XNews

Anthropic / Claude Opus 4.5 Released with Advanced Tool Usage Features

Anthropic / Claude Opus 4.5 Released with Advanced Tool Usage Features

Anthropic has released Claude Opus 4.5, positioned as a high-end coding and agent model, focusing on performance, efficiency, and advanced tool usage.

Key insights

Opus 4.5 Leads in Coding and Agent Benchmarks

Anthropic announced that Opus 4.5 scored 80.9% on the SWE-bench Verified coding benchmark and achieved high scores in internal agent and computer usage tests (OSWorld), demonstrating long-term goal-oriented behavior and the ability to resolve complex multi-system bugs. Internal evaluations show that Opus 4.5 reduces token usage by up to 65% when handling long-term coding tasks and can accomplish tasks that previous versions could not.

Pricing Revolution Makes High-End Capabilities More Affordable

Opus 4.5 has adjusted its pricing to $5 per million input tokens and $25 per million output tokens, down to one-third of the previous Opus 4.1 pricing ($15/$75), making Opus-level capabilities more widely affordable for the first time and available for developers on cloud platforms (Bedrock/Vertex/Azure).

Advanced Tool Usage Module Reduces Context Costs and Enhances Tool Call Accuracy

Anthropic's engineering blog introduced three key technologies: Tool Search Tool (on-demand delayed loading of tools), Programmatic Tool Calling (generating and executing coordinated code), and Tool Use Examples (guidance for complex parameters). These improvements significantly reduced token consumption and latency in experiments, while greatly enhancing the accuracy of tool selection and complex parameter accuracy.

New Operational Parameters and Prompt Engineering Adjustments Needed to Avoid Over-Triggering and Over-Engineering

Opus 4.5 introduces 'effort parameters' to control the investment in reasoning calculations; at the same time, the model is more sensitive to system prompts, where old-style mandatory prompts may lead to tool over-triggering. Prompt engineering suggestions include softening tool triggers, limiting over-engineering, and requiring the model to read relevant documents in full before modifying code.

Market Response Leans Towards Pragmatism, Simple Benchmark Scores No Longer Capture Attention

Despite the significant performance improvements of Opus 4.5, the industry response post-release has been calm. Observations indicate that the industry has entered a 'pragmatism era,' where users are more concerned about whether the model can directly enhance efficiency and deliver ROI. If the model's capability improvements do not lead to significant practical applications, they are unlikely to excite the public.

Source posts