Anthropic: Claude Opus 4.5 and advanced tool-use

Anthropic released Claude Opus 4.5 and an Advanced Tool Use suite for large tool sets and lower token costs. - Claude Opus 4.5 aims at coding, agent reliability and computer use - SWE-bench Verified: 80.9%, first Opus >80 - OSWorld computer-use score: 66.3% - Token consumption reductions up to 65% in long tasks - Pricing: $5 per 1M input tokens, $25 per 1M output tokens (≈1/3 of Opus 4.1 rates) - Opus 4.5 available on Claude apps, API and major clouds (AWS Bedrock, Google Vertex, Azure) - Advanced Tool Use includes Tool Search Tool, Programmatic Tool Calling, Tool Use Examples - Tool Search Tool delays loading to save tokens (reported token savings ~85%) - Programmatic calls output Python for sandboxed multi-tool coordination and parallelism - Tool Use Examples raise parameter/format accuracy to ~90% in complex calls
Key insights
Claude Opus 4.5 targets engineering and agent workloads with efficiency and price drops
Anthropic announced Claude Opus 4.5 with a stated focus on software engineering, agent reliability and multi-system computer use. Internal and public benchmarks report SWE-bench Verified at 80.9% and OSWorld at 66.3%. The release emphasizes large token-efficiency gains — up to 65% token reduction on long coding tasks — and a pricing reset to $5 per 1M input tokens and $25 per 1M output tokens, roughly one-third of Opus 4.1 costs. Anthropic made the model available across Claude apps, public API and major cloud marketplaces.
Advanced Tool Use adds dynamic loading, programmatic calling and examples
Anthropic published an engineering blog explaining Advanced Tool Use: a Tool Search Tool (a meta-tool for on-demand loading), Programmatic Tool Calling (models emit sandboxed Python to coordinate multi-tool workflows) and Tool Use Examples (structured examples clarifying complex parameter formats). The architecture defers loading of many tools to save token budget, uses code orchestration to reduce multi-round reasoning, and attaches examples to increase parameter-accuracy for nested inputs.
Dynamic tool loading significantly reduces token usage and boosts accuracy
Anthropic reported that marking many tools as defer_loading: true and using the Tool Search Tool reduced token consumption from ~77K to ~8.7K in experiments and increased tool selection accuracy (e.g., Claude Opus 4 from 49% to 74%). Programmatic calling reduced tokens and latency, while Tool Use Examples raised complex-parameter accuracy from 72% to ~90% in internal tests.
Opus 4.5 positions Anthropic competitively on coding and agent benchmarks
Anthropic framed Opus 4.5 as a high-end model for professional software engineers and knowledge workers, noting bench strengths and cross-cloud availability. Commentary positioned Opus 4.5 as competitive with Gemini 3 Pro and OpenAI models on coding and long-horizon agent tasks, with a pricing strategy intended to broaden adoption for daily use-cases.