AI
Anthropic Launches Opus 5.5 With 20% Price Cut, But Safety Guardrails May Route Requests to Older Models
The new Opus 5.5 model cuts costs and speeds up agent tasks, but safety classifiers can silently redirect requests mid-chain to less capable predecessors.
The New Stack
Anthropic has released Opus 5.5, its latest frontier model, with aggressive cost reductions and significant speed improvements. The pricing cuts—20% across the board for base tokens and caching—pair with claims of 30% faster output generation and substantially reduced token consumption on real-world tasks.
Pricing and Performance Gains#
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from Opus 5's rates. Cache operations are also cheaper: reads fall to $0.20 per million from $0.50, and writes drop to $5 from $6.25. A fast mode is available at $8/$40 per million tokens for developers who want speeds up to 2.5 times faster.
Anthropic claims overall savings reach closer to 40% when factoring in token efficiency and output generation speed. Enterprise testers reported dramatic improvements: Box found Opus 5.5 consumed roughly one-third the tokens of its predecessor while reducing verbosity by 40%, and GitHub's testing in Copilot CLI and VS Code showed the model completing terminal tasks in less than half the steps.
On several agentic coding benchmarks—Terminal-Bench 4.0 (66.4% vs. Sonnet 5.1's 55.8%) and FrontierCode (54.4% vs. 50.3%)—Opus 5.5 pulls ahead. The company cautiously notes that such margins may not translate to meaningful real-world differences. At default effort on FrontierCode, Opus 5.5 edges GPT-6 Astra at roughly 20% of the per-task cost, and it beats GPT-5.6 Sol on CursorBench at about a third of the cost.
The Safety Classifier Routing Gotcha#
Behind the headline improvements lies a significant architectural detail: Anthropic's safety classifiers can transparently reroute requests mid-chain to older models. When safety checks flag potential issues in cybersecurity, biology, or frontier LLM development, the request may be handled by Opus 4.8 or Opus 5 instead of Opus 5.5.
For developers building multi-turn agent workflows, this introduces unexpected complexity. A single conversation could involve requests handled by models with different capabilities, potentially affecting downstream steps and creating inconsistencies not visible in traditional benchmarks. The routing happens without explicit notification, which could mask performance variations in production systems.
Anthropic is expanding its verification programs to mitigate this. Vetted organizations can join the Life Sciences Verification Program to bypass the biology classifier, and a Cyber Verification Program with three tiers for trusted access is rolling out soon, including access to Claude Mythos models.
