Anthropic releases Claude Opus 4.8, emphasizing honesty and reduced hallucination over raw capability gains

Anthropic shipped Claude Opus 4.8 today. The model achieved the lowest incorrect-rate on every benchmark among six models tested, primarily by abstaining on uncertain questions rather than answering more questions correctly.

Claude Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked. The model now accepts mid-conversation system messages (updated instructions injected after a user turn), which preserves prompt cache hits and reduces input cost on agentic loops (AI systems that autonomously execute tasks). The minimum cacheable prompt length dropped to 1,024 tokens, down from 4,096 tokens on Claude Opus 4.7.

Pricing remains $5 per million input tokens and $25 per million output tokens—unchanged from Opus 4.5/4.6/4.7. Context window is 1,000,000 tokens with a max output of 128,000 tokens. Knowledge cutoff and training data cutoff are both January 2026.