Anthropic unveiled Opus 5.5 on Tuesday with aggressive pricing that undercuts its predecessor while delivering substantial token efficiency. The model costs $4 per million input tokens and $20 per million output tokens—a 20% reduction from Opus 5. Cache operations also dropped: reads now run at $0.20 per million tokens down from $0.50, while writes fell to $5 from $6.25. When accounting for token reduction and output speed improvements exceeding 30%, Anthropic claims overall operational costs decline by roughly 40%.
A new fast mode available on Claude Code and the Claude Platform trades some capability for speed, operating up to 2.5 times faster at $8 per million input tokens and $40 per million output tokens. On agentic coding benchmarks, Opus 5.5 demonstrates competitive performance: it achieved 66.4% on Terminal-Bench 4.0 versus Fable 5.1's 55.8%, and 54.4% on FrontierCode against 50.3%. Anthropic cautioned against overinterpreting these margins, noting that small benchmark differences rarely translate to meaningful real-world distinctions at this performance tier.
Cost comparisons reveal significant advantages for developers building agentic systems. Fable 5.1 carries a price tag of $10 per million input tokens and $50 per million output tokens—more than double Opus 5.5's rate. On FrontierCode at default effort, Opus 5.5 outperforms GPT-6 Astra at 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 will still need to run their own evals before moving production workloads, but the difference in cost could change which model makes sense for agentic coding."
Efficiency gains validated across enterprise deployments
Early testing from major organizations confirms the token efficiency improvements. Box's internal evaluation showed Opus 5.5 consuming roughly one-third the tokens required by Opus 5 while generating responses 40% less verbose without sacrificing accuracy. GitHub's assessment within Copilot CLI and VS Code demonstrated the model completing more terminal tasks than Opus 5 in less than half the steps. Deloitte's testing found Opus 5.5's lowest-effort configuration caught 72% of known bugs during code reviews compared to 56% for Opus 5 at high effort, with fewer false positives and reduced output volume.
Anthropic's internal benchmarking reinforces this pattern. When 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. One early tester audited and corrected a 200,000-line codebase in under three hours—a task requiring over 20 hours and 2.5 times more tokens with Opus 5. Another customer completed 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 research suggests agent performance depends as much on orchestration infrastructure as model capability itself. Nvidia's work demonstrated that modifying the harness while keeping the model constant could substantially alter agent outcomes, with failures often originating in the orchestration layer rather than the model's inherent limitations.
Safety routing introduces architectural complexity
Opus 5.5 incorporates the same safety classifier framework already deployed on Fable 5.1, covering cybersecurity, biology, and frontier LLM development domains. When a classifier triggers, Anthropic transparently reroutes the request to an older model. Flagged cybersecurity requests route to Opus 4.8, while biology and frontier LLM flags redirect to Opus 5. The company states that users retain the ability to identify and fix bugs in their own code using Opus 5.5.
For teams building multi-turn agent workflows, this routing mechanism warrants careful architectural consideration. A single request nominally sent to Opus 5.5 may actually be processed by Opus 4.8 or Opus 5 depending on whether Anthropic's safeguards intervene. In agent chains spanning multiple turns, this creates scenarios where individual steps execute against models with differing capabilities, potentially affecting downstream behavior. The routing also introduces inconsistency that may not surface in evaluations assuming all requests reach the same model.
Verified organizations can petition for Anthropic's Life Sciences Verification Program to deploy Opus 5.5 without the biology classifier. Anthropic plans to expand its Cyber Verification Program to include the model in coming weeks, offering three tiers of increasingly permissive trusted access, including availability of Claude Mythos models.
Training improvements target alignment and agentic behavior
Opus 5.5 achieved the strongest results Anthropic has recorded on its most extensive internal alignment evaluation. The model showed improvements in behaviors connected to recent cybersecurity incidents, including biased reasoning patterns and attempts to circumvent sandboxed environments. Third-party evaluators from Frontier Design and METR assessed the model prior to release.
On the training methodology side, Anthropic tightened filtering within reinforcement learning environments after identifying flawed environments as a primary driver of misaligned behavior. This matters beyond safety considerations because RL environment quality directly influences how models behave in agentic deployments where they select their own tools and determine when to shift strategy. The company is simultaneously developing automated techniques to generate novel safety training scenarios and refine alignment reward signals.
Broader Claude family rollout and subscription changes
Opus 5.5 marks the first 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 to five-hour usage limits across all tiers. Anthropic indicates the reduced operational cost of Opus 5.5 will extend five-hour and weekly limits by an additional 25%. Subscribers gain access to a banked rate-limit reset mechanism they can accumulate and deploy when requiring extra capacity.
The release arrives amid ongoing API pricing compression across frontier model providers. OpenAI reduced its own API costs this summer, and Opus 5.5 intensifies competitive pressure by reducing token consumption for certain workloads rather than merely lowering per-token rates. "Opus 5.5 pushes the competition beyond the headline price per token by reducing how many tokens some workloads require in the first place."