feat: add `len` variable for tier conditions and LLM prompt helper

pkg/billingexpr/billingexpr_test.go

@@ -1000,11 +1000,82 @@ func TestImageAudioZero(t *testing.T) {
 	}
 }
 
+// ---------------------------------------------------------------------------
+// len variable tests — tier conditions based on context length
+// ---------------------------------------------------------------------------
+
+const lenTieredExpr = `len <= 200000 ? tier("standard", p * 3 + c * 15 + cr * 0.3) : tier("long_context", p * 6 + c * 22.5 + cr * 0.6)`
+
+func TestLen_StandardTier(t *testing.T) {
+	params := billingexpr.TokenParams{P: 80000, C: 5000, Len: 100000, CR: 20000}
+	cost, trace, err := billingexpr.RunExpr(lenTieredExpr, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := 80000*3 + 5000*15 + 20000*0.3
+	if math.Abs(cost-want) > 1e-6 {
+		t.Errorf("cost = %f, want %f", cost, want)
+	}
+	if trace.MatchedTier != "standard" {
+		t.Errorf("tier = %q, want standard", trace.MatchedTier)
+	}
+}
+
+func TestLen_LongContextTier(t *testing.T) {
+	// p is low (cache subtracted), but len is high (full context)
+	params := billingexpr.TokenParams{P: 50000, C: 5000, Len: 300000, CR: 250000}
+	cost, trace, err := billingexpr.RunExpr(lenTieredExpr, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := 50000*6 + 5000*22.5 + 250000*0.6
+	if math.Abs(cost-want) > 1e-6 {
+		t.Errorf("cost = %f, want %f", cost, want)
+	}
+	if trace.MatchedTier != "long_context" {
+		t.Errorf("tier = %q, want long_context (len=300000 > 200000)", trace.MatchedTier)
+	}
+}
+
+func TestLen_BoundaryExact(t *testing.T) {
+	params := billingexpr.TokenParams{P: 100000, C: 1000, Len: 200000, CR: 100000}
+	_, trace, err := billingexpr.RunExpr(lenTieredExpr, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if trace.MatchedTier != "standard" {
+		t.Errorf("tier = %q, want standard (len=200000 <= 200000)", trace.MatchedTier)
+	}
+}
+
+func TestLen_BoundaryPlusOne(t *testing.T) {
+	params := billingexpr.TokenParams{P: 100000, C: 1000, Len: 200001, CR: 100001}
+	_, trace, err := billingexpr.RunExpr(lenTieredExpr, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if trace.MatchedTier != "long_context" {
+		t.Errorf("tier = %q, want long_context (len=200001 > 200000)", trace.MatchedTier)
+	}
+}
+
+func TestLen_ZeroDefaultsToZero(t *testing.T) {
+	// len defaults to 0 when not set
+	params := billingexpr.TokenParams{P: 1000, C: 500}
+	_, trace, err := billingexpr.RunExpr(lenTieredExpr, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if trace.MatchedTier != "standard" {
+		t.Errorf("tier = %q, want standard (len=0 <= 200000)", trace.MatchedTier)
+	}
+}
+
 // ---------------------------------------------------------------------------
 // Benchmarks: compile vs cached execution
 // ---------------------------------------------------------------------------
 
-const benchComplexExpr = `p <= 200000 ? tier("standard", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6 + img * 3 + img_o * 30 + ai * 10 + ao * 40) : tier("long_context", p * 6 + c * 22.5 + cr * 0.6 + cc * 7.5 + cc1h * 12 + img * 6 + img_o * 60 + ai * 20 + ao * 80)`
+const benchComplexExpr = `len <= 200000 ? tier("standard", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6 + img * 3 + img_o * 30 + ai * 10 + ao * 40) : tier("long_context", p * 6 + c * 22.5 + cr * 0.6 + cc * 7.5 + cc1h * 12 + img * 6 + img_o * 60 + ai * 20 + ao * 80)`
 
 func BenchmarkExprCompile(b *testing.B) {
 	for i := 0; i < b.N; i++ {
@@ -1015,7 +1086,7 @@ func BenchmarkExprCompile(b *testing.B) {
 
 func BenchmarkExprRunCached(b *testing.B) {
 	billingexpr.CompileFromCache(benchComplexExpr)
-	params := billingexpr.TokenParams{P: 150000, C: 10000, CR: 30000, CC: 5000, Img: 2000, AI: 1000, AO: 500}
+	params := billingexpr.TokenParams{P: 150000, C: 10000, Len: 188000, CR: 30000, CC: 5000, Img: 2000, AI: 1000, AO: 500}
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		billingexpr.RunExpr(benchComplexExpr, params)

pkg/billingexpr/compile.go

@@ -41,6 +41,7 @@ var (
 var compileEnvPrototypeV1 = map[string]interface{}{
 	"p":    float64(0),
 	"c":    float64(0),
+	"len":  float64(0),
 	"cr":   float64(0),
 	"cc":   float64(0),
 	"cc1h": float64(0),

pkg/billingexpr/expr.md

@@ -30,7 +30,8 @@ Powered by [expr-lang/expr](https://github.com/expr-lang/expr). Expressions are
 
 | 变量 | 含义 |
 |------|------|
-| `p` | 输入 token 数。**自动排除**表达式中单独计价的子类别（见下方说明） |
+| `p` | 输入 token 数（**计价用**）。**自动排除**表达式中单独计价的子类别（见下方说明） |
+| `len` | 输入上下文总长度（**条件判断用**）。不受自动排除影响，始终反映完整输入长度。非 Claude：等于原始 `prompt_tokens`；Claude：等于文本输入 + 缓存读取 + 缓存创建 |
 | `cr` | 缓存命中（读取）token 数 |
 | `cc` | 缓存创建 token 数（Claude 5分钟 TTL / 通用） |
 | `cc1h` | 缓存创建 token 数 — 1小时 TTL（Claude 专用） |
@@ -51,6 +52,8 @@ Powered by [expr-lang/expr](https://github.com/expr-lang/expr). Expressions are
 
 **规则：如果表达式使用了某个子类别变量，对应的 token 就从 `p` 或 `c` 中扣除；如果没使用，那些 token 就留在 `p` 或 `c` 里按基础价格计费。**
 
+> **重要：`len` 不受自动排除影响。** `len` 始终代表完整的输入上下文长度，不管表达式是否单独对缓存/图片/音频定价。因此**阶梯条件应使用 `len` 而非 `p`**，以避免缓存命中导致 `p` 降低而误判档位。
+
 举例说明（假设上游返回的原始数据：prompt_tokens=1000，其中包含 200 cache read、100 image）：
 
 | 表达式 | `p` 的值 | 说明 |
@@ -93,8 +96,8 @@ Powered by [expr-lang/expr](https://github.com/expr-lang/expr). Expressions are
 # Simple flat pricing
 tier("base", p * 2.5 + c * 15 + cr * 0.25)
 
-# Multi-tier (Claude Sonnet style)
-p <= 200000
+# Multi-tier (Claude Sonnet style) — use len for tier conditions
+len <= 200000
   ? tier("standard", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6)
   : tier("long_context", p * 6 + c * 22.5 + cr * 0.6 + cc * 7.5 + cc1h * 12)
 
@@ -199,6 +202,16 @@ Example: `p * 2.5 + c * 15 + cr * 0.25`
 - Expression uses `cr` → cache read tokens subtracted from `p`
 - Expression doesn't use `img` → image tokens stay in `p`, priced at $2.50
 
+### `len` — Context Length Variable
+
+`len` represents the total input context length, designed for **tier condition evaluation** (e.g. `len <= 200000 ? ...`). Unlike `p`, `len` is never reduced by sub-category exclusion.
+
+**Computation rules:**
+- **Non-Claude (GPT/OpenAI format)**: `len = prompt_tokens` (the raw total from the upstream response)
+- **Claude format**: `len = input_tokens + cache_read_tokens + cache_creation_tokens` (since Claude's `input_tokens` is text-only, cache must be added back to reflect full context length)
+
+This ensures that heavy cache usage doesn't cause the tier condition to incorrectly evaluate to a lower tier. For example, if a request has 300K total context but 250K is cached, `p` with cache subtracted would be only 50K (standard tier), while `len` correctly reports 300K (long-context tier).
+
 ### Quota Conversion
 
 Expression coefficients are $/1M tokens. Conversion to internal quota:

pkg/billingexpr/run.go

@@ -13,7 +13,8 @@ import (
 
 // RunExpr compiles (with cache) and executes an expression string.
 // The environment exposes:
-//   - p, c             — prompt / completion tokens
+//   - p, c             — prompt / completion tokens (auto-excluding separately-priced sub-categories)
+//   - len              — total input context length for tier conditions (never reduced by sub-category exclusion)
 //   - cr, cc, cc1h     — cache read / creation / creation-1h tokens
 //   - tier(name, value) — trace callback that records which tier matched
 //   - max, min, abs, ceil, floor — standard math helpers
@@ -54,6 +55,7 @@ func runProgram(prog *vm.Program, params TokenParams, request RequestInput) (flo
 	env := map[string]interface{}{
 		"p":    params.P,
 		"c":    params.C,
+		"len":  params.Len,
 		"cr":   params.CR,
 		"cc":   params.CC,
 		"cc1h": params.CC1h,

pkg/billingexpr/types.go

@@ -14,8 +14,9 @@ type RequestInput struct {
 // Fields beyond P and C are optional — when absent they default to 0,
 // which means cache-unaware expressions keep working unchanged.
 type TokenParams struct {
-	P    float64 // prompt tokens (text)
-	C    float64 // completion tokens (text)
+	P    float64 // prompt tokens (text) — auto-excludes sub-categories priced separately
+	C    float64 // completion tokens (text) — auto-excludes sub-categories priced separately
+	Len  float64 // total input context length for tier conditions (non-Claude: raw prompt_tokens; Claude: text + cache read + cache creation)
 	CR   float64 // cache read (hit) tokens
 	CC   float64 // cache creation tokens (5-min TTL for Claude, generic for others)
 	CC1h float64 // cache creation tokens — 1-hour TTL (Claude only)

relay/helper/price.go

@@ -255,8 +255,9 @@ func modelPriceHelperTiered(c *gin.Context, info *relaycommon.RelayInfo, promptT
 	}
 
 	rawCost, trace, err := billingexpr.RunExprWithRequest(exprStr, billingexpr.TokenParams{
-		P: float64(promptTokens),
-		C: float64(estimatedCompletionTokens),
+		P:   float64(promptTokens),
+		C:   float64(estimatedCompletionTokens),
+		Len: float64(promptTokens),
 	}, requestInput)
 	if err != nil {
 		return types.PriceData{}, fmt.Errorf("model %s tiered expr run failed: %w", info.OriginModelName, err)

service/quota.go

@@ -160,8 +160,9 @@ func PostWssConsumeQuota(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, mod
 
 	var tieredResult *billingexpr.TieredResult
 	tieredOk, tieredQuota, tieredRes := TryTieredSettle(relayInfo, billingexpr.TokenParams{
-		P: float64(usage.InputTokens),
-		C: float64(usage.OutputTokens),
+		P:   float64(usage.InputTokens),
+		C:   float64(usage.OutputTokens),
+		Len: float64(usage.InputTokens),
 	})
 	if tieredOk {
 		tieredResult = tieredRes

service/tiered_settle.go

@@ -35,6 +35,14 @@ func BuildTieredTokenParams(usage *dto.Usage, isClaudeUsageSemantic bool, usedVa
 	imgO := float64(usage.CompletionTokenDetails.ImageTokens)
 	ao := float64(usage.CompletionTokenDetails.AudioTokens)
 
+	// len = total input context length for tier condition evaluation.
+	// Non-Claude: prompt_tokens already includes everything.
+	// Claude: input_tokens is text-only, so add cache read + cache creation.
+	inputLen := p
+	if isClaudeUsageSemantic {
+		inputLen = p + cr + cc5m + cc1h
+	}
+
 	if !isClaudeUsageSemantic {
 		if usedVars["cr"] {
 			p -= cr
@@ -69,6 +77,7 @@ func BuildTieredTokenParams(usage *dto.Usage, isClaudeUsageSemantic bool, usedVa
 	return billingexpr.TokenParams{
 		P:    p,
 		C:    c,
+		Len:  inputLen,
 		CR:   cr,
 		CC:   cc5m,
 		CC1h: cc1h,

service/tiered_settle_test.go

@@ -604,6 +604,97 @@ func TestBuildTieredTokenParams_ParityWithRatio_Image(t *testing.T) {
 	}
 }
 
+// ---------------------------------------------------------------------------
+// BuildTieredTokenParams: Len computation tests
+// ---------------------------------------------------------------------------
+
+func TestBuildTieredTokenParams_Len_GPT(t *testing.T) {
+	usage := &dto.Usage{
+		PromptTokens:     10000,
+		CompletionTokens: 2000,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 3000,
+			TextTokens:   7000,
+		},
+	}
+	expr := `tier("base", p * 2.5 + c * 15 + cr * 0.25)`
+	usedVars := billingexpr.UsedVars(expr)
+	params := BuildTieredTokenParams(usage, false, usedVars)
+
+	// Non-Claude: Len = raw PromptTokens
+	if params.Len != 10000 {
+		t.Fatalf("Len = %f, want 10000 (raw PromptTokens)", params.Len)
+	}
+	// P should be reduced by cache
+	if params.P != 7000 {
+		t.Fatalf("P = %f, want 7000 (PromptTokens - CachedTokens)", params.P)
+	}
+}
+
+func TestBuildTieredTokenParams_Len_Claude(t *testing.T) {
+	usage := &dto.Usage{
+		PromptTokens:     5000,
+		CompletionTokens: 2000,
+		UsageSemantic:    "anthropic",
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 3000,
+			TextTokens:   5000,
+		},
+		ClaudeCacheCreation5mTokens: 1000,
+		ClaudeCacheCreation1hTokens: 500,
+	}
+	expr := `tier("base", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6)`
+	usedVars := billingexpr.UsedVars(expr)
+	params := BuildTieredTokenParams(usage, true, usedVars)
+
+	// Claude: Len = PromptTokens + CachedTokens + CacheCreation5m + CacheCreation1h
+	wantLen := float64(5000 + 3000 + 1000 + 500)
+	if params.Len != wantLen {
+		t.Fatalf("Len = %f, want %f (text + cache read + cache creation)", params.Len, wantLen)
+	}
+	// Claude: P is not reduced (isClaudeUsageSemantic = true)
+	if params.P != 5000 {
+		t.Fatalf("P = %f, want 5000 (no subtraction for Claude)", params.P)
+	}
+}
+
+func TestBuildTieredTokenParams_Len_TierCondition(t *testing.T) {
+	// Test that len-based tier conditions work correctly when p is reduced by cache
+	usage := &dto.Usage{
+		PromptTokens:     300000,
+		CompletionTokens: 5000,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 250000,
+			TextTokens:   50000,
+		},
+	}
+	expr := `len <= 200000 ? tier("standard", p * 3 + c * 15 + cr * 0.3) : tier("long_context", p * 6 + c * 22.5 + cr * 0.6)`
+	usedVars := billingexpr.UsedVars(expr)
+	params := BuildTieredTokenParams(usage, false, usedVars)
+
+	// Len = 300000 (raw prompt), P = 50000 (300000 - 250000 cache)
+	if params.Len != 300000 {
+		t.Fatalf("Len = %f, want 300000", params.Len)
+	}
+	if params.P != 50000 {
+		t.Fatalf("P = %f, want 50000", params.P)
+	}
+
+	// Run expression: len=300000 > 200000, so long_context tier
+	cost, trace, err := billingexpr.RunExpr(expr, params)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if trace.MatchedTier != "long_context" {
+		t.Fatalf("tier = %s, want long_context (len=300000 but p=50000)", trace.MatchedTier)
+	}
+	// long_context: 50000*6 + 5000*22.5 + 250000*0.6
+	wantCost := 50000.0*6 + 5000*22.5 + 250000*0.6
+	if math.Abs(cost-wantCost) > 1e-6 {
+		t.Fatalf("cost = %f, want %f", cost, wantCost)
+	}
+}
+
 // ---------------------------------------------------------------------------
 // Stress test: 1000 concurrent goroutines, complex tiered expr vs ratio,
 // random token counts, verify correctness and measure performance

setting/billing_setting/tiered_billing.go

@@ -54,10 +54,10 @@ func SmokeTestExpr(exprStr string) error {
 
 func smokeTestExpr(exprStr string) error {
 	vectors := []billingexpr.TokenParams{
-		{P: 0, C: 0},
-		{P: 1000, C: 1000},
-		{P: 100000, C: 100000},
-		{P: 1000000, C: 1000000},
+		{P: 0, C: 0, Len: 0},
+		{P: 1000, C: 1000, Len: 1000},
+		{P: 100000, C: 100000, Len: 100000},
+		{P: 1000000, C: 1000000, Len: 1000000},
 	}
 	requests := []billingexpr.RequestInput{
 		{},

web/src/components/table/model-pricing/modal/components/DynamicPricingBreakdown.jsx

@@ -21,7 +21,7 @@ import React from 'react';
 import { Avatar, Tag, Table, Typography } from '@douyinfe/semi-ui';
 import { IconPriceTag } from '@douyinfe/semi-icons';
 import { parseTiersFromExpr, getCurrencyConfig } from '../../../../../helpers';
-import { BILLING_VARS } from '../../../../../constants';
+import { BILLING_PRICING_VARS } from '../../../../../constants';
 import {
   splitBillingExprAndRequestRules,
   tryParseRequestRuleExpr,
@@ -113,7 +113,7 @@ export default function DynamicPricingBreakdown({ billingExpr, t }) {
     );
   }
 
-  const priceFields = BILLING_VARS.map((v) => [v.field, v.shortLabel]);
+  const priceFields = BILLING_PRICING_VARS.map((v) => [v.field, v.shortLabel]);
 
   const tierColumns = [
     {

web/src/constants/billing.constants.js

@@ -13,6 +13,7 @@
 export const BILLING_VARS = [
   { key: 'p', field: 'inputPrice', tierField: 'input_unit_cost', label: '输入价格', shortLabel: '输入', side: 'input', isBase: true },
   { key: 'c', field: 'outputPrice', tierField: 'output_unit_cost', label: '补全价格', shortLabel: '补全', side: 'output', isBase: true },
+  { key: 'len', field: null, tierField: null, label: '输入长度', shortLabel: '长度', side: 'condition', isConditionOnly: true },
   { key: 'cr', field: 'cacheReadPrice', tierField: 'cache_read_unit_cost', label: '缓存读取价格', shortLabel: '缓存读', side: 'input', group: 'cache' },
   { key: 'cc', field: 'cacheCreatePrice', tierField: 'cache_create_unit_cost', label: '缓存创建价格', shortLabel: '缓存创建', side: 'input', group: 'cache' },
   { key: 'cc1h', field: 'cacheCreate1hPrice', tierField: 'cache_create_1h_unit_cost', label: '1h缓存创建价格', shortLabel: '1h缓存创建', side: 'input', group: 'cache' },
@@ -24,18 +25,20 @@ export const BILLING_VARS = [
 
 export const BILLING_VAR_KEYS = BILLING_VARS.map((v) => v.key);
 
-export const BILLING_EXTRA_VARS = BILLING_VARS.filter((v) => !v.isBase);
+export const BILLING_PRICING_VARS = BILLING_VARS.filter((v) => !v.isConditionOnly);
+
+export const BILLING_EXTRA_VARS = BILLING_VARS.filter((v) => !v.isBase && !v.isConditionOnly);
 
 export const BILLING_VAR_KEY_TO_FIELD = Object.fromEntries(
-  BILLING_VARS.map((v) => [v.key, v.field]),
+  BILLING_PRICING_VARS.map((v) => [v.key, v.field]),
 );
 
 export const BILLING_VAR_FIELD_TO_LABEL = Object.fromEntries(
-  BILLING_VARS.map((v) => [v.field, v.label]),
+  BILLING_PRICING_VARS.map((v) => [v.field, v.label]),
 );
 
 export const BILLING_VAR_FIELD_TO_SHORT_LABEL = Object.fromEntries(
-  BILLING_VARS.map((v) => [v.field, v.shortLabel]),
+  BILLING_PRICING_VARS.map((v) => [v.field, v.shortLabel]),
 );
 
 export const BILLING_CACHE_VAR_MAP = BILLING_EXTRA_VARS.map((v) => ({
@@ -44,6 +47,10 @@ export const BILLING_CACHE_VAR_MAP = BILLING_EXTRA_VARS.map((v) => ({
 }));
 
 export const BILLING_VAR_REGEX = new RegExp(
-  `\\b(${BILLING_VAR_KEYS.join('|')})\\s*\\*\\s*([\\d.eE+-]+)`,
+  `\\b(${BILLING_PRICING_VARS.map((v) => v.key).join('|')})\\s*\\*\\s*([\\d.eE+-]+)`,
   'g',
 );
+
+export const BILLING_CONDITION_VARS = BILLING_VARS.filter(
+  (v) => v.isBase || v.isConditionOnly,
+).map((v) => v.key);

web/src/helpers/render.jsx

@@ -22,7 +22,7 @@ import { Modal, Tag, Typography, Avatar } from '@douyinfe/semi-ui';
 import { copy, showSuccess } from './utils';
 import { MOBILE_BREAKPOINT } from '../hooks/common/useIsMobile';
 import {
-  BILLING_VARS,
+  BILLING_PRICING_VARS,
   BILLING_VAR_KEY_TO_FIELD,
   BILLING_VAR_REGEX,
 } from '../constants';
@@ -2246,7 +2246,7 @@ export function parseTiersFromExpr(exprStr) {
   if (!exprStr) return [];
   try {
     const { body } = stripExprVersion(exprStr);
-    const condGroup = `((?:(?:p|c)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)(?:\\s*&&\\s*(?:p|c)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)*)`;
+    const condGroup = `((?:(?:p|c|len)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)(?:\\s*&&\\s*(?:p|c|len)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)*)`;
     const tierRe = new RegExp(`(?:${condGroup}\\s*\\?\\s*)?tier\\("([^"]*)",\\s*([^)]+)\\)`, 'g');
     const tiers = [];
     let m;
@@ -2255,7 +2255,7 @@ export function parseTiersFromExpr(exprStr) {
       const conditions = [];
       if (condStr) {
         for (const cp of condStr.split(/\s*&&\s*/)) {
-          const cm = cp.trim().match(/^(p|c)\s*(<|<=|>|>=)\s*([\d.eE+]+)$/);
+          const cm = cp.trim().match(/^(p|c|len)\s*(<|<=|>|>=)\s*([\d.eE+]+)$/);
           if (cm) conditions.push({ var: cm[1], op: cm[2], value: Number(cm[3]) });
         }
       }
@@ -2293,7 +2293,7 @@ export function renderTieredModelPrice(opts) {
   const { symbol, rate } = getCurrencyConfig();
   const gr = groupRatio || 1;
 
-  const priceLines = BILLING_VARS.map((v) => [v.field, v.label]);
+  const priceLines = BILLING_PRICING_VARS.map((v) => [v.field, v.label]);
 
   const lines = [
     buildBillingText('命中档位：{{tier}}', { tier: matchedTier || tier.label }),
@@ -2334,7 +2334,7 @@ export function renderTieredModelPriceSimple(opts) {
     ];
 
     if (tier && isPriceDisplayMode(displayMode)) {
-      const priceSegments = BILLING_VARS.map((v) => [v.field, v.shortLabel]);
+      const priceSegments = BILLING_PRICING_VARS.map((v) => [v.field, v.shortLabel]);
       for (const [field, label] of priceSegments) {
         if (tier[field] > 0) {
           segments.push({

web/src/helpers/utils.jsx

@@ -18,7 +18,7 @@ For commercial licensing, please contact support@quantumnous.com
 */
 
 import { Toast, Pagination } from '@douyinfe/semi-ui';
-import { toastConstants, BILLING_VARS, BILLING_VAR_REGEX } from '../constants';
+import { toastConstants, BILLING_PRICING_VARS, BILLING_VAR_REGEX } from '../constants';
 import React from 'react';
 import { toast } from 'react-toastify';
 import {
@@ -927,7 +927,7 @@ export const formatDynamicPriceSummary = (billingExpr, t, groupRatio = 1) => {
   }
   const hasCoeffs = 'p' in varCoeffs || 'c' in varCoeffs;
 
-  const varLabels = BILLING_VARS.map((v) => [v.key, v.label]);
+  const varLabels = BILLING_PRICING_VARS.map((v) => [v.key, v.label]);
 
   const hasTimeCondition = /\b(?:hour|minute|weekday|month|day)\(/.test(exprBody);
   const hasRequestCondition = /\b(?:param|header)\(/.test(exprBody);

web/src/pages/Setting/Ratio/components/TieredPricingEditor.jsx

@@ -31,9 +31,10 @@ import {
   TextArea,
   Typography,
 } from '@douyinfe/semi-ui';
-import { IconDelete, IconPlus } from '@douyinfe/semi-icons';
+import { IconCopy, IconDelete, IconPlus } from '@douyinfe/semi-icons';
 import { renderQuota } from '../../../../helpers/render';
-import { BILLING_EXTRA_VARS, BILLING_CACHE_VAR_MAP } from '../../../../constants';
+import { copy, showSuccess } from '../../../../helpers';
+import { BILLING_EXTRA_VARS, BILLING_CACHE_VAR_MAP, BILLING_CONDITION_VARS } from '../../../../constants';
 import {
   createEmptyCondition,
   createEmptyTimeCondition,
@@ -70,6 +71,7 @@ function priceToUnitCost(price) {
 
 const OPS = ['<', '<=', '>', '>='];
 const VAR_OPTIONS = [
+  { value: 'len', label: 'len (长度)' },
   { value: 'p', label: 'p (输入)' },
   { value: 'c', label: 'c (输出)' },
 ];
@@ -224,7 +226,7 @@ function tryParseVisualConfig(exprStr) {
     }
 
     // Multi-tier: cond1 ? tier(body) : cond2 ? tier(body) : tier(body)
-    const condGroup = `((?:(?:p|c)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)(?:\\s*&&\\s*(?:p|c)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)*)`;
+    const condGroup = `((?:(?:p|c|len)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)(?:\\s*&&\\s*(?:p|c|len)\\s*(?:<|<=|>|>=)\\s*[\\d.eE+]+)*)`;
     const tierRe = new RegExp(
       `(?:${condGroup}\\s*\\?\\s*)?tier\\("([^"]*)",\\s*${bodyPat}\\)`,
       'g',
@@ -237,7 +239,7 @@ function tryParseVisualConfig(exprStr) {
       if (condStr) {
         const condParts = condStr.split(/\s*&&\s*/);
         for (const cp of condParts) {
-          const cm = cp.trim().match(/^(p|c)\s*(<|<=|>|>=)\s*([\d.eE+]+)$/);
+          const cm = cp.trim().match(/^(p|c|len)\s*(<|<=|>|>=)\s*([\d.eE+]+)$/);
           if (cm) {
             conditions.push({ var: cm[1], op: cm[2], value: Number(cm[3]) });
           }
@@ -283,7 +285,7 @@ function ConditionRow({ cond, onChange, onRemove, t }) {
     }}>
       <Select
         size='small'
-        value={cond.var || 'p'}
+        value={cond.var || 'len'}
         onChange={(val) => onChange({ ...cond, var: val })}
       >
         {VAR_OPTIONS.map((v) => (
@@ -500,7 +502,7 @@ function ExtendedPriceBlock({ tier, index, onUpdate, t }) {
 function VisualTierCard({ tier, index, isLast, isOnly, onUpdate, onRemove, t }) {
   const conditions = tier.conditions || [];
 
-  const varLabel = { p: t('输入'), c: t('输出') };
+  const varLabel = { len: t('长度'), p: t('输入'), c: t('输出') };
   const condSummary = useMemo(() => {
     if (conditions.length === 0) return t('无条件（兜底档）');
     return conditions
@@ -525,7 +527,7 @@ function VisualTierCard({ tier, index, isLast, isOnly, onUpdate, onRemove, t })
   const addCondition = () => {
     if (conditions.length >= 2) return;
     const usedVars = conditions.map((c) => c.var);
-    const nextVar = usedVars.includes('p') ? 'c' : 'p';
+    const nextVar = usedVars.includes('len') ? 'c' : 'len';
     onUpdate(index, 'conditions', [
       ...conditions,
       { var: nextVar, op: '<', value: 200000 },
@@ -694,7 +696,7 @@ function VisualEditor({ visualConfig, onChange, t }) {
     ) {
       newTiers[newTiers.length - 1] = {
         ...newTiers[newTiers.length - 1],
-        conditions: [{ var: 'p', op: '<', value: 200000 }],
+        conditions: [{ var: 'len', op: '<', value: 200000 }],
       };
     }
     newTiers.push({
@@ -723,7 +725,7 @@ function VisualEditor({ visualConfig, onChange, t }) {
     <div>
       <Banner
         type='info'
-        description={t('每个档位可设置 0~2 个条件（对 p 和 c），最后一档为兜底档无需条件。')}
+        description={t('每个档位可设置 0~2 个条件（对 len、p 和 c），最后一档为兜底档无需条件。len 为输入上下文总长度（含缓存），推荐用于阶梯条件。')}
         style={{ marginBottom: 12 }}
       />
 
@@ -762,16 +764,16 @@ const PRESET_GROUPS = [
     presets: [
       { key: 'flat', label: 'Flat', expr: 'tier("base", p * 2 + c * 4)' },
       { key: 'claude-opus', label: 'Claude Opus 4.6', expr: 'tier("base", p * 5 + c * 25 + cr * 0.5 + cc * 6.25 + cc1h * 10)' },
-      { key: 'gpt-5.4', label: 'GPT-5.4', expr: 'p <= 272000 ? tier("standard", p * 2.5 + c * 15 + cr * 0.25) : tier("long_context", p * 5 + c * 22.5 + cr * 0.5)' },
+      { key: 'gpt-5.4', label: 'GPT-5.4', expr: 'len <= 272000 ? tier("standard", p * 2.5 + c * 15 + cr * 0.25) : tier("long_context", p * 5 + c * 22.5 + cr * 0.5)' },
     ],
   },
   {
     group: '阶梯计费',
     presets: [
-      { key: 'claude-sonnet', label: 'Claude Sonnet 4.5', expr: 'p <= 200000 ? tier("standard", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6) : tier("long_context", p * 6 + c * 22.5 + cr * 0.6 + cc * 7.5 + cc1h * 12)' },
-      { key: 'qwen3-max', label: 'Qwen3 Max', expr: 'p <= 32000 ? tier("short", p * 1.2 + c * 6 + cr * 0.24 + cc * 1.5) : p <= 128000 ? tier("mid", p * 2.4 + c * 12 + cr * 0.48 + cc * 3) : tier("long", p * 3 + c * 15 + cr * 0.6 + cc * 3.75)' },
-      { key: 'glm-4.5-air', label: 'GLM-4.5 Air', expr: 'p < 32000 && c < 200 ? tier("short_output", p * 0.8 + c * 2 + cr * 0.16) : p < 32000 && c >= 200 ? tier("long_output", p * 0.8 + c * 6 + cr * 0.16) : tier("mid_context", p * 1.2 + c * 8 + cr * 0.24)' },
-      { key: 'doubao-seed-1.8', label: 'Doubao Seed 1.8', expr: 'p <= 32000 && c <= 200 ? tier("discount", p * 0.8 + c * 2 + cr * 0.16 + cc * 0.17) : p <= 32000 ? tier("short", p * 0.8 + c * 8 + cr * 0.16 + cc * 0.17) : p <= 128000 ? tier("mid", p * 1.2 + c * 16 + cr * 0.16 + cc * 0.17) : tier("long", p * 2.4 + c * 24 + cr * 0.16 + cc * 0.17)' },
+      { key: 'claude-sonnet', label: 'Claude Sonnet 4.5', expr: 'len <= 200000 ? tier("standard", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6) : tier("long_context", p * 6 + c * 22.5 + cr * 0.6 + cc * 7.5 + cc1h * 12)' },
+      { key: 'qwen3-max', label: 'Qwen3 Max', expr: 'len <= 32000 ? tier("short", p * 1.2 + c * 6 + cr * 0.24 + cc * 1.5) : len <= 128000 ? tier("mid", p * 2.4 + c * 12 + cr * 0.48 + cc * 3) : tier("long", p * 3 + c * 15 + cr * 0.6 + cc * 3.75)' },
+      { key: 'glm-4.5-air', label: 'GLM-4.5 Air', expr: 'len < 32000 && c < 200 ? tier("short_output", p * 0.8 + c * 2 + cr * 0.16) : len < 32000 && c >= 200 ? tier("long_output", p * 0.8 + c * 6 + cr * 0.16) : tier("mid_context", p * 1.2 + c * 8 + cr * 0.24)' },
+      { key: 'doubao-seed-1.8', label: 'Doubao Seed 1.8', expr: 'len <= 32000 && c <= 200 ? tier("discount", p * 0.8 + c * 2 + cr * 0.16 + cc * 0.17) : len <= 32000 ? tier("short", p * 0.8 + c * 8 + cr * 0.16 + cc * 0.17) : len <= 128000 ? tier("mid", p * 1.2 + c * 16 + cr * 0.16 + cc * 0.17) : tier("long", p * 2.4 + c * 24 + cr * 0.16 + cc * 0.17)' },
     ],
   },
   {
@@ -793,7 +795,7 @@ const PRESET_GROUPS = [
       },
       {
         key: 'gpt-5.4-tiers', label: 'GPT-5.4 Priority/Flex',
-        expr: 'p <= 272000 ? tier("standard", p * 2.5 + c * 15 + cr * 0.25) : tier("long_context", p * 5 + c * 22.5 + cr * 0.5)',
+        expr: 'len <= 272000 ? tier("standard", p * 2.5 + c * 15 + cr * 0.25) : tier("long_context", p * 5 + c * 22.5 + cr * 0.5)',
         requestRules: [
           { conditions: [{ source: SOURCE_PARAM, path: 'service_tier', mode: MATCH_EQ, value: 'priority' }], multiplier: '2' },
           { conditions: [{ source: SOURCE_PARAM, path: 'service_tier', mode: MATCH_EQ, value: 'flex' }], multiplier: '0.5' },
@@ -880,7 +882,8 @@ function RawExprEditor({ exprString, onChange, t }) {
           <div>
             <div>
               {t('变量')}: <code>p</code> ({t('输入 Token')}), <code>c</code> (
-              {t('输出 Token')}), <code>cr</code> ({t('缓存读取')}),{' '}
+              {t('输出 Token')}), <code>len</code> ({t('输入长度')}),{' '}
+              <code>cr</code> ({t('缓存读取')}),{' '}
               <code>cc</code> ({t('缓存创建')}),{' '}
               <code>cc1h</code> ({t('缓存创建-1小时')})
             </div>
@@ -968,7 +971,11 @@ function evalExprLocally(exprStr, p, c, extraTokenValues) {
       matchedTier = name;
       return value;
     };
-    const env = { p, c, tier: tierFn, max: Math.max, min: Math.min, abs: Math.abs, ceil: Math.ceil, floor: Math.floor };
+    const cacheReadTokens = extraTokenValues.cacheReadTokens || 0;
+    const cacheCreateTokens = extraTokenValues.cacheCreateTokens || 0;
+    const cacheCreate1hTokens = extraTokenValues.cacheCreate1hTokens || 0;
+    const len = p + cacheReadTokens + cacheCreateTokens + cacheCreate1hTokens;
+    const env = { p, c, len, tier: tierFn, max: Math.max, min: Math.min, abs: Math.abs, ceil: Math.ceil, floor: Math.floor };
     for (const field of EXTRA_ESTIMATOR_FIELDS) {
       env[field.var] = extraTokenValues[field.stateKey] || 0;
     }
@@ -1220,6 +1227,146 @@ function RuleGroupCard({ group, index, onChange, onRemove, t }) {
   );
 }
 
+// ---------------------------------------------------------------------------
+// LLM prompt helper — copyable prompt for LLM-assisted expression design
+// ---------------------------------------------------------------------------
+
+const LLM_PROMPT_TEMPLATE = `你是一个 AI API 计费表达式设计助手。用户需要你帮忙设计一个计费表达式（billing expression），用于 AI API 网关的模型计费。
+
+## 表达式语言
+
+表达式基于 expr-lang/expr，支持标准算术运算和三元运算符。
+
+### Token 变量
+
+输入侧：
+- p — 输入 token 数（计价用）。系统会自动排除表达式中单独计价的子类别（如用了 cr，缓存 token 就从 p 中扣除）
+- len — 输入上下文总长度（条件判断用）。不受自动排除影响，始终反映完整输入长度。用于阶梯条件判断
+- cr — 缓存命中（读取）token 数
+- cc — 缓存创建 token 数（5分钟 TTL）
+- cc1h — 缓存创建 token 数（1小时 TTL，Claude 专用）
+- img — 图片输入 token 数
+- ai — 音频输入 token 数
+
+输出侧：
+- c — 输出 token 数。同样会自动排除单独计价的子类别
+- img_o — 图片输出 token 数
+- ao — 音频输出 token 数
+
+### p/c 自动排除机制
+
+p 和 c 是兜底变量，代表所有没有被表达式单独定价的 token。如果表达式使用了某个子类别变量（如 cr），对应 token 就从 p 中扣除，避免重复计费。没用到的子类别 token 则留在 p/c 中按基础价格计费。
+
+重要：len 不受自动排除影响。阶梯条件应使用 len 而非 p，以避免缓存命中导致 p 降低而误判档位。
+
+### 内置函数
+
+- tier(name, value) — 标记计费档位名称，必须包裹费用表达式
+- max(a, b)、min(a, b) — 取大/小值
+- ceil(x)、floor(x)、abs(x) — 向上取整、向下取整、绝对值
+- header(name) — 读取请求头
+- param(path) — 读取请求体 JSON 路径（gjson 语法）
+- has(source, substr) — 子字符串检查
+- hour(tz)、minute(tz)、weekday(tz)、month(tz)、day(tz) — 时间函数，tz 为时区如 "Asia/Shanghai"
+
+### 价格系数
+
+表达式中的数字系数是 $/1M tokens 的价格。例如 p * 2.5 表示输入 $2.50/1M tokens。
+
+## 表达式示例
+
+简单定价：
+tier("base", p * 2.5 + c * 15)
+
+带缓存的定价：
+tier("base", p * 2.5 + c * 15 + cr * 0.25)
+
+多档阶梯（用 len 做条件）：
+len <= 200000
+  ? tier("standard", p * 3 + c * 15 + cr * 0.3 + cc * 3.75 + cc1h * 6)
+  : tier("long_context", p * 6 + c * 22.5 + cr * 0.6 + cc * 7.5 + cc1h * 12)
+
+图片模型：
+tier("base", p * 2 + c * 8 + img * 2.5)
+
+多模态含音频：
+tier("base", p * 0.43 + c * 3.06 + img * 0.78 + ai * 3.81 + ao * 15.11)
+
+三档阶梯示例：
+len <= 128000
+  ? tier("standard", p * 1.1 + c * 4.4)
+  : (len <= 1000000
+    ? tier("medium", p * 2.2 + c * 8.8)
+    : tier("long", p * 4.4 + c * 17.6))
+
+## 规则
+
+1. 每个叶子分支必须用 tier("名称", 费用表达式) 包裹
+2. tier 名称用英文，如 "base"、"standard"、"long_context"
+3. 阶梯条件用 len（不要用 p），支持 <、<=、>、>=
+4. 多档用嵌套三元运算符：条件1 ? tier(...) : (条件2 ? tier(...) : tier(...))
+5. 价格系数直接写供应商官方 $/1M tokens 价格
+6. 不需要缓存/图片/音频单独定价时可以不写对应变量，它们的 token 会自动包含在 p/c 中
+
+请根据用户提供的模型信息和定价需求，生成计费表达式。`;
+
+function LlmPromptHelper({ t, model }) {
+  const [open, setOpen] = useState(false);
+
+  const modelName = model?.name || '';
+  const prompt = useMemo(() => {
+    if (modelName) {
+      return LLM_PROMPT_TEMPLATE + `\n\n当前模型：${modelName}`;
+    }
+    return LLM_PROMPT_TEMPLATE;
+  }, [modelName]);
+
+  const handleCopy = useCallback(async () => {
+    const ok = await copy(prompt);
+    if (ok) showSuccess(t('已复制到剪贴板'));
+  }, [prompt, t]);
+
+  return (
+    <div style={{ marginBottom: 12 }}>
+      <Button
+        theme='borderless'
+        size='small'
+        icon={<IconCopy />}
+        onClick={() => setOpen(!open)}
+        style={{ color: 'var(--semi-color-tertiary)' }}
+      >
+        {t('LLM 辅助设计提示词')}
+      </Button>
+      <Collapsible isOpen={open}>
+        <Card
+          bodyStyle={{ padding: 12 }}
+          style={{ marginTop: 8, background: 'var(--semi-color-fill-0)' }}
+        >
+          <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'center', marginBottom: 8 }}>
+            <Text size='small' type='secondary'>
+              {t('复制以下提示词发送给 LLM（如 ChatGPT / Claude），让它帮你设计计费表达式')}
+            </Text>
+            <Button
+              icon={<IconCopy />}
+              size='small'
+              theme='light'
+              onClick={handleCopy}
+            >
+              {t('复制提示词')}
+            </Button>
+          </div>
+          <TextArea
+            value={prompt}
+            readonly
+            autosize={{ minRows: 6, maxRows: 20 }}
+            style={{ fontFamily: 'monospace', fontSize: 12 }}
+          />
+        </Card>
+      </Collapsible>
+    </div>
+  );
+}
+
 // ---------------------------------------------------------------------------
 // Main component
 // ---------------------------------------------------------------------------
@@ -1543,6 +1690,8 @@ export default function TieredPricingEditor({ model, onExprChange, requestRuleEx
         </div>
       </Card>
 
+      <LlmPromptHelper t={t} model={model} />
+
     </div>
   );
 }