Claude Opus 4.8 Released

Anthropic released Claude Opus 4.8. This update focuses on practical, incremental improvements rather than a dramatic leap in performance over the previous model.

Key Improvements

• Improved Honesty: The model better recognizes when it is uncertain about a question and withholds an answer or unsupported claims less often. In particular, rather than raising the correct-answer rate on benchmarks, it reduces hallucinations by abstaining from answering uncertain questions.
• Code Accuracy: Compared to the previous model, the rate at which it fails to notice flaws in the code it writes decreased by about 4x.

Technical Features and API Updates

• Mid-conversation system messages: You can now add a role: "system" message mid-conversation, allowing instructions to be updated without resending the entire prompt. This improves Prompt Caching efficiency and reduces cost.
• Reduced Minimum Prompt Caching Unit: The minimum cacheable prompt length has been lowered from 4,096 tokens to 1,024 tokens.
• Model Specifications: Context window of 1,000,000 tokens, maximum output of 128,000 tokens, knowledge cutoff of January 2026.

Pricing Pricing remains the same as the existing Opus model ($5/M input, $25/M output). Notably, Fast mode pricing has been significantly cut from the previous $30/$150 to $10/$50.