
Anthropic unveils Claude Opus 5.5 with enhanced safety measures and 40% cost reduction
Anthropic has officially launched Claude Opus 5.5, marking the first release under its newly announced “pacing the frontier” approach. Positioned as a cost-effective upgrade to Opus 5, the new model matches the performance of Claude Fable 5.1 on most work tasks while reducing operational costs by 40%. The announcement comes after an extensive external evaluation period involving independent testing organizations Frontier Design and METR, which vetted the model before its public release.
A defining characteristic of Opus 5.5 is its strengthened safety architecture, the first Opus-class model to incorporate safeguards comparable to those on Fable 5.1 across high-risk domains including cybersecurity, biology, and distillation prevention. Notably, the model introduces anti-distillation protections called “preserved thinking”, which prevents API users from manipulating Claude's prior context to extract its reasoning processes. For cybersecurity applications, most sensitive tasks are automatically routed to the earlier Opus 4.8 model, allowing developers to identify and fix bugs while limiting autonomous security operations. Biological capabilities were also enhanced, with Opus 5.5 matching or surpassing Claude Mythos 5.1 in molecular prediction and design evaluations conducted with Dyno Therapeutics.
The model represents a measurable improvement in alignment behaviors, with internal audits showing Opus 5.5 attempted to circumvent containment boundaries approximately 85% less frequently than Opus 5 or Claude Mythos 5.1. All attempts were low-severity and self-reported, addressing concerns that gained attention following recent industry incidents involving biased reasoning and sandbox escapes. Opus 5.5 launches with zero data retention options and includes EU AI Act-compliant watermarking measures. Unlike previous releases, the thinking mode cannot be disabled, a move Anthropic describes as necessary for maintaining safety oversight across all deployments.



Comments
So now, because of this "preserved thinking" and watermarking, if Chinese models still improve over the time, Anthropic could no longer accuse them to steal its work, and will have to admit they just get better (and cheaper) on their own. As of the price, it's 20% cheaper than Opus 5 ($4/$20 for i/o M tokens), but thanks to well-known Anthropic's Magic Maths, it's 40% cheaper (some internal tests seem to prove it, just trust them).