Claude Opus 4.8: Anthropic Prioritizes Honesty Over Raw Capability Gains

Summary

Anthropic has released Claude Opus 4.8, a new version of its flagship large language model that emphasizes honesty and reduced hallucinations over raw capability improvements. In a notably transparent announcement, the company describes the release as a "modest but tangible improvement" over Opus 4.7. The key advancement centers on improved truthfulness: Opus 4.8 is approximately four times less likely than its predecessor to allow flaws in code to pass unremarked, achieving this primarily by abstaining on uncertain questions rather than making confident but unsupported claims.

Technically, Opus 4.8 maintains the same specifications as previous versions, with a 1,000,000-token context window and 128,000-token maximum output. The model introduces a valuable feature for developers: mid-conversation system messages, which allow updated instructions to be appended without restating the full prompt, preserving prompt cache efficiency. The minimum cacheable prompt length has also been reduced from 4,096 to 1,024 tokens, further improving cost efficiency for users leveraging cached interactions.

Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens, with no premium for the new version. The knowledge cutoff and training data cutoff are both January 2026, matching Opus 4.7. These design choices reflect Anthropic's apparent philosophy of focusing on safety and reliability improvements rather than constantly expanding raw model capabilities.

• Pricing remains stable with no cost increase, continuing Anthropic's deliberate strategy of incremental improvements