Opus 4.8 scored 81 in my benchmark. I still wouldn’t default to it. (The full breakdown + Nate’s Community Slack) Claude Opus 4.8 achieved the highest score of 81 in the current benchmark suite, outperforming competitors like GPT-5.5. However, it did not win every task, showing weaknesses in visual and front-end areas, and sometimes performing worse on long-horizon tasks when using maximum effort. The article emphasizes that model selection depends on specific workflow needs, such as task duration, source material requirements, and the need for human oversight, rather than simply seeking the ‘smartest’ model.

• Claude Opus 4.8 leads the benchmark suite with a score of 81, surpassing GPT-5.5 (71) and other models.
• Opus 4.8 excels in areas like source discipline, operational judgment, and self-correction, which are critical for professional AI output.
• Despite its high score, Opus 4.8 has weaknesses, including visual and front-end issues, and was outperformed by other models on specific tasks like the Artemis visualization.
• The article questions the ‘smartest model’ approach, suggesting that effective model choice requires considering factors like work type, task length, source material needs, tool usage, and error handling.
• The author will not treat Opus 4.8 as a universal replacement but will use it aggressively for specific tasks and provide guidance on choosing between Opus 4.8, Codex/5.5, and GPT-5.5.
• The article promises a breakdown of tests, discussion on the ‘effort-level trap,’ guidance on choosing daily tools, and role-specific advice for builders, leaders, and executives. Continue reading https://natesnewsletter.substack.com/p/opus-48-benchmark-model-selection