feat(billing): 长上下文阶梯定价(边际费率,与缓存折扣/峰段可叠加)

## 影响面实测

审计里 1000 条请求:**883 条(88.3%)prompt 超过 128k**,中位数 246849,
p90 298179,最大 354251。这不是小众场景,是 agent 负载的主流形态。

按「模型 × 源」拆开:其中真正按 token 计价的只有 alittokenplan 的 10 次
(deepseek-v4.1-flash,约 350k prompt),其余走 $70 套餐(commandcode)或
免费源(localzen),记 0 是正确的,对阶梯无影响。

那 10 次当前少算 54.9%($0.3148 → $0.4877)。绝对值不大,但比例严重,
而且这个偏差只会随上下文变长而变大。

## 边际而非混合平均

阶梯只对**超过阈值的部分**重新计价,不是把整段按高价算。理由是可叠加性:
缓存读取同时享受缓存折扣**和**长上下文阶梯价,这与供应商的计费方式一致。
混合平均必须二选一,选错就正好低估「缓存命中的长请求」——而这恰是缓存
存在的意义所在(本部署缓存命中率 88%)。

超出部分按缓存/新鲜的比例拆分,不重不漏。

## DSL 两种写法

供应商两种发布方式都支持:绝对费率(OpenAI 报 \$3.00/M past 200k)和
倍数(DeepSeek 报 2x past 128k)。两者都缺必须报错 —— 一个没有费率的阶梯
会编译成一条「看起来处理了长上下文、实际没处理」的规则,静默少算且无迹可寻。
两者都给时以绝对费率为准,因为那是供应商真正报出的数字。

## 判据

Lua 侧两条(边际性、与缓存折扣叠加)+ Go 侧一条(两种写法与拒绝规则)。
变异自证三次:整段按高价 / 缓存部分不叠加折扣 / 缓存部分不计阶梯,
分别命中 0.40 vs 0.272、0.165 vs 0.2325、0.09 vs 0.2325。

规则编辑器加了三个输入(阈值 + 绝对费率 + 倍数),三项都留空 = 无阶梯 ——
与「阶梯费率为 0」刻意区分,所以是三个独立输入而不是从 0 推断是否存在阶梯。

Co-Authored-By: ModelRouter <noreply@modelrouter.dev>
This commit is contained in:
JianFeeeee
2026-10-04 14:36:38 +08:00
parent b93799a919
commit 314f4da8f2
6 changed files with 343 additions and 2 deletions

View File

@ -463,6 +463,9 @@ type BillingToken struct {
// setting it to a multiple of Completion reprices the thinking share rather
// than billing it twice.
Reasoning string `yaml:"reasoning,omitempty" json:"reasoning,omitempty"`
// LongContext is the prompt-price step above a token threshold. Nil = one
// flat rate, which is what a rule means when it says nothing.
LongContext *BillingLongContext `yaml:"long_context,omitempty" json:"long_context,omitempty"`
}
// BillingPeak declares peak-hour pricing.
@ -474,6 +477,31 @@ type BillingPeak struct {
Hours [][2]int `yaml:"hours,omitempty" json:"hours,omitempty"`
}
// BillingLongContext is a prompt-price step for long inputs.
//
// Every major provider charges the input tokens ABOVE a threshold at a higher
// rate — DeepSeek doubles past 128k, OpenAI and Anthropic step up as well. A
// flat prompt price therefore under-bills exactly the traffic an agent workload
// generates, and it under-bills it silently: the number is plausible, the bill
// adds up, and nothing says the rate was wrong.
//
// The step applies to the PROMPT only. Output and cache reads are billed at
// their own rates regardless of input length, which is how providers publish
// it — the higher rate is for the extra context you made the model read, not for
// what it wrote back.
type BillingLongContext struct {
// Above is the token count the step starts at. A request with exactly this
// many prompt tokens is still at the base rate; the step covers the tokens
// BEYOND it. Inclusive-below matches how the providers document it.
Above int `yaml:"above" json:"above"`
// Prompt is the USD-per-million rate for the tokens above Above.
Prompt string `yaml:"prompt" json:"prompt"`
// Multiplier is an alternative to Prompt: a factor applied to the base rate
// (DeepSeek's published step is exactly 2x). When both are set Prompt wins,
// and setting neither is a validation error rather than a silent no-op.
Multiplier float64 `yaml:"multiplier,omitempty" json:"multiplier,omitempty"`
}
// Validate rejects a BillingDSL that could not work as written.
func (b *BillingDSL) Validate() error {
if b == nil {