The short answer
Opus 5.5 makes Anthropic’s top production model meaningfully cheaper and faster, and for most teams the upgrade from Opus 5 will pay for itself. List prices drop 20%, cache reads drop 60%, and Anthropic claims a 40% lower cost to run on typical workloads plus more than 30% faster output. If you build on Claude through the Anthropic API, AWS, Google Cloud or Azure, the new model ID is claude-opus-5-5.
The part that deserves engineering attention is the new safeguard layer. Requests classified as sensitive cybersecurity work are handed to the older Opus 4.8, and API customers must opt in to that fallback — until they do, a flagged call returns HTTP 200 with a stop reason instead of an answer. In an AI agent pipeline, a successful status code with no usable output is exactly the kind of failure that slips past monitoring.
What changed in Opus 5.5?
The headline is cost. Anthropic cut Opus list pricing to $4 per million input tokens and $20 per million output tokens, a 20% reduction, while cache reads fell 60% to $0.20 per million tokens. Because long-running agents re-read large cached contexts on almost every step, the cache cut matters more than the list price for many production workloads — which is how Anthropic arrives at its “about 40% cheaper to run” figure. A new fast mode trades money for latency: up to 2.5 times the standard speed at double the price.
On capability, Anthropic positions Opus 5.5 as matching its larger Claude Fable 5.1 on most tasks. VentureBeat reported Anthropic-published scores of 66.4% on Terminal-Bench 4.0 versus 55.8% for Fable 5.1, and 57.8% versus 51.8% on CursorBench 4.0. These are vendor-run numbers, so treat them as a reason to test, not as a verdict. The model keeps a 1-million-token context window, and Anthropic says it writes more plainly and front-loads the important information in its answers.
Bloomberg framed the launch against Anthropic’s expected IPO: a cheaper flagship widens the gap between what enterprises pay per task and what they paid a year ago. For buyers, the practical signal is the pace — a major Opus release every two months means model upgrades are now a recurring engineering task, not an annual event.
How does the cyber fallback work?
Opus 5.5 ships with the same class of safety classifiers Anthropic already runs on Fable 5.1. When a classifier flags a request as sensitive cybersecurity work, the request is served by Claude Opus 4.8 instead. Anthropic says ordinary work such as finding and fixing bugs in your own code is still handled normally; the restriction targets higher-risk security tasks. Organizations doing legitimate defensive work can apply to the Cyber Verification Program, which now has three access tiers, to get full capability.
The behavior differs by surface. In the Claude apps the switch is automatic: the user sees a notice and the response is labeled with the model that answered. On the API, Anthropic’s help center says customers must opt into and configure fallbacks, and until they do, a flagged request returns a 200 response with a stop reason. That is a sensible safety default, but it means a security-adjacent agent can start returning empty results without a single error in your logs.
What it means for US & EU software teams
First, the unit economics of agents improve. Agent workloads are dominated by repeated context reads, so a 60% cheaper cache read changes which designs are affordable. Workflows that were routed to smaller models purely on cost — code review over large repositories, multi-step research, long customer-support threads — are worth re-costing on Opus 5.5. At the same time, the fast mode makes it easy to double spend without noticing; decide explicitly which paths genuinely need lower latency.
Second, “which model answered?” becomes an observability question. With fallbacks enabled, some requests in the same workflow may be served by Opus 4.8; without them, some may end in a safety stop. Both need to be visible. Treat the stop reason as a first-class signal in your agent loop, alert on unexpected stops, and record the model reported for each response alongside the one you requested.
Third, compliance documentation needs to match reality. Under the EU AI Act, GDPR accountability rules and most enterprise vendor-risk reviews, you are expected to know which system produced an output. A model inventory that says “Opus 5.5” while some answers come from Opus 4.8 is a small but real documentation gap. It is also a product question for anyone selling security tooling built on Claude: test your core prompts against the classifiers now, and decide whether you need Cyber Verification Program access before customers find the edge cases for you.
What to do before you switch
- Re-run your eval set. Compare Opus 5 and Opus 5.5 on your own tasks, including tone and output length — Anthropic changed the writing style on purpose.
- Check for API changes. VentureBeat reports that adaptive thinking is always on (controlled via the
effortparameter), forced tool use may return errors, and an older computer-use tool version is rejected. Verify against Anthropic’s migration notes before flipping production traffic. - Handle stop reasons explicitly. Make sure your agent loop distinguishes a completed answer from a safety stop and does not pass an empty result downstream as success.
- Decide on fallbacks. If your product touches security topics, choose whether to enable the Opus 4.8 fallback, apply for the Cyber Verification Program, or route those tasks elsewhere.
- Re-cost caching and fast mode. Measure cache hit rates after the switch, and gate fast mode behind the paths that actually need it.
Frequently asked questions
What did Anthropic release on September 22, 2026?
Anthropic released Claude Opus 5.5, the successor to Opus 5 (launched July 24, 2026). It is available through the Claude API as claude-opus-5-5 and on AWS, Google Cloud and Microsoft Azure. Anthropic says it matches Claude Fable 5.1 on most work, generates output more than 30% faster than Opus 5, and costs about 40% less to run on typical workloads. Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks.
How much does Claude Opus 5.5 cost?
Standard API pricing is $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads fell from $0.50 to $0.20 per million tokens and cache writes cost $5 per million tokens. A fast mode that runs up to 2.5 times the standard speed costs $8 per million input tokens and $40 per million output tokens.
What is the Opus 4.8 fallback for cybersecurity requests?
Opus 5.5 ships with safety classifiers for sensitive cybersecurity work. When a request is flagged, it falls back to the older Claude Opus 4.8. In Claude apps this happens automatically and the response is labeled with the model that answered. API customers must opt into and configure fallbacks; until they do, a flagged call returns a 200 response with a stop reason instead of an answer. Ordinary bug finding and fixing in your own code is still supported.
Should we migrate from Opus 5 to Opus 5.5 right away?
For most workloads the lower price and faster output make the upgrade attractive, but treat it as a model change, not a config tweak. Re-run your evaluation set, check prompt-cache hit rates, and review reported API changes: adaptive thinking is always on and controlled through the effort parameter, forced tool use may return errors, and an older computer-use tool version is rejected. Security-tooling products should also test which requests trigger the Opus 4.8 fallback.
Does the model fallback matter for compliance?
It can. If your audit trail, model inventory or EU AI Act documentation records that a system runs on Opus 5.5, some responses may in fact come from Opus 4.8 once fallbacks are enabled, and others may end in a safety stop. Teams in regulated sectors should log the model and stop reason reported for each response, not just the model requested, and reflect this behavior in their model documentation and risk assessments.
Sources
Anthropic — Introducing Claude Opus 5.5 (company announcement)
Claude Help Center — Why Claude switched models in your conversation
TechCrunch — Anthropic releases Opus 5.5 with lower prices and Fable-level performance
Bloomberg — Anthropic Unveils More Cost-Efficient Opus 5.5 Model Before IPO
VentureBeat — Anthropic releases Claude Opus 5.5, beating Fable 5.1 on key agentic benchmarks