Claude Opus 5.5 Launches: 30% Faster, 40% Cheaper, Blocks AI Brain Copying

On September 23, 2026, Anthropic released Claude Opus 5.5. This upgrade targets what developers care about even more than benchmark scores: cost and security.

Cheaper, Faster, Same Quality

Opus 5.5 matches the performance of Claude Fable 5.1 on most workplace tasks, but running costs drop 40% compared to Opus 5. The new pricing is $4 per million input tokens and $20 per million output tokens, with cached reads 60% cheaper. Token output speed is up more than 30%, the context window reaches 1 million tokens, and maximum output is 128K tokens. The model ships with adaptive thinking enabled by default at a medium level, so you do not need to turn on extended reasoning mode separately.

Preserved Thinking: Locking Down the AI's Brain

Preserved thinking is an anti-distillation mechanism that first appeared in Fable 5.1 and now ships in Opus 5.5 too. It encrypts Claude's thinking blocks: a block can only be reused when the system prompt, tools, and prior messages stay exactly the same. Change even one detail and the API returns a 400 error. This blocks attackers who tweak the context to extract Claude's reasoning chain, then use it to train another model that inherits the capability without inheriting the safety guardrails, such as refusing cyberattack instructions or weapons guidance.

Available Everywhere From Day One

Opus 5.5 launched simultaneously on the Claude API, AWS, Google Cloud, and Microsoft Azure, available to API accounts created after August 31, 2026. If you are building coding agents or enterprise applications, this upgrade is worth trying: prices are sharply lower, and the model's IP is protected more tightly.

Are you planning to move your agents to Opus 5.5? Let us know in the comments.