Claude Sonnet 与 Opus 2026：停止为错误的模型支付过高费用
📄 中文摘要
在行业标准编码基准 SWE-bench Verified 中，Claude Sonnet 4.6 的得分为 79.6%，而 Opus 4.6 的得分为 80.8%，两者之间仅相差 1.2 个百分点。然而，Opus 的每个 token 成本大约高出 67%。在实际使用中，用户在 59% 的情况下更倾向于选择 Sonnet 4.6 而非 Opus 4.5。许多开发者为微小的性能提升支付了额外的费用，可能并未察觉到这些差异。尽管如此，Opus 在特定任务上确实表现更佳。该指南详细分析了何时值得支付额外费用，何时则是在浪费金钱。
📄 English Summary
Claude Sonnet vs Opus 2026: Stop Overpaying for the Wrong Model
On the SWE-bench Verified coding benchmark, Claude Sonnet 4.6 scores 79.6%, while Opus 4.6 scores 80.8%, reflecting a mere 1.2 percentage point difference. However, Opus costs approximately 67% more per token. In practical usage, users preferred Sonnet 4.6 over Opus 4.5 a significant 59% of the time, indicating that many developers are paying a premium for negligible gains. Nevertheless, Opus does excel in specific tasks. This guide provides a detailed breakdown of when the premium is justified and when spending extra money is unnecessary.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等