Claude Opus 5.5 is a proprietary language model from Anthropic for agentic coding and demanding knowledge work—in other words, multi-step tasks in which the model uses tools autonomously and works toward an outcome over a longer period. In all nine tests Anthropic reported for its release in September 2026, it ranks ahead of Claude Fable 5.1 and its predecessor Claude Opus 5. According to Anthropic, operating costs drop by around 40% versus Opus 5 in typical use, and output is over 30% faster.
For businesses, data protection is the key factor: Like earlier Opus models, Opus 5.5 is available with Zero Data Retention, meaning requests are not stored after processing. With Amazon Bedrock, processing can also be restricted to European data centers. There are no open model weights. You use the model via Anthropic, the major cloud platforms, or an access route such as our MaaS Proxy.
Claude Opus 5.5
Anthropic
September 2026
Proprietary (no open weights)
Proprietary, reasoning permanently enabled, input text and image, output text
Not yet released
Not yet released. Adaptive Reasoning, controllable via five effort levels
1,000,000 Tokens (max. 128,000 output tokens)
The standard model with the API identifier claude-opus-5-5. Available via the Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Azure, as well as via our MaaS Proxy.
The same model with faster output at double the price per token. Useful when short response times matter more than cost.
With the identifier eu.anthropic.claude-opus-5-5, processing runs only in Amazon Bedrock’s European regions: Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland, London, and Paris. Zurich and London are outside the EU.
Opus 5.5 is powerful and significantly cheaper than its predecessor, but it is not the most economical choice for every task. We will work with you to determine where it pays off and how your costs are likely to develop realistically.
Anthropic publishes neither parameter count nor architecture and provides no information on Opus 5.5’s training data or training methods. The model’s knowledge cutoff is June 2026. What is documented is that the model uses fewer tokens per task than Opus 5 and that this efficiency is reflected in the price.
Reasoning—i.e., step-by-step thinking through a task before answering—is permanently enabled and cannot be switched off. Instead, you control depth of thinking via five effort levels from low to max; the default is medium. Even the lowest level delivers strong results, according to Anthropic: In code review, Opus 5.5 found 72% of known errors there, while Opus 5 found 56% on the high level.
Data protection: Opus 5.5 is available with Zero Data Retention and can be restricted to European regions via Amazon Bedrock. Claude Fable 5.1, by contrast, is a Covered Model with at least 30 days of retention.
Safeguards: If they intervene for cybersecurity requests, Claude Opus 4.8 takes over. Restrictions apply for biology. Outputs carry watermarks under the EU AI Act.
Behavior: According to Anthropic, Opus 5.5 attempted to bypass system boundaries about 85% less often than Opus 5 and achieved the lowest success rate in Gray Swan’s prompt-injection test. Because the model often suspects it is being tested, Anthropic says the security assessment is only of limited significance.
For demanding tasks that require depth and reliability, including multi-hour agent runs without ongoing supervision. Opus 5.5 is less suitable for large volumes of standard tasks such as classification and for scenarios that require full control over the infrastructure. In those cases, smaller or open models in your own data center are the better choice. We would be happy to clarify this with you in an AI consultation.
Leading in Anthropic’s benchmarks: In all nine reported tests, Opus 5.5 ranks ahead of Fable 5.1 and Opus 5.
Cheaper than its predecessors: 20% lower per token than Opus 5 and 60% lower than Fable 5.1.
Data protection: Zero Data Retention and processing in European regions via Amazon Bedrock.
Large context window: One million input tokens and 128,000 output tokens.
Robust research: In Anthropic’s strict source-check, Opus 5.5 passed 16 of 18 reports; Fable 5.1 and Opus 5 passed none.
No on-premises operation: The model is proprietary and can only be used via providers.
Reasoning cannot be disabled: Latency and token usage should be remeasured when switching.
Rerouting sensitive requests: Cybersecurity and biology requests are restricted or routed to other models.
Not leading in every test: GPT-6 Astra scores higher on AutomationBench and Terminal-Bench-Science.
We integrate Opus 5.5 into your systems via our MaaS Proxy, with centralized budget control, transparent billing in euros, and gateway operation in German data centers. If desired, we can also assess which applications are better run on open models in your own infrastructure.
The price per token is 20% lower than Claude Opus 5 and 60% lower than Claude Fable 5.1. Cached request parts (cache reads) are 60% cheaper than with Opus 5. Because Opus 5.5 also uses fewer tokens per task, total costs drop by around 40% in typical use, according to Anthropic. Via our MaaS Proxy, you are billed in euros.
Opus 5.5 ranks ahead of Claude Fable 5.1 in all nine tests reported by Anthropic and is 60% cheaper per token. Important for businesses: Fable 5.1 is a Covered Model with at least 30 days of retention, whereas Opus 5.5 is available with Zero Data Retention. In addition, Amazon Bedrock offers a European profile for Opus 5.5, but not for Fable 5.1.
The prerequisites are good: Zero Data Retention is available, and with Amazon Bedrock, processing can be restricted to European regions. The specific assessment depends on your use case, your data, and the chosen procurement route. We will advise you individually.
No. The model is proprietary and can only be used via Anthropic, connected cloud platforms, or an access point like our MaaS Proxy. Those who need full control over the infrastructure are better served with open models, such as GLM-5 or DeepSeek-V4-Pro.
According to Anthropic, operating costs drop by around 40%, output is over 30% faster, and Opus 5.5 leads in all reported tests. Before switching, check whether your integration is affected by Preserved Thinking and how permanently enabled reasoning impacts your response times.
Would you like individual advice?
Our AI experts are here for you!
