Anthropic Releases Claude Opus 5 With Enhanced Agentic Capabilities

Anthropic released Claude Opus 5 this week, positioning it as the new flagship model for its Opus tier and a successor to Claude Opus 4.8. The company has maintained the existing pricing structure for Opus, which remains at $5 per million input tokens and $25 per million output tokens. Anthropic states that Claude Opus 5 approaches the intelligence of Claude Fable 5 but at half the cost, and it is now the default model for Claude Max and the most capable model available on Claude Pro.

A significant change in Claude Opus 5 is that its "thinking" capability is now enabled by default for all requests. Previously, on Claude Opus 4.8, users had to explicitly enable thinking by setting `thinking: {"type": "adaptive"}`. With Opus 5, the `effort` parameter controls the depth of this thinking process. This change necessitates a review of existing `max_tokens` settings, as they now cap both the thinking process and the response text. A breaking change has also been implemented: setting `thinking: {"type": "disabled"}` with `effort` set to `xhigh` or `max` will now result in a 400 error. Developers are advised to either cap the effort at `high` or remove the `thinking` field entirely for requests.

Anthropic has also instructed developers to remove verification prompts from their instructions. The company notes that with Opus 5's inherent self-verification capabilities, explicit instructions for a final verification step can lead to over-verification. The updated prompting guide for Opus 5 offers details on tuning patterns for this new behavior. The model ID for this release is `claude-opus-5`.

The context window for Claude Opus 5 is set at 1 million tokens, serving as both the default and maximum capacity, with no smaller variants available. The synchronous Messages API supports a maximum output of 128,000 tokens. For higher throughput, the Message Batches API can achieve up to 300,000 tokens, accessible via the `output-300k-2026-03-24` beta header. Additionally, the minimum cacheable prompt size has been reduced from 1,024 tokens to 512 tokens.