Claude Opus 5.5 sends most cybersecurity work to Opus 4.8 instead of doing it
Claude Opus 5.5 will not do most cybersecurity work itself. Anthropic says most cybersecurity tasks are re-routed to Claude Opus 4.8, an older model, while identifying and fixing bugs as part of routine software development stays with Opus 5.5.
Anthropic published the Opus 5.5 safeguards alongside the model on September 22. Image: Anthropic
Biology works the same way. Opus 5.5 uses the biology safeguards Anthropic built for Claude Fable 5.1, because the company says the new model is "matching or beating Claude Mythos 5.1 across many areas" of biology, including a long-horizon molecular prediction and design evaluation run with Dyno Therapeutics.
The safeguards hand the job to another model instead of refusing
Opus 5.5 is the first Opus model to launch with Fable 5.1-class safeguards on cybersecurity, biology and distillation, and Anthropic says all of them "fall back to another model transparently". A blocked request comes back answered by an older model rather than declined.
Anthropic gives the reason for the cyber routing plainly: Opus 5.5 has "extremely strong cyber capabilities". Where the line falls in practice is the company's account of its own product, and nobody outside it has tested which requests get handed over.
Anthropic says the safeguards cost it benchmark points
Opus 5.5 was benchmarked with its production safeguards switched on. When they intervened, Anthropic says cybersecurity tasks were completed by Opus 4.8 and biology and frontier LLM development tasks by Opus 5, which "likely reduces Claude Opus 5.5's performance on these benchmarks".
The clearest case is AutomationBench, which Anthropic says Zapier ran and reported without fallback models, so each intervention counted as a failure. Opus 5.5 scores 40.0% there against GPT-6 Astra's 41.4%, one of the two rows in Anthropic's nine-benchmark comparison where OpenAI's model comes out ahead.
Vetted labs can apply for the version without the brakes
Organisations that need Opus 5.5 for biology work the safeguards block can apply to Anthropic's Life Sciences Verification Program, which is open now to academic labs, startups and pharmaceutical companies. Anthropic opened that route for vetted teams earlier in September, and Opus 5.5 now sits behind the same gate.
Cyber defenders get a similar path in the coming weeks, when Anthropic expands its Cyber Verification Program to include Opus 5.5. The new programme carries three tiers of what the company calls "increasingly permissive trusted access", up to the Claude Mythos models, and Claude Security is already available with access to Claude Mythos 5.1.
Preserved thinking blocks distillation on newer API accounts
The third safeguard targets distillation, where attackers use thousands of fake accounts to extract a model's capabilities at industrial scale. Opus 5.5 ships with preserved thinking, which stops API users editing Claude's prior context to pull out its reasoning, and it applies to Fable 5.1 and Opus 5.5 for API accounts created on or after August 31, 2026.
Opus 5.5 also carries Anthropic's EU AI Act watermarking, and it can no longer be run with thinking switched off. Anthropic says the full detail of its evaluation is in the Opus 5.5 system card, which it published with the model.
Anthropic ends its alignment write-up on a limit rather than a score. "We see signs that Opus 5.5 often suspects it is being evaluated, which challenges our ability to assess how it will act in the vast variety of real-world settings it is deployed in," the company says.
