跳到主要内容Skip to main content

模型 API 怎么谈 SLA 与退出条款How to Negotiate Model API SLAs and Exit Terms

微型指南 · 约 9 分钟Guide · about 9 min · 引用核验 2026-11-19(仅引用链接与日期) · refs checked 2026-11-19 (reference links and dates only)

模型 API 的坑不在价格,在可用性、限流、数据权利和「想走时走不掉」。The trap in a model API is not price; it is availability, rate limits, data rights, and not being able to leave.

怎么做How to

可用性:写清月度可用率、故障响应与赔付;不要接受「尽力而为」。

限流与配额:RPM、TPM、并发、突发;超限是排队还是报错,要写进合同。

数据权利:输入输出归谁、是否训练、留存多久、能否删除;与隐私政策对齐。

模型变更:供应商换底座或改版本要提前通知;你有权做回归测试与终止。

退出:数据导出格式、迁移协助、缓存与备份清理、最后计费日。

价格:单价、阶梯、最低消费、涨价上限;不要只看首月折扣。

责任:内容侵权、数据泄露、服务中断的责任边界与保险。

Availability: write monthly uptime, incident response and credits; do not accept “best efforts”.

Rate limits and quota: RPM, TPM, concurrency, bursts; say whether overage queues or errors, and put it in the contract.

Data rights: who owns input and output, is it used for training, how long is it kept, can it be deleted; align with the privacy policy.

Model changes: the vendor must give notice before changing the base or version; you keep the right to regression-test and terminate.

Exit: export format, migration help, cache and backup deletion, final billing date.

Price: unit price, tiers, minimum spend, cap on increases; do not look only at the first-month discount.

Liability: boundaries and insurance for IP infringement, data breach and service interruption.

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.

把合同里的「99.9% 可用」换算成每月允许 43 分钟停机,再问清超限是排队还是报错、赔付怎么算;同时写一条「供应商换底座需提前 30 天通知并允许 30 天回归测试」。Translate "99.9% uptime" into 43 minutes of allowed downtime per month, then ask whether throttling queues or errors and how credits are calculated. Add a clause requiring 30 days' notice before a base-model change plus a 30-day regression window.

做完例子后怎么检查How to check your practice

先用自己的话解释上面的判断,再检查输入条件,最后回原文验证。说不清原因时回概念;缺日期、字段或权限时补输入;结论无出处时先撤回结论。Explain the example in your own words, inspect input constraints, then verify against the source. Revisit concepts if you cannot explain why; supply missing dates, fields or permissions; withdraw unsupported conclusions.

按这篇主题,先从“概念没分清”练习查起For this topic, start with concept confusion · 查看全部六个情境与理由See all six scenarios and their reasoning

出处Sources
  • OpenAI · Business termsOpenAI · Business terms ↗
  • Anthropic · Commercial termsAnthropic · Commercial terms ↗
  • 学 AI · 怎么把安全要求写进采购条款Learn AI · Writing security requirements into procurement ↗