— 定价Pricing
每次优化真实消耗 GPU 分钟和 token，所以按量计费。诊断阶段诊断和验证不计费——看到加速比之后，你再决定要不要付。 Every run burns real GPU minutes and real tokens, so billing is metered. Diagnosis is free: you decide whether to pay once you've seen the speedup and the quote.
为什么订阅是双倍 credits。订阅让我们能预测负载、提前锁定 GPU 容量，这部分节省直接还给你。代价是订阅 credits 按计费周期清零——不清零的话，订一个月囤一年的量，容量预测就不成立了。充值的 credits 永不过期。 Why subscriptions get 2× credits. A subscription lets us forecast load and reserve GPU capacity ahead of time, and that saving goes straight back to you. The trade-off: subscription credits reset at the end of each billing period. If they rolled over, one month's subscription could stockpile a year of capacity and the forecast stops meaning anything. Topped-up credits never expire.
— 费率表Rate card
没有加价系数，没有最低消费。你付的就是这次运行真实烧掉的 GPU 时间和 token，按下面的费率折算成 credits。No markup multiplier, no minimum spend. You pay for the GPU time and tokens the run actually burned, converted to credits at the rates below.
| NVIDIA T4 | 0.59 credits / h |
| NVIDIA L4 | 0.80 credits / h |
| NVIDIA A10 | 1.10 credits / h |
| NVIDIA L40S | 1.95 credits / h |
| NVIDIA A100 · 40GB | 2.10 credits / h |
| NVIDIA A100 · 80GB | 2.50 credits / h |
| NVIDIA H100 | 3.95 credits / h |
| NVIDIA H200 | 4.54 credits / h |
| NVIDIA B200 | 6.25 credits / h |
| CPU 阶段（分诊、装环境、诊断）CPU stages (triage, env, diagnosis) | 不计费free |
按卡型计价，1 credit = $1，按你选的机器占用时长计。诊断和验证不计费 —— 只有在你自己的程序里端到端复测确实更快了，才会扣。Priced by card class at 1 credit = $1, charged for the time the machine is held. Diagnosis and verification are not billed: the charge happens only once it is measurably faster end to end in your own program.
| Opus 5 输入input | 5.00 cr per MTok |
| Opus 5 输出output | 25.00 cr per MTok |
| Opus 5 缓存写入cache write | 6.25 cr per MTok |
| Opus 5 缓存读取cache read | 0.50 cr per MTok |
= Anthropic 官方 API 价目表。不论 agent 实际调用什么模型，一律按 Opus 5 费率计——费率表只有这一行要记。缓存命中按 0.1× 折算，长会话越跑越便宜。= Anthropic's list API pricing. Whatever model the agent actually uses, tokens bill at the Opus 5 rate — one line to remember. Cache hits bill at 0.1×, so long sessions get cheaper as they run.
| 项目Item | 用量Quantity | Credits |
|---|---|---|
| H100 GPU | 13.2 min | 0.87 |
| Opus 5 缓存读取cache read | 1.50 MTok | 0.75 |
| Opus 5 新增输入fresh input | 0.30 MTok | 1.50 |
| Opus 5 输出output | 0.15 MTok | 3.75 |
| 合计Total | = $6.87 | 6.87 |
这次拿到 1.61× 端到端加速。订阅(2×)下实付 $3.44。 That run returned a 1.61× end-to-end speedup. On a subscription (2×), you actually pay $3.44.
— 计费规则Billing rules
| 诊断阶段Diagnosis | 不计费。跑基线、建评测台、找瓶颈，这一段我们承担。Free. Baseline, profiling and bottleneck ranking are on us. |
| 报价Quote | 诊断完成后给出预期加速比和 credit 报价。你确认后才开始改代码。After diagnosis you get a projected speedup and a credit quote. Nothing is modified until you accept. |
| 达标Target met | 按实际消耗扣费，且不超过报价。超出部分我们吃掉。Charged at actual consumption, and never above the quote. We absorb any overrun. |
| 未达标Target missed | credits 全额退回账户，不是退款到卡。失败报告照样给你。Credits are returned to your balance in full — not refunded to your card. You still get the failure report. |
| 环境跑不起来Env won't build | 完全不扣费，连诊断都不算。No charge at all, not even for diagnosis. |
| 扣费顺序Deduction order | 先扣本周期的订阅 credits，再扣永久有效的充值 credits。This period's subscription credits are spent first, then your permanent topped-up balance. |
加速比怎么算：训练看单步墙钟时间的比值，推理看单位时间吞吐的比值。报告里同时给出启动时间变化和每步耗时变化，测量条件锁死，给的是分布不是单值。How speedup is measured: for training, the ratio of per-step wall-clock time; for inference, the ratio of throughput per unit time. The report also gives the change in startup time and per-step latency. Conditions are frozen and we report a distribution, not a single number.