Anthropic Claudev4

22 May 2025

Claude 4 — Opus 4, Sonnet 4.6, Extended Thinking

Anthropic's Claude 4 family ships Opus 4 as the strongest reasoning model available, Sonnet 4.6 as the default production workhorse, and introduces Extended Thinking — allowing the model to reason through hard problems with a scratchpad before answering.

• 1Claude Opus 4 — top-tier reasoning and coding, achieves 72.5% on SWE-bench (real-world software engineering tasks), best-in-class for agentic long-context tasks
• 2Claude Sonnet 4.6 — daily production default; 2× faster than Opus 4 with 90% of the quality at lower cost
• 3Claude Haiku 4.5 — ultra-low latency, 3× cheaper than Haiku 3, suitable for high-volume classification and triage
• 4Extended Thinking — models expose reasoning tokens (the scratchpad) in the API response, enabling verifiable chain-of-thought
• 5200k context window across all Claude 4 models with improved instruction following at long contexts
• 6Computer use (beta) improved — lower hallucination on UI coordinates, higher task completion rate
• 7Native tool use streaming — results stream token-by-token instead of arriving as a single batch
• 8Prompt caching extended to 1-hour TTL for large system prompts — cost reduction up to 90% for high-reuse workloads