Anthropic Launches Claude Opus 5: Tops Coding Benchmarks, Reduces Safety Block Rate by 85%
机器之心 · wechat · 2026-07-25
Anthropic has officially released its next-generation model, Claude Opus 5. Maintaining previous pricing ($5/$25 per million input/output tokens), it significantly boosts performance at half the cost of Fable 5, introducing a 2.5x faster 'Fast mode'.
- Benchmarks: Sets new records in coding and agentic evaluations like Frontier-Bench. Scores three times the second-place model on ARC-AGI 3, leading at equivalent cost points in OSWorld 2.0.
- Autonomy & Research: Demonstrates strong self-correction and end-to-end engineering capabilities (e.g., writing custom CV pipelines to rebuild 3D parts, creating its own test harnesses). Shows notable improvements in life sciences and visualization.
- Safety & Alignment: Described as Anthropic's most aligned model yet. Cybersecurity guardrails are significantly relaxed, allowing source-code vulnerability detection while dropping block rates by 85% compared to Fable 5. Blocked requests automatically fall back to Opus 4.8.
Related event: Anthropic Releases Claude Opus 5(89 posts)→
More from coding & agent
- ChatGPT Voice plus Codex makes a surprisingly good hands-free coding workflow — petergyang · 2026-07-25
- LM Studio’s Bionic launches as a local-first agent for docs, coding and voice — nicolascraske · 2026-07-25
- 'Clean Code' Author Rebuts: True Engineering Beats Vibecoding — burny_tech · 2026-07-25
- AI Security Agent Achieves RCE on GitLab Default Configuration via Dependency Chain — andreamichi · 2026-07-25
- MongoDB guide maps the production stack that makes AI agents work — TheTuringPost · 2026-07-25
- A Type Proposal Would Make Inline React Arrays Fail at Build Time — aidenybai · 2026-07-25