Briefing

Claude Opus 4.8 Released – Minor Improvements, Major Honesty Boost

ai-dev
Claude Anthropic

Test mid‑conversation system messages and the new 1,024‑token prompt cache minimum; update your SDK to the latest commit and adjust caching logic accordingly.

What to do now

Update your Anthropic Python SDK to the latest commit, test mid‑conversation system messages, and adjust your prompt caching logic to use the new 1,024‑token minimum.

Summary

Claude Opus 4.8 was released today, building on 4.7 with modest but tangible improvements. The release emphasizes honesty, with evaluations showing a four‑fold reduction in hallucinations and more frequent uncertainty flags. Pricing remains $5/million input and $25/million output, but fast mode has halved its cost to $10/million input and $50/million output, a significant drop from the $30/$150 of 4.6/4.7. The model keeps a 1,000,000‑token context window and 128,000‑token maximum output, with knowledge and training cutoffs set to January 2026.

New API features include mid‑conversation system messages that can be inserted after a user turn, allowing updated instructions without re‑sending the full system prompt and preserving prompt cache hits. The prompt cache minimum has been lowered from 4,096 to 1,024 tokens, reducing input cost for agentic loops. Fast mode is now limited to research preview accounts, requiring account‑manager approval. The Anthropic Python SDK has been updated to support these changes, and developers should migrate to the latest commit.

Key changes

  • Mid‑conversation system messages accepted after user turns, enabling updated instructions without re‑sending the full system prompt
  • Prompt cache minimum reduced from 4,096 to 1,024 tokens, lowering input cost
  • Fast mode pricing halved to $10/million input and $50/million output, a 50% drop from previous $30/$150
  • Honesty improvements: 4× fewer hallucinations, more uncertainty flags, per system card evaluation
  • Context window unchanged at 1,000,000 tokens; max output 128,000 tokens
  • Knowledge cutoff and training data cutoff both set to January 2026
  • Same pricing as Opus 4.5‑4.7 for standard mode
  • Fast mode only available to research preview accounts; contact account manager for access

Affects

internal

Customer impact

Analyzing matches…

Ask about this story

Impact on an agency? Which customers? Compare historically Risks of waiting