XNews

Anthropic / Claude: Opus 4.5 release and advanced tool-use features

Anthropic / Claude: Opus 4.5 release and advanced tool-use features

Anthropic released Opus 4.5 and an advanced tool-use framework for agent scalability and efficiency.

Key insights

Claude Opus 4.5: performance, efficiency and price cuts

Anthropic announced Claude Opus 4.5, reporting SWE-bench Verified 80.9% and improvements in coding, agent and computer-use tasks. Opus 4.5 reportedly reduces token consumption by up to 65% on long coding tasks, supports 'effort' API parameter to balance compute/latency, and launched at substantially lower pricing (input token $5/M, output token $25/M) compared with Opus 4.1.

Advanced tool-use reduces context costs and improves accuracy

Anthropic's engineering blog describes three innovations—Tool Search Tool (deferred loading), Programmatic Tool Calling (code-generated multi-tool orchestration) and Tool Use Examples (call templates)—that collectively reduce token usage, speed up calls, and increase tool-calling accuracy. Internal tests reported token savings and accuracy improvements across MCP and GIA benchmarks.

Skills/Agent architecture emphasizes prompt-driven specializations

Claude's 'skills' system is a prompt-driven meta-tool architecture: SKILL.md packages include YAML metadata and Markdown instructions which are injected at runtime to reconfigure Claude into domain specialists, with pre-authorized minimal tool permissions and explicit step instructions. The skills discovery mechanism dynamically lists available skills and injects them into conversation without code execution, supporting hot updates and constrained privileges.

Market reaction reflects pragmatism over hype

Observers note Claude Opus 4.5 generated modest public excitement compared to earlier model launches, signaling an industry shift from model-race enthusiasm to ROI and application-driven adoption; users now prioritize stability, cost-effectiveness and real-task productivity over raw benchmark supremacy.

Source posts