Opus 4.5 dropped.

Anthropic released Opus 4.5 today, and one of the notable takeaways is the resistance to prompt injection.

In Anthropic's published tests, Opus 4.5 had the lowest attack success rate across major thinking models.

Prompt-injection is one of the biggest practical issues for real-world LLM adoption. It's where a user or attacker attempts to manipulate a model into ignoring its instructions or leaking information.

Despite persistent vulnerabilities in most models under pressure, Opus 4.5 achieved a meaningful drop in failure rates. This advancement improves safety, reliability, and adoption, especially for handling sensitive information.

The improvements come from better training and evaluation methods, not size. The release also comes with new integrations for Chrome and Excel, and as the product suite expands, the durability should continue to expand alongside it.