Anthropic launches Claude Opus 5: Key Updates and Features

Availability and Integration

Claude Opus 5, Anthropic’s new flagship model, became available on July 24. The model is immediately integrated into the API, Amazon Bedrock, Google Cloud, Microsoft Foundry, as well as the Claude apps. Anthropic positions Opus 5 as a significant leap forward compared to previous versions.

Plans and User Access

With the Claude Max plan, Opus 5 is now the default model. For Pro subscribers, this is the most powerful model available. Free-tier users continue to work with Sonnet 5 and do not have access to Opus 5. Unlike some competitors, Anthropic maintains a policy where its best models are only available through paid subscriptions.

Pricing

Opus 5 is priced at $5 per million input tokens and $25 per million output tokens via the API, matching the price of Opus 4.8. For comparison, this is less expensive than the Fable 5 model. When the Fast mode is enabled, the cost doubles, and this mode is only available through the Claude API. For individual users, token pricing is less relevant since payment is based on subscription rather than per-token usage.

Automatic Internal Reasoning

Opus 5 now has internal reasoning enabled by default, a feature that previously required separate activation. The model independently determines the depth of task analysis. There are five reasoning levels available: low, medium, high, xhigh, and max, with high set as the default. Using low or medium is recommended to reduce token consumption. Enabling this feature leads to longer responses and additional model verification, which may use up a user's quota more quickly.

Context and Data Processing

Opus 5 supports a context window of one million tokens, which is both the standard and the maximum value. No smaller options are provided. Output text can reach up to 128,000 tokens. This allows for processing lengthy documents—such as contracts, manuals, or large project folders—without the need to split them.

Safety and Filtering

Opus 5 is described as Anthropic’s most refined model, with minimal erroneous behavior. At the same time, safety restrictions have been relaxed: cybersecurity filters trigger about 85% less often than in Fable 5. If a request is blocked, the older Opus 4.8 version is automatically used. In tests involving the creation of malicious code, Opus 5 is less robust than the Mythos 5 model.

Comparisons and Independent Evaluations

All reported metrics are based on Anthropic’s internal measurements. Independent comparisons with models like OpenAI’s GPT-5.6 and Chinese counterparts have not yet been published. Opus 4.8 remains available across all platforms.