Opus 5.5 四个破坏性改动，Jev 帮你自动调参
作者列出 Opus 5.5 会 400 报错的四处 API 变更（thinking 常开、禁用强制工具调用等），并称 Jev 按轮次路由 effort，避免模型切换导致缓存失效多花 31% 的钱。
Jev + Opus 5.5: Anthropic's new model outscores GPT-6 Astra by 8.5 points, and 4 quiet API changes will 400 your agent before it ships a single line fix these before you touch the model ID: step 1 → 66.4% on Terminal-Bench 4.0 vs GPT-6 Astra 57.9% and Opus 5 52.3%. +14.1 points in one release step 2 → $4 / $20 per 1M, down from $5 / $25. cache reads drop to $0.20. at default effort it beats Opus 5 at max effort for ~1/5 the cost step 3 → thinking is always on now. send thinking: disabled and the call fails with a 400. remove the field and control depth with effort step 4 → forced tool use is gone. tool_choice any or tool returns a 400. use auto and tell the model when the tool applies step 5 → rewriting history breaks thinking. edit system, tools or earlier messages and you get a 400. append new messages instead, or turn on drop_block step 6 → the old computer-use tool is rejected on the API and Google Cloud. swap computer_20251124 for computer_toolset_20260801 step 7 → hop Opus 5.5 → Sonnet 5 → Opus 5.5 and you pay 4.36 instead of 3.32. +31%, cold cache, and Sonnet can't read Opus's reasoning. Jev routes effort per turn instead, so the cache stays warm 4 breaking changes. 3 knobs. Jev turns them for you full breakdown in the article below ↓
这条是引用着另一条写的：「长文：如何搭好 Jev 驱动的 agent 技术栈」