换模型前怎么迁移评估How to Assess a Model Migration
微型指南 · 约 7 分钟Guide · about 7 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
换模型不是改个名字：同一套提示词在新模型上可能失效。这篇讲换之前要过的四道检查。Switching models is not renaming: the same prompt set can break on the new one. Four checks to pass before you switch.
先跑自建评测对比新旧：公共榜单只说明普遍水平，你的真实输入才说明能不能换。
逐项核接口差异：参数名、流式行为、结构化输出支持、上下文长度——兼容接口不等于行为一致。
重测依赖提示词的功能：同一提示词在不同模型上表现可以差很多，边界条款尤其容易漂。
核许可与数据条款：新模型的许可证、数据保留政策、区域可用性可能和旧模型不同。
留回滚路径：旧模型配置保留到确认期结束，出问题能一键切回。
确认期用线上指标收尾：任务完成率、拒答率、成本三项对比基线，达标才下线旧模型。
Run your own eval on old vs new first: public leaderboards show general level; only your real inputs say whether the switch works.
Diff the API surface item by item: parameter names, streaming behavior, structured-output support, context length — a compatible interface is not identical behavior.
Re-test everything that depends on prompts: the same prompt can behave very differently across models; boundary clauses drift most.
Check licenses and data terms: the new model's license, data-retention policy, and regional availability may differ from the old one.
Keep a rollback path: retain the old model's configuration until the confirmation window closes, so you can switch back with one click.
Close out with live metrics during the confirmation window: task completion, refusal rate, and cost against baseline — retire the old model only when they pass.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
先保存旧配置，再用固定输入比较新旧模型的事实、格式、工具调用与成本。提前约定什么情况回滚，并验证旧配置仍可恢复。迁移完成的依据是任务验收，而不是新接口返回了 200。Save the old configuration and compare facts, formats, tool calls and cost on fixed inputs. Define rollback triggers and verify recoverability. Task acceptance, not an HTTP 200, establishes a completed migration.
迁移评估Migration Evaluation 模型版本与变体Model Family, Version, Variant OpenAI 兼容接口OpenAI-compatible API 许可证License