Anthropic Unveils Claude Opus 5.5 with Enhanced Performance and 40% Lower Operational Costs

The new model demonstrates major improvements in agentic tasks, including coding and knowledge work. In early testing, the model completed a 680,000-line code migration in less than a day, a task estimated to take an engineering team weeks. It also showed high proficiency in diverse benchmarks, such as Terminal-Bench 4.0, where it recorded a score of 66.4%, surpassing both Claude Opus 5 and OpenAI's GPT-6 Astra.

Efficiency is a core feature of the update. Anthropic states that Claude Opus 5.5 requires significantly less compute, resulting in a 40% reduction in total cost for typical workloads compared to Opus 5. Specific pricing for the API is set at $4 per 1 million input tokens and $20 per 1 million output tokens. Additionally, cache read costs have been reduced by 60% to $0.20 per million tokens. For high-speed requirements, a "Fast mode" is available via Claude Code and the Claude Platform, offering up to 2.5x speed.

Safety features have been bolstered to match the standards of the advanced Claude Fable 5.1. This includes implementing safeguards for cybersecurity and biology, where high-risk requests may be re-routed to more specialized or restricted models like Claude Opus 4.8. The model also introduces "preserved thinking" to defend against distillation attacks—a method where attackers attempt to extract a model's reasoning through repeated queries. This feature prevents API users from editing the prior context to harvest reasoning information.

Claude Opus 5.5 is currently available on all major platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic also announced that Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to be released within the coming weeks.

PLUS ULTRAby Amenoyomi

The mentioned "agentic tasks" represent a transition where the AI does not merely suggest code or answer a prompt, but autonomously manages the entire lifecycle of a complex project.

In practical terms, this allows for the completion of large-scale engineering tasks without constant human intervention. For instance, early testers completed a 680,000-line code migration in less than a day—work that previously required weeks from an engineering team. Other users reported the model autonomously investigating, designing, and implementing fixes for service-layer bugs overnight, shifting the human role from manual coding to the review of completed results.

This autonomy is enabled by a shift in economic efficiency, particularly in how the model handles repeated context. Cache read costs have been reduced by 60% to $0.20 per million tokens, which is critical because cache reads constitute the majority of the cost in agentic workflows where the AI must repeatedly access the same codebase or documentation to verify its own work.