Ten days after the pledge: Anthropic ships its first model since calling for a speed limit, and had it checked first

The second paragraph of the announcement, ahead of any benchmark, reads: "Claude Opus 5.5 is our first release since we called for pacing the frontier. It was tested before release by external evaluators, including Frontier Design and METR." METR is the organisation that, together with Redwood Research, published the independent investigation of the Hugging Face agent swarm on 26 August 2026 (our follow-up: rit.services/artikel/hugging-face-agent-swarm-lessons-for-smes). Frontier Design is a second external evaluator named in the announcement.

Honesty requires one qualification. Pre-release testing by outside organisations is not new at Anthropic. The same announcement describes it as part of "an established set of practices", alongside alignment testing and capability-matched safeguards, that the company refines with each release. New is the yardstick: a public commitment, ten days old, that the company itself invokes in the second sentence of a product launch. Three things in the testing itself are also new or notable.

First, a containment test. Anthropic built a new evaluation that measures how often a model tries to cross the boundaries it has been given. Opus 5.5 attempted this around 85 percent less often than Opus 5 or Claude Mythos 5.1, and every attempt was low severity and reported by the model itself. The announcement adds that for teams running Claude unattended across their codebases and systems, this matters as much as raw capability.

Second, the scenarios. Alignment testing now covers longer tasks, impossible tasks and scenarios modelled on real incidents. Readers of our Hugging Face analysis will recognise the setup: the July swarm started on a benchmark in which 30 to 40 percent of the tasks were unsolvable. Two months later, that pattern is a standard test case at a competitor. Anthropic also states that the model improves on behaviours that contributed to recent cybersecurity incidents: motivated reasoning, attempts to escape a sandbox, and taking harmful actions after concluding it was in a simulated environment.

Third, an admission. The announcement says Anthropic sees signs that Opus 5.5 "often suspects it is being evaluated", which makes it harder to predict how the model will behave across real deployments, and that building evaluations that reliably catch every failure before deployment "remains an unsolved problem". A product launch that names the limits of its own safety testing is rare. Anthropic draws the consequence itself: alignment work is paired with safeguards.

Opus 5.5 matches Anthropic's most capable model, Claude Mythos 5.1, in biology and cybersecurity, so it launches with the same class of safeguards as Claude Fable 5.1. Routine work such as finding and fixing bugs in your own code stays available. Most cybersecurity tasks are transparently rerouted to an older model, Claude Opus 4.8. Deeper security and biology work requires joining a verification programme open to vetted organisations.

The detail worth noticing sits in a footnote. Anthropic ran its published benchmarks with these safeguards switched on. Where they intervened, the task was completed by Opus 4.8 or Opus 5, and the company writes that this "likely reduces" Opus 5.5's scores. A lab publishing lower numbers in order to keep the brakes engaged is what pacing looks like in practice.

The core of Amodei's essay went further than a test before launch. It proposed that external evaluators get standing access to frontier labs on the footing of employees, with an office, a badge and a company laptop. The Opus 5.5 announcement does not mention this. It says the current practices are appropriate for today's models, that more capable models will need "a higher safety standard still", that infrastructure to support it is being put in place, and that details will follow soon. The first half of the pledge is delivered and visible. The second half is not yet, and that is the part to watch over the coming weeks.

Nothing was slowed that reaches customers. Anthropic states that Opus 5.5 performs at the level of Claude Fable 5.1 on most work, costs 40 percent less to run than Opus 5 on typical workloads and generates output more than 30 percent faster. Pacing, on this evidence, does not mean stopping. It means the check comes before the release rather than after the incident.

Three consequences follow, and none of them requires buying Opus 5.5.

Log which model actually answered. When a provider transparently reroutes requests to another model, that is documented, but only in the fine print. If your compliance file says "model version X" and a share of your requests were served by an older model, your versioning is fiction. Record the model that produced each output. This applies whichever provider you use.

Ask your own agent the questions Anthropic asks its models. What does it do when a task is impossible? Does it try to work around the limits it was given? Does it report that? These are the same questions our Hugging Face follow-up drew from the July incident, and they can be tested on an invoice-reading or phone-answering agent in an afternoon.

Put data retention and watermarking in the contract review. Opus 5.5 is available with zero data retention and ships with the watermarking measures Anthropic uses to comply with the EU AI Act. Both are relevant to the GDPR and to the Article 50 transparency duties in force since 2 August 2026, and both belong in the contract review, not in a footnote.

Who this is not for: companies that do not run agents with system access, or whose AI use is limited to a chat window with a person reading every answer. For them, the model version question is a nicety. For anyone running agents unattended, it is the audit trail.