a log that answers the question when someone asks it six months later.

Enable anthropic/claude-opus-5 in your organization's model settings to start. If the external-routing acceptance is not on file, the first request will tell you so, which is the correct order for that conversation to happen in.