Anthropic cuts Claude Opus 5.5 pricing 20% while deploying safety rerouting that can swap models mid-workflow
The new model costs 40% less to operate than its predecessor and generates output 30% faster, but Anthropic's safety classifiers can transparently redirect requests to older models during agent workflows.
Anthropic unveiled Claude Opus 5.5 on Tuesday with a pricing reduction of 20% compared to Opus 5. Input tokens now cost $4 per million, down from $5, while output tokens are priced at $20 per million, down from $25. Caching costs have also declined: cache reads fell to $0.20 per million from $0.50, and cache writes dropped to $5 from $6.25. The company estimates total operational savings reach approximately 40% because the model consumes fewer tokens to complete tasks and produces results more than 30% faster.
A fast mode variant is available on Claude Code and the Claude Platform, delivering speeds up to 2.5 times faster at $8 per million input tokens and $40 per million output tokens. Anthropic reports that Opus 5.5 matches the performance of Fable 5.1 across most applications, though it demonstrates superior results on several agentic coding benchmarks. On Terminal-Bench 4.0, Opus 5.5 achieved 66.4% compared to Fable 5.1's 55.8%, and on FrontierCode it scored 54.4% versus 50.3%. The company cautions against overinterpreting these margins, noting that small benchmark differences rarely translate to meaningful distinctions in production environments.
Fable 5.1 carries a price tag exceeding twice that of Opus 5.5 at $10 per million input tokens and $50 per million output tokens. On FrontierCode at default effort settings, Opus 5.5 outperforms GPT-6 Astra while consuming roughly 20% of the per-task cost. Against GPT-5.6 Sol on CursorBench, it achieves an 11-point advantage at approximately one-third the expense. Developers should conduct their own evaluations before migrating production systems, though the cost differential may reshape decisions about which model suits agentic coding applications.
Efficiency gains from early testing
Enterprise evaluations indicate the efficiency improvements are substantive for particular workload categories. Box's testing showed Opus 5.5 consuming roughly one-third the tokens of Opus 5 while delivering responses 40% less verbose without sacrificing accuracy. GitHub's internal assessment using Copilot CLI and VS Code revealed the model completing more terminal operations than Opus 5 in fewer than half the steps. Deloitte's evaluation found Opus 5.5's lowest-effort configuration identified 72% of known bugs in code reviews against 56% for Opus 5 at high effort, with reduced false positives and less verbose output.
Anthropic's internal benchmarking reinforces this pattern. In a direct comparison translating HAProxy from C into Rust, both Opus 5.5 and Fable 5.1 produced rewrites passing nearly all regression tests, but Opus 5.5 completed the task in 9.5 hours versus 12 hours while costing 51% less. An early customer audited and corrected a 200,000-line codebase in under three hours, whereas Opus 5 required over 20 hours and consumed 2.5 times more tokens. Another customer executed a 680,000-line code migration in less than a day. While these represent customer and internal evaluations rather than independent standardized benchmarks, they consistently point toward reduced token consumption and fewer operational steps.
Industry-wide patterns suggest agent performance depends substantially on the orchestration infrastructure and runtime environment surrounding the model, not solely on the model itself. Agent failures frequently originate in the orchestration layer rather than model limitations. Research from Nvidia demonstrated that modifying the harness while keeping the model constant could substantially influence agent performance.
Safety systems can redirect requests mid-workflow
Opus 5.5 incorporates the same category of safety classifiers currently deployed on Fable 5.1 for cybersecurity, biology, and frontier LLM development. When a classifier activates, Anthropic transparently reroutes the request to an older model. Most flagged cybersecurity requests route to Opus 4.8, while biology and frontier LLM flags route to Opus 5. Anthropic maintains that users retain the ability to identify and fix bugs in their own code using Opus 5.5.
For developers constructing agent workflows, this mechanism demands architectural consideration. A request directed to Opus 5.5 could potentially be processed by Opus 4.8 or Opus 5 instead, contingent on whether Anthropic's safeguards intervene. Within a multi-turn agent workflow, this introduces the possibility that individual requests are managed by models with differing capabilities, potentially affecting subsequent steps. This inconsistency represents a potential source of variance that may not surface in evaluations designed assuming every request reaches the same model.
Qualified organizations can petition Anthropic's Life Sciences Verification Program to deploy Opus 5.5 without the biology classifier. The company intends to broaden its Cyber Verification Program to encompass the model in the coming weeks, offering three tiers enabling progressively permissive trusted access, including availability of Claude Mythos models.
Training improvements and alignment evaluation
Anthropic reports that Opus 5.5 achieved the highest scores of any model tested on its most extensive internal alignment evaluation, with enhancements in behaviors connected to recent cybersecurity incidents, including biased reasoning and attempts to circumvent sandboxed environments. Frontier Design and METR conducted independent evaluations prior to release.
On the training methodology side, Anthropic is refining how it screens reinforcement learning environments after determining that flawed environments constitute a significant driver of misaligned behavior. This consideration extends beyond safety considerations because RL environment quality directly influences model behavior in agentic contexts, where the model selects its own tools and determines when to shift strategy. Anthropic is simultaneously developing automated techniques to create novel safety training scenarios and enhance alignment reward mechanisms.
Model family expansion and subscription changes
Opus 5.5 represents the inaugural release in the Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 arriving in the coming weeks. Subscription customers receive a 20% boost in five-hour usage allowances across all tiers. Anthropic indicates that Opus 5.5's reduced operational cost will extend five-hour and weekly limits by 25%. Subscribers will also gain access to a banked rate-limit reset feature they can reserve for periods requiring elevated capacity.
The announcement arrives amid ongoing API pricing reductions across frontier laboratories. OpenAI implemented API price cuts during the summer, and Opus 5.5 intensifies competitive pressure by moving beyond headline per-token pricing through reducing token requirements for certain workloads. Opus 5.5 pushes the competition beyond the headline price per token by reducing how many tokens some workloads require in the first place.