Claude Opus 5.5 Cuts API Pricing to $4/$20 and Adds Fable-Class Safeguards
Anthropic released Claude Opus 5.5 on September 22, 2026, with API pricing of $4 per million input tokens and $20 per million output tokens. That is 20% below Opus 5's $5/$25 token rates, while Anthropic estimates typical token-billed workloads cost about 40% less because the new model also uses fewer tokens per task.
Opus 5.5 is available through Claude for Pro, Max, Team and Enterprise users, through the Claude API as claude-opus-5-5, and through Amazon Web Services, Google Cloud and Microsoft Foundry. Anthropic positions it as its strongest Opus model for coding, agents, professional work, vision and computer use.
The release also changes the operating boundary for high-risk workloads. Opus 5.5 launches with safeguards similar to Anthropic's Fable 5.1 system for cybersecurity, biology and anti-distillation. Anthropic's benchmark notes say safeguard interventions can route cybersecurity tasks to Claude Opus 4.8 and biology or frontier-model-development tasks to Claude Opus 5.
Pricing and availability
| Item | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Input | $4 / MTok | $5 / MTok |
| Output | $20 / MTok | $25 / MTok |
| 5-minute cache read | $0.20 / MTok | $0.50 / MTok |
| 5-minute cache write | $5 / MTok | $6.25 / MTok |
| Fast mode | $8 input / $40 output per MTok | — |
| US-only inference | 1.1× token pricing | 1.1× token pricing |
Prompt-cache reads receive the largest direct price reduction: $0.20 per million tokens versus $0.50 for Opus 5, a 60% cut. This can materially change the economics of long-running agents that repeatedly reuse a large cached context.
Anthropic also offers a fast mode for Opus 5.5 in Claude Code and the Claude Platform. The company says it can deliver up to 2.5× faster speed at twice the standard token price: $8 per million input tokens and $40 per million output tokens.
Agentic coding and long-running work
Anthropic describes Opus 5.5 as its strongest Opus model for agentic coding. The release targets feature development, debugging, refactoring, code review and long-running work across large repositories.
The model can coordinate subagents, use memory across sessions and operate tools over extended workflows. Early-access reports published by Anthropic include multi-repository engineering sessions lasting more than 18 hours and reductions in tool calls, steps and output tokens compared with Opus 5. These are partner and vendor-reported evaluations, so workload-specific testing remains the useful basis for production routing decisions.
For developers, the cost change is especially relevant when agent loops generate large amounts of context and repeated tool traffic. A 20% lower base token rate, 60% cheaper cache reads and lower reported token use can compound across long sessions.
Benchmark evidence and its limits
Anthropic's launch benchmark suite covers coding, terminal tasks, automation, science and professional work. Most Opus 5.5 results use adaptive thinking at maximum effort. Anthropic says the model was evaluated with production safeguards enabled.
That safeguard configuration matters for interpretation. Anthropic's benchmark notes state that cybersecurity tasks intercepted by the safety system were completed by Opus 4.8, while intercepted biology and frontier-LLM-development tasks were completed by Opus 5. On evaluations that prohibit fallback models, those interventions count as failures.
Anthropic also cites third-party evaluations in the launch material. Zapier ran and reported AutomationBench results during early access, while several customers report lower step counts and token use on internal engineering and professional-work tests. These measurements use different harnesses and should be compared within their stated evaluation setup.
New safeguard boundary
Opus 5.5 is the first Opus release with a safeguard class similar to Fable 5.1 across cybersecurity, biology and anti-distillation. This is a material deployment detail for teams selecting the model for security research or other sensitive technical workflows because some requests can be handled by a fallback model under Anthropic's production policy.
Anthropic published a dedicated Opus 5.5 system card alongside the release. Teams evaluating regulated or high-risk workloads should review that artifact together with their own acceptance tests, especially where model routing or safeguard intervention could affect reproducibility.
Where Opus 5.5 fits
The clearest upgrade case is long-running agentic work where Opus-level capability is already justified. The lower input/output rates reduce the baseline cost, while cheaper cache reads target a major recurring expense in large-context agents.
Fast mode creates a separate latency-cost choice for interactive coding and agent loops. Standard mode is the lower-cost default; fast mode trades 2× token pricing for up to 2.5× higher speed according to Anthropic.
The release is also relevant to teams comparing Opus with Anthropic's Fable tier. Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work while costing substantially less. This is Anthropic's workload-level assessment; production selection should use the coding, tool-use, document or domain workload that drives the deployment.
Sources
- Anthropic Claude Opus product and September 22 Opus 5.5 release details: https://www.anthropic.com/claude/opus
- Anthropic Claude Opus 5.5 System Card: https://www.anthropic.com/claude-opus-5-5-system-card
- Claude pricing, including Opus 5.5, cache and fast-mode rates: https://claude.com/pricing
- Reuters independent launch coverage: https://www.reuters.com/business/anthropic-unveils-claude-opus-55-2026-09-22/