Claude Opus 4.8: "a modest but tangible improvement"

2026-05-28 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, short

Summary

Anthropic shipped Claude Opus 4.8 on May 28, 2026, describing it as a "modest but tangible improvement" over its predecessor, with a primary focus on enhanced "honesty." Evaluations indicate Opus 4.8 is four times less likely to overlook code flaws and exhibits the lowest incorrect-rate across benchmarks, largely by abstaining when uncertain rather than answering more questions correctly. Pricing remains \$5/million input and \$25/million output, but "Fast mode" for 4.8 is now \$10/million input and \$50/million output, a significant reduction from the \$30/\$150 for previous models. The model retains a January 2026 knowledge cutoff, a 1,000,000 token context window, and 128,000 max output tokens. New features include mid-conversation system messages for dynamic instruction updates and a reduced prompt cache minimum from 4,096 to 1,024 tokens, improving efficiency for agentic loops.

Key takeaway

For AI Engineers building agentic systems, Claude Opus 4.8 offers tangible improvements in reliability and cost-efficiency. Its enhanced "honesty" and ability to flag uncertainties mean you can trust its outputs more, especially for code generation. Utilize mid-conversation system messages to dynamically steer agents and gain from the lower prompt cache minimum, reducing input costs on long-running tasks. Consider "Fast mode" for applicable use cases to further optimize expenses.

Key insights

Claude Opus 4.8 prioritizes "honesty" and uncertainty flagging, improving reliability by abstaining rather than hallucinating.

Principles

• Honesty reduces unsupported claims.
• Abstaining improves factual accuracy.
• Dynamic prompts enhance agentic loops.

Method

Mid-conversation system messages allow appending updated instructions after a user turn, preserving prompt cache hits and reducing input costs in long-running agentic conversations.

In practice

• Utilize mid-conversation system messages.
• Benefit from lower prompt cache minimum.
• Test "Fast mode" for cost savings.

Topics

• Claude Opus 4.8
• LLM Honesty
• Prompt Engineering
• Agentic AI
• Cost Optimization
• Anthropic

Code references

• anthropics/anthropic-sdk-python
• simonw/llm-anthropic
• simonw/tools

Best for: CTO, VP of Engineering/Data, AI Architect, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Anthropic Shipped Claude Opus 4.8 Yesterday. — Claude Opus 4.8 enhances AI honesty and introduces Dynamic Workflows for complex, large-scale coding tasks.
• 😺 Claude Opus 4.8 got safer today — Claude Opus 4.8 balances enhanced capability with safety through effort controls and dynamic agent workflows.
• Claude's Opus 4.7 Review — Claude Opus 4.7 significantly improves agentic coding, instruction following, and multimodal understanding.
• Claude Opus 4.8 is now available in Microsoft Foundry — Claude Opus 4.8 in Microsoft Foundry enhances AI for complex coding, agentic workflows, and deep enterprise analysis.
• Anthropic ships Claude Opus 4.8 as a "modest but tangible improvement" that tops GPT-5.5 in most benchmarks — Claude Opus 4.8 improves benchmarks, honesty, and introduces dynamic workflows and user effort controls.
• Claude Opus 4.8: The System Card — Claude Opus 4.8 improves intelligence and honesty but regresses in prompt injection and computer use safety due to training changes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.