我实际编码三天后，在真实会话中对比 Opus 4.7 和 4.6 的表现
1 分•作者: agentseal•3 个月前
今天我花了一些时间，使用我自己的使用数据比较了 Opus 4.6 和 4.7，看看它们并排的表现如何。
虽然 4.7 还在早期阶段，但有些事情让我感到惊讶。
在我的会话中，4.7 第一次就做对的次数比 4.6 少。一次成功率约为 74.5%，而 4.6 为 83.8%，而且我看到每次编辑的重试次数大约翻了一番（0.46 次 vs 0.22 次）。
它每次调用的输出也多得多，大约 800 个 token，而 4.6 为 372 个，这使得它明显更贵。每次调用的成本是 0.185 美元，而 4.6 是 0.112 美元。
当我按任务类型细分时，编码和调试在 4.7 上都表现较弱。编码一次成功率从 84.7% 下降到 75.4%，调试从 85.3% 下降到 76.5%。功能开发在 4.7 上略有改善（75% vs 71.4%），但样本量很小。委托任务显示出很大差距（100% vs 33.3%），尽管 4.7 方面只有 3 个样本，所以我不会太在意这一点。
4.7 每次轮次使用的工具也更少（1.83 vs 2.77），并且几乎不委托给子代理（0.6% vs 3.1%）。目前还不确定这是一种风格差异，还是仅仅因为样本量较小。
有几点需要说明。这是大约 3 天的 4.7 数据（3,592 次调用）与 8 天的 4.6 数据（8,020 次调用）的对比。某些类别只有少数例子。这些数字会随着更多使用而变化，并且您的结果可能会因您所做的工作类型而异。
npx codeburn compare
查看原文
I spent some time today comparing Opus 4.6 and 4.7 using my own usage data to see how they actually behave side by side.<p>still pretty early for 4.7, but a few things surprised me.<p>In my sessions, 4.7 gets things right on the first try less often than 4.6. One-shot rate sits around 74.5% vs 83.8%, and I am seeing roughly double the retries per edit (0.46 vs 0.22).<p>It also produces a lot more output per call, about 800 tokens vs 372 on 4.6, which makes it noticeably more expensive. cost per call is $0.185 vs $0.112.<p>when I broke it down by task type, coding and debugging both looked weaker on 4.7. Coding one-shot dropped from 84.7% to 75.4%, debugging from 85.3% to 76.5%. Feature work was slightly better on 4.7 (75% vs 71.4%), but the sample is small. Delegation showed a big gap (100% vs 33.3%), though that one only has 3 samples on the 4.7 side so I wouldnt read much into it yet.<p>4.7 also uses fewer tools per turn (1.83 vs 2.77) and barely delegates to subagents (0.6% vs 3.1%). Not sure yet if that's a style difference or just the smaller sample.<p>A couple of caveats. This is about 3 days of 4.7 data (3,592 calls) vs 8 days of 4.6 (8,020 calls). Some categories only have a handful of examples. These numbers will shift with more usage, and your results will probably look different depending on what kind of work you do.<p>npx codeburn compare