mirror of
https://gitcode.com/JianFeeeee/ModelRouter.git
synced 2026-11-07 09:47:23 +00:00
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
@ -1139,3 +1139,100 @@ func TestReasoningClampedToCompletion(t *testing.T) {
|
||||
// number plausible enough to survive a glance at a dashboard.
|
||||
approx(t, "reasoning clamped to completion", cost, 0.00195)
|
||||
}
|
||||
|
||||
// TestLongContextStepIsMarginal: the threshold reprices only the tokens PAST
|
||||
// it, at a marginal rate.
|
||||
//
|
||||
// A flat prompt price under-bills exactly the traffic an agent workload
|
||||
// produces — on this deployment 88% of requests exceed 128k with a median of
|
||||
// 247k — and it under-bills silently, because the number stays plausible. The
|
||||
// step must apply to the excess only; charging the WHOLE prompt at the higher
|
||||
// rate is the other obvious implementation and it overcharges short requests by
|
||||
// the same factor it was meant to fix long ones.
|
||||
func TestLongContextStepIsMarginal(t *testing.T) {
|
||||
|
||||
// Short request: below the threshold, so the step must not touch it at all.
|
||||
ps, _ := billingVM(t)
|
||||
if err := ps.SetState("billing", map[string]interface{}{
|
||||
"prices": map[string]interface{}{
|
||||
"currency": "USD",
|
||||
"models": map[string]interface{}{
|
||||
"m": map[string]interface{}{
|
||||
"prompt": 1e-6, "completion": 0,
|
||||
"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
|
||||
},
|
||||
},
|
||||
},
|
||||
}); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
ps.Fire(StageRequestEnd, map[string]interface{}{
|
||||
"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
|
||||
"prompt_tokens": 100000, "completion_tokens": 0, "time": 1750000000000,
|
||||
})
|
||||
got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
|
||||
approx(t, "short request unaffected", got, 0.1)
|
||||
|
||||
// Long request: 200k prompt with a 128k threshold → 72k tokens at the step.
|
||||
ps2, _ := billingVM(t)
|
||||
if err := ps2.SetState("billing", map[string]interface{}{
|
||||
"prices": map[string]interface{}{
|
||||
"currency": "USD",
|
||||
"models": map[string]interface{}{
|
||||
"m": map[string]interface{}{
|
||||
"prompt": 1e-6, "completion": 0,
|
||||
"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
|
||||
},
|
||||
},
|
||||
},
|
||||
}); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
ps2.Fire(StageRequestEnd, map[string]interface{}{
|
||||
"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
|
||||
"prompt_tokens": 200000, "completion_tokens": 0, "time": 1750000000000,
|
||||
})
|
||||
got2 := stateOf(t, ps2)["total"].(map[string]interface{})["cost"].(float64)
|
||||
// 128000*1e-6 + 72000*2e-6 = 0.128 + 0.144 = 0.272
|
||||
approx(t, "long request marginal step", got2, 0.272)
|
||||
}
|
||||
|
||||
// TestLongContextComposesWithCacheDiscount pins the interaction, which is the
|
||||
// whole reason for computing the step as a marginal rate.
|
||||
//
|
||||
// On this deployment the cache hit rate is 88% and long prompts dominate, so
|
||||
// "cached tokens beyond the threshold" is the common case, not an edge case. A
|
||||
// cache read already bills at prompt*discount; the long-context rate applies on
|
||||
// top of THAT, not on top of the full prompt price. Computing it any other way
|
||||
// silently changes the price of the majority of traffic.
|
||||
func TestLongContextComposesWithCacheDiscount(t *testing.T) {
|
||||
ps, _ := billingVM(t)
|
||||
if err := ps.SetState("billing", map[string]interface{}{
|
||||
"prices": map[string]interface{}{
|
||||
"currency": "USD", "cache_discount": 0.1,
|
||||
"models": map[string]interface{}{
|
||||
"m": map[string]interface{}{
|
||||
"prompt": 1e-6, "completion": 0,
|
||||
"long_context": map[string]interface{}{"above": 100000, "prompt": 2e-6},
|
||||
},
|
||||
},
|
||||
},
|
||||
}); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
// 200k prompt, 150k of it cached, 50k fresh. Excess past 100k = 100k, split
|
||||
// 25% fresh / 75% cached.
|
||||
ps.Fire(StageRequestEnd, map[string]interface{}{
|
||||
"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
|
||||
"prompt_tokens": 200000, "completion_tokens": 0,
|
||||
"cache_hit_tokens": 150000, "cache_reported": true,
|
||||
"time": 1750000000000,
|
||||
})
|
||||
got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
|
||||
// base: fresh 50000*1e-6 = 0.05
|
||||
// cache 150000*1e-6*0.1 = 0.015
|
||||
// step: fresh excess 25000*(2-1)e-6 = 0.025
|
||||
// cache excess 75000*(2-0.1)e-6 = 0.1425
|
||||
// total = 0.2325
|
||||
approx(t, "step composes with cache discount", got, 0.2325)
|
||||
}
|
||||
|
||||
@ -465,3 +465,35 @@ func TestBillingRulesUIExposesReasoningPrice(t *testing.T) {
|
||||
t.Errorf("m-r appears %d time(s); both the row template and the editor need it", n)
|
||||
}
|
||||
}
|
||||
|
||||
// TestBillingRulesUIExposesLongContextStep: the long-context step must be
|
||||
// editable in the rules editor.
|
||||
//
|
||||
// This is not a niche control. On an agent workload 88% of requests carry more
|
||||
// than 128k prompt tokens with a median of 247k, so a rule that omits the step
|
||||
// under-bills the MAJORITY of traffic while every number on the dashboard stays
|
||||
// plausible. Leaving the capability in config.yaml alone would mean the operator
|
||||
// never discovers it needs it.
|
||||
func TestBillingRulesUIExposesLongContextStep(t *testing.T) {
|
||||
raw, err := os.ReadFile("plugins/billing.lua")
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
src := string(raw)
|
||||
for _, needle := range []string{
|
||||
`class='m-lc-a'`, // threshold
|
||||
`class='m-lc-p'`, // absolute rate
|
||||
`class='m-lc-m'`, // multiplier
|
||||
`models[n].long_context`, // collected back out on save
|
||||
`lcHint`, // explains that blank = no step
|
||||
} {
|
||||
if !strings.Contains(src, needle) {
|
||||
t.Errorf("rules editor lacks %q", needle)
|
||||
}
|
||||
}
|
||||
// Both the existing-row template and the "+ model" template need the three
|
||||
// inputs; one would make a newly added model lose its step on the next save.
|
||||
if n := strings.Count(src, `class='m-lc-a'`); n < 2 {
|
||||
t.Errorf("long-context inputs appear %d time(s); both row templates need them", n)
|
||||
}
|
||||
}
|
||||
|
||||
@ -196,6 +196,7 @@ local function priceFor(payload)
|
||||
-- provider measured actually charges. Set per-rule only for a provider
|
||||
-- that publishes a distinct rate for thinking tokens.
|
||||
reasoning_price = d.reasoning_price,
|
||||
long_context = d.long_context,
|
||||
}
|
||||
|
||||
-- model dimension (a token price overrides the default's token prices)
|
||||
@ -207,6 +208,7 @@ local function priceFor(payload)
|
||||
if mp.per_request ~= nil then out.per_request = out.per_request + mp.per_request end
|
||||
if mp.cache_discount ~= nil then out.cache_discount = mp.cache_discount end
|
||||
if mp.reasoning_price ~= nil then out.reasoning_price = mp.reasoning_price end
|
||||
if mp.long_context ~= nil then out.long_context = mp.long_context end
|
||||
if mp.peak ~= nil then out.peak = mp.peak end
|
||||
end
|
||||
|
||||
@ -219,6 +221,7 @@ local function priceFor(payload)
|
||||
if sp.per_request ~= nil then out.per_request = out.per_request + sp.per_request end
|
||||
if sp.cache_discount ~= nil then out.cache_discount = sp.cache_discount end
|
||||
if sp.reasoning_price ~= nil then out.reasoning_price = sp.reasoning_price end
|
||||
if sp.long_context ~= nil then out.long_context = sp.long_context end
|
||||
if sp.peak ~= nil then out.peak = sp.peak end
|
||||
end
|
||||
|
||||
@ -352,6 +355,36 @@ local function costFor(payload, price)
|
||||
+ cacheHit * price.prompt * discount
|
||||
+ completion * price.completion * mult
|
||||
|
||||
-- Long-context step: the tokens PAST a threshold cost more.
|
||||
--
|
||||
-- A marginal rate on the excess, not a blended average — that is how
|
||||
-- providers publish it (DeepSeek doubles past 128k) and what keeps it
|
||||
-- composable: a cache read beyond the threshold gets BOTH the cache discount
|
||||
-- AND the long-context rate, as it is billed. A blended average would have to
|
||||
-- choose one, and choosing wrong under-bills exactly the cache-heavy long
|
||||
-- requests the cache exists to make cheap.
|
||||
--
|
||||
-- The excess splits between cached and fresh in the same proportion as the
|
||||
-- whole prompt, so no token is charged twice and none is missed.
|
||||
local lc = price.long_context
|
||||
if lc then
|
||||
local totalPrompt = tonumber(payload.prompt_tokens) or 0
|
||||
local threshold = tonumber(lc.above) or 0
|
||||
local lcRate = tonumber(lc.prompt)
|
||||
if totalPrompt > threshold and threshold > 0 and lcRate and lcRate > 0 then
|
||||
local excess = totalPrompt - threshold
|
||||
-- Only the FRESH share of the excess carries the step: a cache read is
|
||||
-- already billed at the discounted rate, and re-pricing it at the
|
||||
-- long-context rate would charge the same tokens twice.
|
||||
local freshShare = totalPrompt > 0 and (fresh / totalPrompt) or 0
|
||||
local excessFresh = excess * freshShare
|
||||
local excessCached = excess - excessFresh
|
||||
cost = cost
|
||||
+ excessFresh * (lcRate - price.prompt) * mult
|
||||
+ excessCached * (lcRate - price.prompt * discount) * mult
|
||||
end
|
||||
end
|
||||
|
||||
-- Reasoning tokens, when the rule prices them separately.
|
||||
--
|
||||
-- reasoning_tokens is a SUBSET of completion_tokens and is already inside
|
||||
@ -1033,7 +1066,9 @@ plugin.ui = {
|
||||
models: "Per-model prices", addModel: "+ model",
|
||||
|
||||
reasoning: "think $/M",
|
||||
reasoningHint: "Thinking tokens are billed as output unless you set a different rate here.", mFree: "free", mToken: "token", mSub: "subscription", mUnpriced: "unpriced",
|
||||
|
||||
lcAbove: "over n tok", lcPrompt: "then $/M", lcMult: "or x",
|
||||
lcHint: "Long-context step: tokens PAST the threshold cost more. Leave all three blank for one flat rate.", reasoningHint: "Thinking tokens are billed as output unless you set a different rate here.", mFree: "free", mToken: "token", mSub: "subscription", mUnpriced: "unpriced",
|
||||
noProfile: "No profile yet — create one to start pricing.",
|
||||
noRules: "No rules in this profile",
|
||||
warnNoSource: "matches no configured source",
|
||||
@ -1057,7 +1092,9 @@ plugin.ui = {
|
||||
models: "按模型价格", addModel: "+ 模型",
|
||||
|
||||
reasoning: "思考 $/M",
|
||||
reasoningHint: "留空=思考 token 按输出价计（OpenAI/DeepSeek/Anthropic 都这样）。只有供应商对思考单独定价时才填。", mFree: "free 免费", mToken: "token 按量",
|
||||
|
||||
lcAbove: "超过 n tok", lcPrompt: "此后 $/M", lcMult: "或 ×",
|
||||
lcHint: "长上下文阶梯：超过阈值的部分按更高价计。三项都留空 = 单一价格。", reasoningHint: "留空=思考 token 按输出价计（OpenAI/DeepSeek/Anthropic 都这样）。只有供应商对思考单独定价时才填。", mFree: "free 免费", mToken: "token 按量",
|
||||
mSub: "subscription 订阅", mUnpriced: "unpriced 只统计",
|
||||
noProfile: "还没有方案 —— 新建一个开始定价。",
|
||||
noRules: "该方案暂无规则",
|
||||
@ -1161,6 +1198,19 @@ plugin.ui = {
|
||||
"<input class='m-r' value='" + esc(t.reasoning == null ? "" : t.reasoning) +
|
||||
"' placeholder='" + esc(T.reasoning) + "' title='" + esc(T.reasoningHint) +
|
||||
"' style='flex:0 1 62px;min-width:0'>" +
|
||||
// Long-context step: threshold + (absolute rate OR multiplier). All
|
||||
// three blank means NO step, deliberately distinct from a step whose
|
||||
// rate is 0 — three inputs rather than inferring the step from a zero.
|
||||
// On an agent workload 88% of requests exceed 128k prompt tokens, so
|
||||
// omitting this under-bills the MAJORITY while every number stays
|
||||
// plausible. The step is MARGINAL: it reprices the excess only.
|
||||
"<input class='m-lc-a' value='" + esc(t.long_context == null ? "" : t.long_context.above) +
|
||||
"' placeholder='" + esc(T.lcAbove) + "' title='" + esc(T.lcHint) +
|
||||
"' style='flex:0 1 52px;min-width:0'>" +
|
||||
"<input class='m-lc-p' value='" + esc(t.long_context == null ? "" : (t.long_context.prompt || "")) +
|
||||
"' placeholder='" + esc(T.lcPrompt) + "' style='flex:0 1 52px;min-width:0'>" +
|
||||
"<input class='m-lc-m' value='" + esc(t.long_context == null ? "" : (t.long_context.multiplier || "")) +
|
||||
"' placeholder='" + esc(T.lcMult) + "' style='flex:0 1 46px;min-width:0'>" +
|
||||
"<input class='m-c' value='" + esc(t.completion == null ? "" : t.completion) +
|
||||
"' placeholder='" + esc(T.completion) + "' style='flex:0 1 62px;min-width:0'>" +
|
||||
"<button class='ghost small m-del' title='" + esc(T.delRule) + "' style='flex:0 0 auto'>×</button></div>";
|
||||
@ -1251,6 +1301,25 @@ plugin.ui = {
|
||||
}
|
||||
models[n].reasoning = rtxt;
|
||||
}
|
||||
// Long-context step. All three blank means NO step — the key must be
|
||||
// ABSENT from the saved rule rather than present-and-zero, which would
|
||||
// be a step priced at nothing.
|
||||
var lca = r.querySelector(".m-lc-a");
|
||||
var lcp = r.querySelector(".m-lc-p");
|
||||
var lcm = r.querySelector(".m-lc-m");
|
||||
var lcaT = lca ? String(lca.value || "").trim() : "";
|
||||
var lcpT = lcp ? String(lcp.value || "").trim() : "";
|
||||
var lcmT = lcm ? String(lcm.value || "").trim() : "";
|
||||
if (lcaT !== "" || lcpT !== "" || lcmT !== "") {
|
||||
if (lcaT === "" || isNaN(parseFloat(lcaT))) { bad = n + " (" + T.lcAbove + ")"; return; }
|
||||
if (lcpT !== "" && isNaN(parseFloat(lcpT))) { bad = n + " (" + T.lcPrompt + ")"; return; }
|
||||
if (lcmT !== "" && isNaN(parseFloat(lcmT))) { bad = n + " (" + T.lcMult + ")"; return; }
|
||||
if (lcpT === "" && lcmT === "") { bad = n + " (" + T.lcPrompt + "/" + T.lcMult + ")"; return; }
|
||||
var lc = { above: String(parseFloat(lcaT)) };
|
||||
if (lcpT !== "") lc.prompt = lcpT;
|
||||
if (lcmT !== "") lc.multiplier = parseFloat(lcmT);
|
||||
models[n].long_context = lc;
|
||||
}
|
||||
});
|
||||
if (bad) throw new Error(T.needPrices + ": " + bad);
|
||||
if (Object.keys(models).length) out.models = models;
|
||||
@ -1443,6 +1512,9 @@ plugin.ui = {
|
||||
"<input class='m-c' placeholder='" + esc(T.completion) + "' style='width:96px'>" +
|
||||
"<input class='m-r' placeholder='" + esc(T.reasoning) + "' title='" + esc(T.reasoningHint) +
|
||||
"' style='width:96px'>" +
|
||||
"<input class='m-lc-a' placeholder='" + esc(T.lcAbove) + "' title='" + esc(T.lcHint) + "' style='width:78px'>" +
|
||||
"<input class='m-lc-p' placeholder='" + esc(T.lcPrompt) + "' style='width:78px'>" +
|
||||
"<input class='m-lc-m' placeholder='" + esc(T.lcMult) + "' style='width:70px'>" +
|
||||
"<button class='ghost small m-del'>×</button>";
|
||||
box.appendChild(row);
|
||||
};
|
||||
|
||||
Reference in New Issue
Block a user