Anthropic: Claude / Opus 4.5, agent tooling and skills

Anthropic announced Claude Opus 4.5, Opus‑level coding and agent improvements, and a developer platform feature set for large tool ecosystems. - Opus 4.5 reaches 80.9% on SWE‑bench Verified and beats leading coding models in reported internal tests. - Opus 4.5 reduces token usage up to 65% and cuts pricing to $5/1M input, $25/1M output (one‑third prior cost). - Opus 4.5 introduces mixed inference modes and improved memory for long contexts and agent orchestration. - Anthropic published an engineering blog on 'advanced tool use' (tool search tool, programmatic calling, examples) to handle hundreds/thousands of tools with token and accuracy gains. - Tool Search Tool deferred loading saved >85% token usage and improved tool selection accuracy (Opus 4->4.5 examples). - Programmatic Tool Calling enables sandboxed code orchestration to reduce rounds and token use. - Claude 'skills' architecture is a prompt‑driven, modular system where SKILL.md injects domain rules and allowed tools; skills act as context rewrite rather than external code execution. - Skill files include YAML metadata and Markdown instructions; skill discovery and dynamic injection enable safe, hot‑updated domain experts.
Key insights
Opus 4.5 targets coding and agent use with large efficiency gains
Anthropic's Opus 4.5 reports an 80.9% score on SWE‑bench Verified, claims improved long‑context memory, mixed inference modes and token efficiency (up to 65% reduction in long tasks). Anthropic published pricing at $5/1M input and $25/1M output, representing a two‑thirds price cut versus Opus 4.1. The release positions Opus 4.5 for professional knowledge work (software engineering, financial analysis) and for use as master agent orchestration in multi‑agent workflows. (supporting posts: [13538,13754])
Advanced tool use reduces token cost and improves tool selection
Anthropic's engineering blog details three techniques—Tool Search Tool (deferred loading), Programmatic Tool Calling (sandboxed code orchestration), and Tool Use Examples—that together cut token consumption dramatically and increased tool call accuracy. Deferred loading converted 77K token tool lists down to ~8.7K and raised tool selection accuracy (example: Opus 4 from 49% to 74%). Programmatic calling reduces multi‑round reasoning and token shuffling by returning structured results. (supporting_posts: [13534])
Claude 'skills' are prompt‑driven, safe domain plugins
Claude's 'skills' system packages domain behavior as a SKILL.md file with YAML metadata, allowed‑tool lists and Markdown instructions. Skills operate by injecting high‑priority system prompts into the running conversation, altering Claude's behavior and tool permissions without granting direct code execution. Discovery runs via a 'Skill' meta‑tool that scans available skills and injects the matching skill at runtime. The design prioritizes least privilege, hot updates, and safety by keeping logic in prompts rather than external execution. (supporting_posts: [13316])
Market reaction muted: practical ROI now governs excitement
Observers reported muted public excitement after Opus 4.5 despite strong benchmarks, arguing the industry has shifted from model‑worship to pragmatism: buyers now focus on task ROI, cost, stability and tooling integration rather than raw peak metrics. The commentary notes that developer‑facing features and applications that demonstrate clear productivity gains generate more visible user enthusiasm than model leaderboards. (supporting_posts: [13754,13538])
Opus 4.5 targets enterprise toolchains and cloud availability
Anthropic made Opus 4.5 available via Claude app, API and through major cloud platforms (Amazon Bedrock, Google Vertex AI, Microsoft Azure) with model identifier claude-opus-4-5-20251101. The model introduces an 'effort parameter' for balancing latency, compute and response depth for customers. Internal tests report improved agent capabilities on multi‑system bugs and long tasks. (supporting_posts: [13538])