However, Anthropic has rolled out Claude Opus 5.5, the first of its newly developed 5.5 series, and the context around it is just as important as the model itself. It is the company’s first new launch following the public statement by its CEO, Dario Amodei, to “pace the frontier,” in which he urged for a slower pace of development to ensure that capabilities do not surpass safety efforts. Opus 5.5 is about equally capable of performing tasks as Claude Fable 5.1 but is 40% cheaper to operate than Opus 5.
Also read: Zoho wants Zia to become the interface for enterprise software
In terms of benchmarks, Opus 5.5 dominates agentic coding, computer usage, and knowledge work, surpassing its competitors GPT-6 Astra and GPT-5.6 Sol on tests such as Terminal-Bench 4.0 and GDPval-AA v2.1. But Anthropic itself warns that, at this level, benchmark differences are less of a trustworthy indicator. Efficiency is where the story lies: one tester was able to migrate a 680,000-line codebase in less than a day, while another tested auditing a 200,000-line codebase in less than three hours compared to more than 20 hours by Opus 5, all the while using 2.5x less tokens. Pricing follows suit. Input and output tokens are 20% cheaper, and cache reads, which represent the majority of agentic workload cost, are 60% cheaper, at $0.20 per million tokens.
The next headline gain is communications. According to Anthropic, Opus 5.5 communicates more effectively, highlights the most critical aspects of information, and does not use a verbose and jargon-laden language that was criticized in Opus 5. The number of tokens and steps required to achieve the same goal were significantly lower in tests conducted by enterprise users, such as GitHub, Stripe, and Box. In one of Stripe’s test cases, the system successfully rebased a 40-pull-request code in a dozen sub-sessions, where all changes went through CI at night.
Also read: Snapdragon 8 Elite Gen 6 vs 8 Elite Extreme Gen 6: Who really needs the Extreme?
The next topic where the “pacing” theme finds itself reflected is safety. Opus 5.5 achieved the best results that Anthropic ever had with its behavioral audit, consisting of almost 2,000 simulated scenarios. The new version is less likely to take irreversible actions or bypass containment borders; the number of such actions decreased by 85% compared to Opus 5 and Claude Mythos 5.1. It is also more robust against prompt injection attacks, showing the same level of resistance as Fable 5.1 tested by Gray Swan.
As the model is equally good as Claude Mythos 5.1 biologically and in cybersecurity, it is being released with the same level of protection as Fable 5.1. Almost all cybersecurity actions are automatically transferred to the older version – Opus 4.8, and sophisticated biology experiments require special access through Anthropic Life Sciences Verification Program. Moreover, Opus 5.5 includes “preserved thinking” which means that Claude’s thinking cannot be distilled through the API.
An important feature of the new version is that it is not hiding the limitation of the previous versions but announcing it explicitly. Opus 5.5 tends to notice whether it is being tested more and more often which makes difficult for Anthropic to foresee its behavior in reality, and the issue will become more and more serious as the models become more efficient and diverse deployment environments appear.
Claude Sonnet 5.5 and Claude Haiku 5.5 will be launched in the nearest future. In the meantime, Opus 5.5 is available at Claude Platform and in major clouds, and it can be seen as the proof that pacing the frontier does not imply to slow down but move forward carefully.
Also read: GPT-6 Astra vs Sol vs Luna: Which One Is for You