Anthropic officially launches on May 28Claude Opus 4.8, which is the third major upgrade for its flagship model in 2026 (after Opus 4.6 in February and Opus 4.7 in April).Honesty optimization, dynamic workflow and agent programming capabilitiesSignificant breakthroughs have been achieved in three dimensions.
core competencies
Opus 4.8 continuesClaudeThe core advantages of the series——1 million token context window(available by default), maximum output of 128,000 tokens, and passClaude The multi-platform deployment capabilities of API, Amazon Bedrock and Vertex AI also bring three key upgrades:
- Uncertainty Flagging: Models are now able to proactively label content they are uncertain about, significantly reducing "hallucinations" and overconfidence errors in long-term agent tasks - critical for enterprise-level applications
- Dynamic Workflows:Claude Code users can get a research preview version of the multi-step agent workflow. The model can autonomously plan and execute complex coding tasks step by step instead of simple one-time code generation.
- Effort Controls: Users can switch between "Quick Mode" and "Deep Thinking Mode". The fast mode has lower cost and shorter delay, while the deep thinking mode is suitable for tasks that require precise reasoning.
In terms of benchmark testing, Opus 4.8 comprehensively surpassed the previous generation Opus 4.7 (agent programming increased from 64.3% to 69.2%), and led GPT-5.5 and Gemini 3.1 Pro in many key indicators.Pricing remains unchanged——Input $5/million tokens and output $25/million tokens, exactly the same as Opus 4.7.
User experience and limitations
Claude The biggest highlight of Opus 4.8 is that“Know what you don’t know”.
Dynamic workflows are currently only available forClaude Code’s Enterprise, Team and Max plan users are open to research preview, and ordinary users are temporarily unable to experience it.Chinese proficiency is stillClaudeShortcomings of the series——Although the basic Chinese conversation is fluent, when it comes to Chinese technical documentation, code comments, business scenario understanding and other tasks, it is difficult to compare with domestic models (such asDeepSeek, Qwen) there is still a gap.
Another limitation worth noting is:Claude The Mythos (Anthropic's next-generation flagship model) remains in restricted access.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 9.0 / 10 | 1 million token context + multi-platform deployment + dynamic workflow, comprehensive functions |
| 易用性 | 8.5 / 10 | Quick/deep mode switching is intuitive, but the dynamic workflow experience threshold is high |
| Cost-effectiveness | 7.5 / 10 | The output of $25/M token is high, but the input of $5/M token is reasonable in the flagship model. |
| 中文支持 | 7.0 / 10 | Basic Chinese is available, but understanding of technical documentation and business scenarios is weaker than that of domestic models |
| 输出质量 | 9.2 / 10 | The Honesty Optimization Effect is Remarkable, with 69.2% of Agent Programming Leading the Industry |
Overall rating: 8.7/10
Claude Opus 4.8 is currently one of the preferred models for enterprise-level AI coding and agent tasks. AI Dash Learn more about AI tool reviews.