Anthropic released Opus 4.8 just six weeks after Opus 4.7, without changing standard prices: $5 per million input tokens and $25 per million output tokens. Opus 4.8 is faster, smarter, and introduces a fast mode that triples previous speed at reduced cost, though still much more expensive than Chinese competitors. On key benchmarks, it outperforms OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro: 69.2% on SWE-bench Pro (software engineering tasks), 49.8–57.9% on Humanity’s Last Exam (academic questions), and 83.4% on OSWorld-Verified (real-world software use). Its only relative weakness is Terminal-Bench 2.1 (command-line tasks), where it trails GPT-5.5. New user controls allow adjustment of model effort per task, balancing cost and accuracy. Opus 4.8 uses more tokens per task, potentially increasing costs versus Anthropic’s less-capable Claude Sonnet. Alignment and safety have improved: lower deception and misuse rates, better bug detection, and prosocial behaviors. Dynamic workflows, now in research preview, enable Claude to orchestrate and verify subagent tasks in one session. Despite Anthropic’s superior benchmarks and safety, DeepSeek V4 Pro and Xiaomi MiMo V2.5 Pro offer dramatically lower prices, making Opus 4.8 primarily attractive for regulated or safety-critical industries.