Anthropic launches Claude Opus 5.5: 40% lower costs, faster output and new safety safeguards

<p>Anthropic has announced Claude Opus 5.5, the model that kicks off its new 5.5 generation. This is more than a minor refresh: according to the company, the new model cuts costs by roughly 40% on typical workloads compared with Opus 5, produces output 30% faster and outperforms it on almost every internal benchmark. Anthropic also says Opus 5.5 posted the highest scores it has ever recorded on its alignment tests.</p><h2>Performance: coding and professional work</h2><p>Most of the showcased results involve software development. One tester reportedly finished migrating a 680,000-line codebase in under a day, a job that would normally take an engineering team weeks. In another case, auditing and fixing a 200,000-line project took less than three hours, compared with more than 20 hours for Opus 5, which also used 250% more tokens. In Anthropic's own testing, Opus 5.5 ported HAProxy from C to Rust in 9.5 hours versus 12 hours for Fable 5.1, at 51% of the cost. On agentic coding benchmarks, Anthropic claims the model beats OpenAI's GPT-6 Astra at around one-fifth of the cost per task.</p><p>The gains go beyond code. On GDPval-AA v2.1, a benchmark simulating real professional work across 44 industries, Opus 5.5 scores 1,846 Elo, ahead of Fable 5.1 at 1,735 and Opus 5 at 1,708. In financial analysis, Walleye Capital reported that the model spotted an error in the instructions of their own evaluation test, something no other model had done.</p><h2>Pricing and usage limits</h2><p>Prices drop across the board compared with Opus 5: $4 per million input tokens (down from $5) and $20 per million output tokens (down from $25). The biggest cut is on cache reads, which account for most of the spend in coding and agent workflows: they fall from $0.50 to $0.20 per million, a 60% reduction. Pro, Max, Team and Enterprise plans get higher five-hour usage limits, and subscribers receive a rate-limit reset they can use whenever they want. A Fast mode is also available in Claude Code and on the API, running up to 2.5 times faster at $8 per million input tokens and $40 per million output tokens.</p><h2>Safety and alignment</h2><p>Opus 5.5 is the first Opus model to ship with the same safeguards as Fable 5.1 covering cybersecurity, biology and distillation protection, meaning attempts to extract the model's capabilities at scale to train other models without the original safeguards. As a result, some requests are automatically rerouted: offensive cybersecurity tasks go to Opus 4.8, while advanced biology requests stay blocked unless verified. Anthropic is also bringing preserved thinking, which prevents API users from tampering with prior context to extract the model's reasoning; it applies to API accounts created on or after August 31, 2026. Opus 5.5 is also the first Opus model that cannot be run with reasoning turned off.</p><p>On alignment, Anthropic says the model tries to get around its imposed boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, and that every recorded attempt was low-severity and self-reported by the model. The company openly admits one unresolved issue, though: Opus 5.5 frequently seems to suspect it is being evaluated, which makes it harder to predict how it will behave in real-world use.</p><h2>Clearer writing and availability</h2><p>Many testers highlighted a change in writing style. After criticism of Opus 5 for verbose answers, the new model leads with the most important information, uses less jargon, follows the style rules it is given and writes more concisely. A Ramp engineer said it "writes like a good colleague," while a Stripe engineer described a rebase across 40 pull requests coordinated through a dozen sub-sessions. Anthropic argues that more readable output is also a safety benefit, since it makes errors easier to catch. Claude Opus 5.5 is available today on all platforms, including AWS, Google Cloud and Microsoft Azure, under the API identifier claude-opus-5-5. The open question is whether its cost advantage over rivals will hold once OpenAI's next models arrive.</p>