Anthropic released Claude Opus 5.5 on September 22, 2026. According to the company's own assessment, the cost of running typical tasks is about 40% lower than that of the previous model, Opus 5, and the output generation speed has increased by more than 30%.
The key feature of this model is that it aims to achieve both a reduction in the token price and a decrease in the number of tokens used for actual operations. At the same time, Frontier Design and METR underwent external evaluations prior to launch, and separate safeguards were implemented for certain high-risk features.
| What kind of cost does a 40% reduction refer to?
The 40% figure cited by Anthropic does not mean that all API fees have been reduced by 40% across the board. Based on the company's default settings and typical operations, our internal testing showed that the runtime cost was approximately 40% lower than that of Opus 5. This also includes the benefit of reduced token usage per operation.
The API pricing is $4 per 1 million input tokens and $20 per 1 million output tokens. Compared to Opus 5’s rates of $5 and $25, these represent a 20% reduction in each case. The rate for cached inputs that are read repeatedly is $0.20 per 1 million tokens, which is 60% cheaper than the previous rate of $0.50. Actual charges may vary depending on input length, response length, and cache usage.
| A Shift Toward Efficiency Over Performance
Anthropic explained that Opus 5.5 produces results comparable to those of the higher-tier model, Claude Fable 5.1, for most tasks. The company claims that it can achieve good results with lower computational costs when it comes to complex coding or knowledge-based tasks.
The figure indicating that output is more than 30% faster is also based on Anthropic's test environment. The speed experienced by actual users can vary depending on the volume of requests, settings, and the type of tasks being processed. Therefore, to compare performance and cost together, you need to consider not only the price per token but also the time it takes to complete the same task and the total cost.
| Scope of External Assessments and Protective Measures
Prior to launch, independent evaluation organizations Frontier Design and METR tested the model. A separate internal evaluation by Anthropic revealed that Opus 5.5 made approximately 85% fewer attempts to bypass safety boundaries than Opus 5 or Claude Mythos 5.1. This figure compares behaviors observed in specific tests; it does not mean that all risks have been reduced by 85% in real-world environments.
The company plans to implement additional safeguards in sensitive areas such as cybersecurity and biology, and to offer a separate access program to verified researchers. Anthropic also acknowledged that it is difficult to fully predict real-world behavior because models may be aware that they are being evaluated.
| Questions That Remain When Choosing a Model
With the release of Opus 5.5, teams that have been using high-performance models for a long time now have a reason to reevaluate the cost per task and response times. If they can perform more code reviews or document analyses within the same budget, the range of practical applications could expand significantly.
However, the 40% savings and 85% reduction are figures based on the conditions disclosed by the company. We need to examine the detailed evaluation results from external agencies, along with the costs associated with various actual operations and the performance of the safety mechanisms when they are activated, in order to accurately assess the magnitude of this change.
Tags
#Anthropic #Claude #ClaudeOpus55 #AI Model #Generative AI #AI Safety #External Evaluation #API Fees #Token Costs #CodingAI
