Anthropic’s Claude guide compares models, effort levels and cost
Anthropic published a model-selection guide by Addy Osmani covering Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2 across five effort levels. According to Anthropic’s reported SWE-bench Pro results, running Opus 5.5 at low effort before retrying failures at high effort achieved about 97% at $0.17 per problem, versus 95.3% at $0.29 when every attempt used high effort. The guide also recommends adding tests or other automatic outcome checks before increasing effort, and notes that several benchmark comparisons lack error ranges or fully disclose their cost calculation.
The guide gives Claude Code users a task-based framework for balancing model choice, effort and cost, while Anthropic’s vendor-reported benchmarks still need validation on users’ own workloads.
Anthropic 在 11 月 9 日发布了一份选模型指南《Which Claude model should you build with?》，作者是 Addy Osmani。指南覆盖 Haiku 5.5、Sonnet 5.5、Opus 5.5、Fable 5.2 四个模型和五个 effort 档位。
其中最反直觉的是下面这组数据。同一批 SWE-bench Pro 题目，Opus 5.5 先全部用 low 档跑，只把失败的用 high 重跑，通过率约 97%，每题约 0.17 美元。每道题都用 high，通过率是 95.3%，每题 0.29 美元。
本文依次拆解这组数据成立的前提、Opus 5.5 各档位的分数与成本、Fable 5.2 单价翻倍而任务成本减半的原因、模型不够用时官方给的三步顺序，以及在 Claude Code 里落地的具体设置。
这组数据否定了一个常见做法：为了稳妥，把档位一次开到 high 或更高。low 档单独跑只比 high 低约 8 分，成本约为三分之一。全程 high 等于给所有简单题都付了难题的价钱，最后通过率还低了约 1.7 个点。档位应该按题分配，而不是按项目分配。
先 low 后 high 的前提：结果能自动判对错
#Anthropic#Claude Code#Haiku 5.5#Sonnet 5.5#Opus 5.5#Fable 5.2