XNews

Anthropic: Claude Opus 4.5, Agent Skills & Advanced Tool Use

Anthropic: Claude Opus 4.5, Agent Skills & Advanced Tool Use

Anthropic announced Claude Opus 4.5, positioned for coding and agent tasks with major improvements in efficiency, tool use, and pricing. • Opus 4.5 targets coding, long-running agents and complex tool workflows. • SWE-bench Verified score: 80.9% (reported lead over Gemini 3 Pro and GPT-5.1 codex-max). • OSWorld agent/computer-use score: 66.3%. • Token consumption reductions claimed: up to 65% in long coding tasks. • New 'advanced tool use' features: Tool Search Tool (deferred loading), Programmatic Tool Calling (code orchestration), Tool Use Examples (input examples for clarity). • Tool Search Tool reduced token usage from ~77K to ~8.7K and improved tool-selection accuracy (e.g., Opus 4 accuracy from 49%→74%). • Programmatic calls (Python execution sandbox) reduced tokens by ~37% and lowered latency for multi-step workflows. • Tool Use Examples raised complex-parameter accuracy from 72% to 90%. • Pricing change: $5 per 1M input tokens; $25 per 1M output tokens (roughly one-third of Opus 4.1 pricing). • Anthropic's 'skill' architecture lets developers bundle prompt-driven domain behaviors as SKILL.md packages with allowed tools, discovery and hot-reload.

Key insights

Opus 4.5 delivers major coding and agent performance gains

Anthropic released Opus 4.5 with reported benchmarks showing high software engineering and agent capabilities: SWE-bench Verified at 80.9% and OSWorld agent/computer-use at 66.3%. The model emphasizes token-efficiency (up to 65% reduction in long coding tasks) and mixed-mode 'effort' parameters to trade latency and cost. Anthropic positions Opus 4.5 for professional knowledge workers and heavy agent workflows.

Advanced tool use introduces deferred loading and programmatic orchestration

Anthropic described 'advanced tool use' comprising a Tool Search Tool (deferred tool loading), Programmatic Tool Calling (model emits sandboxed code to coordinate multi-tool actions), and Tool Use Examples (concrete input examples). Deferred loading cut token consumption dramatically (reported from 77K to 8.7K), while programmatic calls reduced token usage ~37% and improved latency and accuracy on multi-API tasks.

Skill system: dynamic prompt-driven capabilities with controlled tool permissions

Anthropic's 'skill' system packages domain expertise as prompt-based SKILL.md folders containing YAML metadata (name, description, allowed-tools, model) and detailed markdown instructions. A Skill meta-tool enumerates available skills, and Claude dynamically injects the SKILL.md content into the conversation upon invocation. Skills enable scoped tool pre-authorization, hot updates, and do not require direct code execution privileges.

Pricing and platform availability broaden enterprise access

Opus 4.5 pricing set at $5 per 1M input tokens and $25 per 1M output tokens (approx. one-third of Opus 4.1), and the model is available across Claude apps, APIs, and major cloud providers (Amazon Bedrock, Google Vertex AI, Microsoft Azure). Anthropic reports internal benchmarks showing high productivity for software engineers and enterprise workflows.

Source posts