chore(gui): 版本号 1.9.3 → 1.9.4

## 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>

ARCHIVE.md

@@ -1,32 +0,0 @@
-# archive/pre-v1.9-quota-refactor
-
-**这是一个归档分支，不要从它继续开发。**
-
-它冻结的是 v1.9 之前的一段插件重构历史，保存于 2026-10-05。
-
-## 为什么保留
-
-分支上有 16 个提交没有以等价形式进入主线。逐个 patch-id 核对后，其中 **3 个
-是远端完全没有的**：
-
-| commit | 内容 | 为什么没进主线 |
-|---|---|---|
-| `a6c6e6e` | 可见性、state 分区、注册式 HTTP 端点 | 后继版本 `35b9422` 是它的 rebase 等价版，但这个 hash 上还挂着另一些只存在于此分支的改动 |
-| `6bc3ff8` | ccquota 显示每月额度 | ccquota 后来被通用 `quota` 插件取代，这段额度 UI 没进主线 |
-| `b5f6cac` | ccquota 页面首次打开是空的 | 同上，是 ccquota 生命周期内的修复 |
-
-其余 13 个已在主线以等价 patch 存在（`35b9422` / `cab3361` / `ba10f03` 等），
-它们保留在这里是为了让这段历史可追溯，不是为了再次使用。
-
-## 现在的功能落点
-
-| 当时在这里 | 现在在哪 |
-|---|---|
-| ccquota 多账户额度插件 | `internal/lua/plugins/quota.lua`（通用套餐余量插件） |
-| 可见性 / state 分区 / 按密钥授权 | `internal/lua/plugin_access.go`（`AccessRule`） |
-| 注册式 HTTP 端点 | `internal/lua/plugin_http_handler.go` + `AccessRule` |
-
-## 当前发布线
-
-- `main` — 下一个未发布版本
-- `release/v1.9.x` — v1.9.x 发布线（最新 tag `v1.9.5`）

cmd/gui/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.0",
+  "version": "1.9.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "modelrouter-gui",
-      "version": "1.9.0",
+      "version": "1.9.3",
       "license": "Proprietary",
       "devDependencies": {
         "electron": "^33.0.0",

cmd/gui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "modelrouter-gui",
-  "version": "1.9.0",
+  "version": "1.9.4",
   "description": "ModelRouter Desktop — embedded ModelRouter core with tray",
   "author": "ModelRouter",
   "main": "main.js",

internal/billing/compile.go

@@ -195,6 +195,36 @@ func applyRule(prices map[string]interface{}, rule *config.BillingRule, srcName
 			if t.CacheDiscount != nil {
 				entry["cache_discount"] = *t.CacheDiscount
 			}
+			// reasoning is optional and stays ABSENT (not nil, not 0) when the
+			// rule does not set it, which is what makes the billing plugin skip
+			// the repricing branch entirely. Writing an explicit 0 here would
+			// mean "thinking is free" and silently under-bill every request
+			// from a provider that bills it like ordinary output.
+			if strings.TrimSpace(t.Reasoning) != "" {
+				rp, err := usdPerM(t.Reasoning)
+				if err != nil {
+					return fmt.Errorf("model %q reasoning: %w", modelID, err)
+				}
+				entry["reasoning_price"] = rp
+			}
+			if lc := t.LongContext; lc != nil {
+				if lc.Above <= 0 {
+					return fmt.Errorf("model %q long_context: above must be > 0", modelID)
+				}
+				above, err := longContextRate(lc, prompt)
+				if err != nil {
+					return fmt.Errorf("model %q long_context: %w", modelID, err)
+				}
+				entry["long_context"] = map[string]interface{}{
+					"above": lc.Above,
+					// Marginal, not blended. The plugin charges the tokens past the
+					// threshold at this rate and the rest at the base rate, so the
+					// threshold composes with a cache discount (which applies to the
+					// cached portion) instead of forcing the plugin to guess how
+					// the two interact.
+					"prompt": above,
+				}
+			}
 			if rule.Peak != nil {
 				entry["peak"] = compilePeak(rule.Peak)
 			}
@@ -262,3 +292,32 @@ func compilePeak(p *config.BillingPeak) map[string]interface{} {
 	}
 	return out
 }
+
+// longContextRate resolves the marginal per-token rate for the tokens ABOVE a
+// threshold, given the model's base per-million prompt price.
+//
+// Two ways to express it are supported because providers publish both: an
+// absolute rate (OpenAI quotes \$3.00/M past 200k) and a multiple (DeepSeek
+// quotes exactly 2x past 128k). Neither may be missing: a step with no rate is a
+// silent no-op, and an operator who wrote one meant something.
+func longContextRate(lc *config.BillingLongContext, basePerToken float64) (float64, error) {
+	hasPrompt := strings.TrimSpace(lc.Prompt) != ""
+	hasMult := lc.Multiplier > 0
+	switch {
+	case hasPrompt && hasMult:
+		p, err := usdPerM(lc.Prompt)
+		if err != nil {
+			return 0, err
+		}
+		// Both given: the absolute rate wins, because it is the one the provider
+		// actually quotes. Silently multiplying instead would produce a number
+		// the operator never wrote.
+		return p, nil
+	case hasPrompt:
+		return usdPerM(lc.Prompt)
+	case hasMult:
+		return basePerToken * lc.Multiplier, nil
+	default:
+		return 0, fmt.Errorf("needs either prompt or multiplier")
+	}
+}

internal/billing/wildcard_test.go

@@ -86,3 +86,129 @@ func TestSpecificURLTokenRuleStillMarksItsSource(t *testing.T) {
 		t.Error("an unrelated source was marked priced by a specific-url rule")
 	}
 }
+
+// TestReasoningPriceIsOmittedUnlessSet pins the absence convention for the
+// thinking-token rate.
+//
+// A rule that says nothing about reasoning must produce a price entry with NO
+// reasoning_price key, so the billing plugin skips the repricing branch and
+// bills thinking at the ordinary output rate — which is what every provider
+// actually charges. Writing an explicit 0 instead would read as "thinking is
+// free" and silently under-bill every request from that model.
+func TestReasoningPriceIsOmittedUnlessSet(t *testing.T) {
+	sources := []config.Source{{Name: "up", BaseURL: "https://up.example/v1"}}
+
+	// Case 1: no reasoning declared → the key must not appear.
+	p1 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+		URL: "https://up.example/v1", Mode: "token",
+		Models: map[string]config.BillingToken{
+			"m": {Prompt: "1.00", Completion: "2.00"},
+		},
+	}}}
+	prices, err := Compile(p1, sources)
+	if err != nil {
+		t.Fatal(err)
+	}
+	entry := prices["models"].(map[string]interface{})["m"].(map[string]interface{})
+	if _, present := entry["reasoning_price"]; present {
+		t.Errorf("reasoning_price present (%v) without a rule; the plugin would treat 0 as free",
+			entry["reasoning_price"])
+	}
+
+	// Case 2: declared → present, and converted to per-token like every other rate.
+	p2 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+		URL: "https://up.example/v1", Mode: "token",
+		Models: map[string]config.BillingToken{
+			"m": {Prompt: "1.00", Completion: "2.00", Reasoning: "6.00"},
+		},
+	}}}
+	prices2, err := Compile(p2, sources)
+	if err != nil {
+		t.Fatal(err)
+	}
+	e2 := prices2["models"].(map[string]interface{})["m"].(map[string]interface{})
+	rp, ok := e2["reasoning_price"].(float64)
+	if !ok {
+		t.Fatalf("reasoning_price missing or wrong type: %#v", e2["reasoning_price"])
+	}
+	if rp != 6e-6 {
+		t.Errorf("reasoning_price = %v, want 6e-6 (USD per TOKEN)", rp)
+	}
+	// A malformed rate must be rejected, not silently ignored: an operator who
+	// typed "6 USD" expects either that value or an error, never a rule that
+	// looks applied and bills nothing.
+	p3 := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+		URL: "https://up.example/v1", Mode: "token",
+		Models: map[string]config.BillingToken{
+			"m": {Prompt: "1.00", Completion: "2.00", Reasoning: "six dollars"},
+		},
+	}}}
+	if _, err := Compile(p3, sources); err == nil {
+		t.Error("a non-numeric reasoning rate must fail the compile")
+	}
+}
+
+// TestLongContextStepCompiles pins the long-context step's two accepted forms
+// and its rejection of a step that would silently do nothing.
+//
+// Providers publish both shapes — DeepSeek says "2x past 128k", OpenAI quotes
+// an absolute "$3.00/M past 200k" — so both must compile. Neither may be
+// missing: a step with no rate is a rule that LOOKS like it handles long
+// context and does not, which under-bills without any visible sign.
+func TestLongContextStepCompiles(t *testing.T) {
+	sources := []config.Source{{Name: "up", BaseURL: "https://up.example/v1"}}
+
+	compile := func(lc config.BillingLongContext) (map[string]interface{}, error) {
+		p := &config.BillingProfile{ID: "p", Rules: []config.BillingRule{{
+			URL: "https://up.example/v1", Mode: "token",
+			Models: map[string]config.BillingToken{
+				"m": {Prompt: "1.00", Completion: "2.00", LongContext: &lc},
+			},
+		}}}
+		prices, err := Compile(p, sources)
+		if err != nil {
+			return nil, err
+		}
+		e := prices["models"].(map[string]interface{})["m"].(map[string]interface{})
+		return e["long_context"].(map[string]interface{}), nil
+	}
+
+	// Absolute rate (OpenAI's published shape).
+	lc1, err := compile(config.BillingLongContext{Above: 200000, Prompt: "3.00"})
+	if err != nil {
+		t.Fatalf("absolute long_context rate rejected: %v", err)
+	}
+	if lc1["above"] != 200000 || lc1["prompt"] != 3e-6 {
+		t.Errorf("absolute step = %#v, want above=200000 prompt=3e-6", lc1)
+	}
+
+	// Multiplier (DeepSeek's shape): 2x of the 1.00/M base.
+	lc2, err := compile(config.BillingLongContext{Above: 128000, Multiplier: 2})
+	if err != nil {
+		t.Fatalf("multiplier long_context rejected: %v", err)
+	}
+	if lc2["prompt"] != 2e-6 {
+		t.Errorf("multiplier step = %#v, want prompt=2e-6", lc2)
+	}
+
+	// Both given: the absolute rate wins, because that is the number the
+	// provider quotes. Multiplying instead would produce a figure the operator
+	// never wrote.
+	lc3, err := compile(config.BillingLongContext{Above: 128000, Prompt: "5.00", Multiplier: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if lc3["prompt"] != 5e-6 {
+		t.Errorf("both given = %#v, want the quoted absolute rate 5e-6", lc3)
+	}
+
+	// Neither: a no-op step must fail, not compile into a rule that silently
+	// bills long context at the base rate.
+	if _, err := compile(config.BillingLongContext{Above: 128000}); err == nil {
+		t.Error("a step with neither prompt nor multiplier must be rejected")
+	}
+	// A zero threshold would make every request "long".
+	if _, err := compile(config.BillingLongContext{Above: 0, Multiplier: 2}); err == nil {
+		t.Error("above=0 must be rejected")
+	}
+}

internal/config/config.go

@@ -453,6 +453,19 @@ type BillingToken struct {
 	Prompt        string   `yaml:"prompt" json:"prompt"`         // USD per million tokens, e.g. "0.15"
 	Completion    string   `yaml:"completion" json:"completion"` // USD per million tokens
 	CacheDiscount *float64 `yaml:"cache_discount,omitempty" json:"cache_discount,omitempty"`
+	// Reasoning is the USD-per-million rate for thinking tokens, for the rare
+	// provider that charges them differently from ordinary output. Omit it and
+	// reasoning is billed at Completion — which is what OpenAI, DeepSeek and
+	// Anthropic all do, and what every source on this deployment reports.
+	//
+	// It is a rate for a SUBSET of completion tokens, never an extra charge: the
+	// billing plugin applies only the difference between this and Completion, so
+	// 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.
@@ -464,6 +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 {

internal/gateway/chat.go

@@ -789,6 +789,18 @@ func recordChatUsage(rec *Req, req *types.ChatRequest, resp *types.UnifiedRespon
 	// Cache accounting: whichever source format the adapter normalized into
 	// (prompt_tokens_details.cached_tokens or legacy prompt_cache_hit_tokens),
 	// read it back so the request record carries the hit/miss split.
+	// reasoning_tokens is a SUBSET of completion_tokens and is billed at the
+	// completion rate by every provider measured, so it must NOT be added to
+	// Compl. Recording it separately is what lets a price rule split thinking
+	// from output for the providers that do price them differently, and what
+	// keeps the billing plugin from guessing.
+	//
+	// Leaving it unrecorded is worse than useless: the adapters already parse it
+	// and forward it (opencodego/gemini/anthropic all do), so the number exists
+	// in the response and is thrown away at the one place that could keep it.
+	if d := resp.TokenUsage.CompletionTokensDetails; d != nil && d.ReasoningTokens > 0 {
+		rec.Reasoning = int64(d.ReasoningTokens)
+	}
 	if d := resp.TokenUsage.PromptTokensDetails; d != nil && d.CachedTokens > 0 {
 		rec.CacheHit = int64(d.CachedTokens)
 	}
@@ -1044,6 +1056,9 @@ func (g *Gateway) fireEnd(rec *Req) {
 		// carried this (rec.CacheReported); the plugin payload did not, so any
 		// plugin-level cache accounting had to guess.
 		"cache_reported": rec.CacheReported,
+		// A SUBSET of completion_tokens, not an extra. Billing plugins that
+		// add it to completion would double-charge every thinking token.
+		"reasoning_tokens": rec.Reasoning,
 		"image_count":    rec.ImageCount,
 		"error":          rec.Err,
 		"time":           rec.Time,

internal/gateway/chat_reasoning_test.go

@@ -0,0 +1,71 @@
+package gateway
+
+import (
+	"testing"
+
+	"llmsproxy/internal/types"
+)
+
+// TestRecordChatUsageKeepsReasoningTokens pins the reasoning-token plumbing.
+//
+// The adapters already parse completion_tokens_details.reasoning_tokens and
+// forward it (opencodego, gemini, anthropic all do), and types.go has carried
+// the field since it was defined — but nothing ever READ it. The number arrived
+// in the response and was discarded at the one place that could have kept it,
+// so a pricing rule had no way to split thinking from output.
+//
+// The property that matters most: reasoning is a SUBSET of completion. Every
+// provider measured bills it at the completion rate, so a billing plugin that
+// added it to completion_tokens would double-charge every thinking token. The
+// test asserts completion is left alone.
+func TestRecordChatUsageKeepsReasoningTokens(t *testing.T) {
+	rec := &Req{}
+	resp := &types.UnifiedResponse{
+		TokenUsage: types.TokenUsage{
+			Prompt:     100,
+			Completion: 500,
+			CompletionTokensDetails: &types.CompletionTokensDetails{
+				ReasoningTokens: 320,
+			},
+		},
+	}
+	recordChatUsage(rec, &types.ChatRequest{}, resp)
+
+	if rec.Reasoning != 320 {
+		t.Errorf("Reasoning = %d, want 320", rec.Reasoning)
+	}
+	// The subset must not be folded into completion — it is already counted there.
+	if rec.Compl != 500 {
+		t.Errorf("Compl = %d, want 500 (reasoning is a subset, never additive)", rec.Compl)
+	}
+	if rec.Prompt != 100 {
+		t.Errorf("Prompt = %d, want 100", rec.Prompt)
+	}
+}
+
+// TestRecordChatUsageReasoningAbsentStaysZero: "not reported" and "reported as
+// zero" must both leave the field at 0, and neither may invent a number. A
+// plugin reading Reasoning cannot tell them apart, which is why the field is
+// documented as "0 means not reported" rather than "0 means no thinking".
+func TestRecordChatUsageReasoningAbsentStaysZero(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		u    types.TokenUsage
+	}{
+		{"no details object", types.TokenUsage{Prompt: 10, Completion: 20}},
+		{"details without reasoning", types.TokenUsage{
+			Prompt:                  10,
+			Completion:              20,
+			CompletionTokensDetails: &types.CompletionTokensDetails{},
+		}},
+	} {
+		rec := &Req{}
+		recordChatUsage(rec, &types.ChatRequest{}, &types.UnifiedResponse{TokenUsage: tc.u})
+		if rec.Reasoning != 0 {
+			t.Errorf("%s: Reasoning = %d, want 0", tc.name, rec.Reasoning)
+		}
+		if rec.Compl != 20 {
+			t.Errorf("%s: Compl = %d, want 20", tc.name, rec.Compl)
+		}
+	}
+}

internal/gateway/stats.go

@@ -43,6 +43,16 @@ type Req struct {
 	// accounting at all (even when the hit count is 0). The WebUI shows
 	// "0%" instead of "—" for such rows.
 	CacheReported bool `json:"cache_reported,omitempty"`
+	// Reasoning is completion_tokens_details.reasoning_tokens.
+	//
+	// It is a SUBSET of Compl, not an addition: reasoning tokens are billed at
+	// the completion rate and already counted there. This field exists so a
+	// pricing rule can split them when a provider prices thinking tokens
+	// differently (most do NOT — OpenAI and DeepSeek bill them as ordinary
+	// output), and so a reader can see what fraction of a bill is thinking.
+	// 0 means "not reported", which is NOT the same as "no reasoning happened" —
+	// three of the sources measured return no such field at all.
+	Reasoning int64 `json:"reasoning_tokens,omitempty"`
 	OK            bool `json:"ok"`
 	// Status http status code
 	Status int `json:"status"`

internal/lua/billing_test.go

@@ -994,3 +994,245 @@ func TestBillingCacheHitClampedInStats(t *testing.T) {
 			tot["cache_fresh_tokens"])
 	}
 }
+
+// TestUnpricedIsTrackedPerDay pins the day-keyed unpriced tally.
+//
+// The lifetime counters cannot answer the question an operator actually has —
+// "is it STILL happening?" — because they keep the debt from before a rule was
+// added. After pricing a mispriced source the lifetime number never moves, so a
+// dashboard reading "unpriced: 6376" is indistinguishable from a system that is
+// still billing nothing. Day buckets are what make a pricing fix observable: the
+// offending day keeps its count and the next day starts at zero for it.
+func TestUnpricedIsTrackedPerDay(t *testing.T) {
+	ps, _ := billingVM(t)
+	// Nothing is priced in the default config, so every request is unpriced.
+	// Two requests on two different UTC days.
+	for _, ms := range []int64{1790000000000, 1790100000000} {
+		ps.Fire(StageRequestEnd, map[string]interface{}{
+			"model": "mystery-model", "source": "nowhere", "key": "***aaaaaa",
+			"ok": true, "prompt_tokens": 10, "completion_tokens": 5,
+			"time": ms,
+		})
+	}
+	st := stateOf(t, ps)
+	byDay, ok := st["unpriced_by_day"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("state has no unpriced_by_day map: %T", st["unpriced_by_day"])
+	}
+	if len(byDay) != 2 {
+		t.Fatalf("expected 2 day buckets, got %d: %v", len(byDay), byDay)
+	}
+	var sum float64
+	for day, bucket := range byDay {
+		b := bucket.(map[string]interface{})
+		sum += b["reqs"].(float64)
+		models := b["models"].(map[string]interface{})
+		if models["mystery-model"].(float64) != 1 {
+			t.Errorf("%s: model tally = %v, want 1", day, models["mystery-model"])
+		}
+	}
+	if sum != 2 {
+		t.Errorf("day buckets sum to %v, want 2 (the lifetime counter must not be double-counted)", sum)
+	}
+	// The lifetime counters still exist and still agree — this is additive, not
+	// a replacement.
+	if st["unpriced_reqs"].(float64) != 2 {
+		t.Errorf("unpriced_reqs = %v, want 2", st["unpriced_reqs"])
+	}
+}
+
+// TestReasoningBilledAtCompletionRateByDefault is the property that keeps the
+// standard formula matching a real invoice.
+//
+// reasoning_tokens is a subset of completion_tokens and every provider measured
+// (OpenAI, DeepSeek, Anthropic) charges it at the ordinary output rate. A rule
+// with no reasoning_price must therefore produce EXACTLY the number the
+// pre-reasoning formula produced — the field exists for a provider that prices
+// thinking differently, not as a new charge.
+func TestReasoningBilledAtCompletionRateByDefault(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				// USD PER TOKEN (compile.go divides the per-million figure by 1e6).
+				// prompt 0.15/M, completion 0.60/M. No reasoning_price anywhere.
+				"m": map[string]interface{}{"prompt": 0.15e-6, "completion": 0.60e-6},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000,
+		"reasoning_tokens": 800, // all of it thinking; still billed as output
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	approx(t, "reasoning at completion rate", cost, 0.00075)
+}
+
+// TestReasoningPriceOverridesOnlyWhenSet: with a rule that DOES price thinking
+// differently, only the difference is applied. Charging the reasoning portion a
+// second time in full is the obvious wrong implementation and it doubles the
+// completion line.
+func TestReasoningPriceOverridesOnlyWhenSet(t *testing.T) {
+	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": 0.15e-6, "completion": 0.60e-6,
+					// Thinking costs 3x ordinary output: 1.80/M.
+					"reasoning_price": 1.80e-6,
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000, "reasoning_tokens": 800,
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	// base = 1000*0.15e-6 + 1000*0.60e-6 = 0.00075
+	//      + 800 * (1.80-0.60)e-6   = 0.00096  →  0.00171
+	approx(t, "reasoning repriced to 1.80/M", cost, 0.00171)
+}
+
+// TestReasoningClampedToCompletion: a provider claiming MORE reasoning than
+// completion is malformed. Without a clamp the visible-output count goes
+// negative and the delta exceeds the whole request, so the bill silently
+// explodes on bad data instead of degrading.
+func TestReasoningClampedToCompletion(t *testing.T) {
+	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": 0.15e-6, "completion": 0.60e-6, "reasoning_price": 1.80e-6,
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 1000, "completion_tokens": 1000, "reasoning_tokens": 99999,
+		"time": 1750000000000,
+	})
+	st := stateOf(t, ps)
+	cost := st["total"].(map[string]interface{})["cost"].(float64)
+	// reasoning is clamped from 99999 down to completion (1000):
+	//   base  = 1000*0.15e-6 + 1000*0.60e-6            = 0.00075
+	//   delta = 1000 * (1.80-0.60)e-6                  = 0.00120
+	//   total                                          = 0.00195
+	// Without the clamp this would be 0.18 — 92x the whole request, and a
+	// 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)
+}

internal/lua/billing_ui_test.go

@@ -414,3 +414,86 @@ func TestRuleTableLocksColumnWidthsInAColgroup(t *testing.T) {
 			"control and must never be squeezed by the flexible inputs")
 	}
 }
+
+// TestBillingUIShowsRecentUnpriced: the page must be able to distinguish a
+// settled pricing debt from a live leak. The lifetime unpriced counter answers
+// neither — it keeps every request from before a rule was added, so it reads the
+// same number in both situations. The day buckets exist to make the difference
+// visible, and a bucket nothing renders is a bucket nobody benefits from.
+func TestBillingUIShowsRecentUnpriced(t *testing.T) {
+	raw, err := os.ReadFile("plugins/billing.lua")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(raw)
+	for _, needle := range []string{
+		"unpriced_by_day",  // the tally is written
+		"recentUnpriced",   // it is reduced to a window
+		"unpriced_recent",  // it reaches the view
+		"unpriced7d",       // the card labels it
+		"unpricedClean",    // and says so when the window is clear
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("billing UI lacks %q", needle)
+		}
+	}
+}
+
+// TestBillingRulesUIExposesReasoningPrice: the thinking-token rate must be
+// editable in the rules editor, not only settable by hand-editing config.yaml.
+// A pricing capability that lives only in the config file is one the operator
+// will assume does not exist.
+func TestBillingRulesUIExposesReasoningPrice(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-r'`,       // the input, in both the existing rows and the "+ model" row
+		`models[n].reasoning`, // collected back out of the DOM on save
+		`reasoningHint`,     // and it says what leaving it blank means
+	} {
+		if !strings.Contains(src, needle) {
+			t.Errorf("rules editor lacks %q", needle)
+		}
+	}
+	// Two occurrences: an existing model's row and the template for a new one.
+	// One would mean the "+ model" button produces a row that silently drops
+	// the rate on the next save.
+	if n := strings.Count(src, `class='m-r'`); n < 2 {
+		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)
+	}
+}

internal/lua/plugins/billing.lua

@@ -192,6 +192,11 @@ local function priceFor(payload)
   out = {
     prompt = d.prompt or 0, completion = d.completion or 0,
     per_request = 0, cache_discount = d.cache_discount, peak = d.peak,
+      -- nil = bill reasoning at the completion rate, which is what every
+      -- 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)
@@ -202,6 +207,8 @@ local function priceFor(payload)
     if mp.completion ~= nil then out.completion = mp.completion end
     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
 
@@ -213,6 +220,8 @@ local function priceFor(payload)
     if sp.completion ~= nil then out.completion = sp.completion end
     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
 
@@ -346,6 +355,60 @@ 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
+    -- the term above. This block only matters when a provider charges thinking
+    -- at a DIFFERENT rate — and every provider measured does not: OpenAI,
+    -- DeepSeek and Anthropic all bill reasoning at the ordinary output rate. So
+    -- reasoning_price is nil by default and this is skipped entirely, which is
+    -- what makes the standard formula match the invoice.
+    --
+    -- When it IS set, only the DIFFERENCE is applied: the reasoning portion is
+    -- re-priced from the normal rate to the reasoning rate. Charging it a
+    -- second time in full would double-bill every thinking token.
+    local reasoningPrice = tonumber(price.reasoning_price)
+    if reasoningPrice and reasoningPrice ~= price.completion then
+      local reasoning = tonumber(payload.reasoning_tokens) or 0
+      if reasoning > 0 then
+        -- Defensive clamp: a provider reporting more reasoning than completion
+        -- would otherwise produce a negative visible-output count and a cost
+        -- delta larger than the whole request.
+        if reasoning > completion then reasoning = completion end
+        cost = cost + reasoning * (reasoningPrice - price.completion) * mult
+      end
+    end
+
   local flat = price.per_request
   if not payload.ok and not plugin.count_failures then
     flat = 0
@@ -442,6 +505,21 @@ function plugin.on_request_end(payload)
     s.unpriced_reqs = s.unpriced_reqs + 1
     local m = payload.model or "?"
     s.unpriced_models[m] = (s.unpriced_models[m] or 0) + 1
+
+  -- Day-keyed unpriced tally. The lifetime counters above cannot answer the
+  -- question an operator actually has — "is it STILL happening?" — because they
+  -- keep the debt from before a pricing rule was added. After pricing a source
+  -- that was mispriced, the lifetime number never moves, so "unpriced: 6376" on
+  -- a dashboard is indistinguishable from a system that is still billing
+  -- nothing. Counting by day makes the fix observable: the offending day keeps
+  -- its number, new days start at zero.
+  if s.unpriced_by_day == nil then s.unpriced_by_day = {} end
+  -- dayKey wants epoch SECONDS; payload.time is milliseconds.
+  local uday = dayKey(math.floor((tonumber(payload.time) or (os.time() * 1000)) / 1000))
+  local ubucket = s.unpriced_by_day[uday]
+  if ubucket == nil then ubucket = { reqs = 0, models = {} }; s.unpriced_by_day[uday] = ubucket end
+  ubucket.reqs = ubucket.reqs + 1
+  ubucket.models[m] = (ubucket.models[m] or 0) + 1
   end
   if s.degraded_reqs == nil then s.degraded_reqs = 0 end
   -- Degradation is counted here rather than in the chain_step hook because
@@ -517,7 +595,7 @@ plugin.ui = {
     global = {
       "total", "by_source", "by_model", "by_day", "by_day_src",
       "by_day_model", "by_tier_served", "skip_reasons", "degraded_reqs",
-      "unpriced_models", "unpriced_reqs", "started",
+      "unpriced_models", "unpriced_reqs", "unpriced_by_day", "started",
     },
     depth = { by_key = 1, by_day_key = 2 },
   },
@@ -593,6 +671,7 @@ plugin.ui = {
     en: {
       total: "Total", requests: "Requests", degraded: "Degraded",
       unpriced: "Unpriced", prompt: "Prompt tokens",
+      unpriced7d: "Unpriced (7d)", unpricedClean: "none in last 7 days",
       completion: "Completion tokens", failures: "Failures",
       cacheRate: "Cache hit rate", cacheTokens: "Cache read tokens",
       perSource: "Per source", perModel: "Per model",
@@ -614,6 +693,7 @@ plugin.ui = {
     zh: {
       total: "总开销", requests: "请求数", degraded: "降级",
       unpriced: "未定价", prompt: "输入 tokens",
+      unpriced7d: "未定价（近 7 天）", unpricedClean: "近 7 天无",
       completion: "输出 tokens", failures: "失败",
       cacheRate: "缓存命中率", cacheTokens: "缓存读取 tokens",
       perSource: "按源", perModel: "按模型",
@@ -734,13 +814,35 @@ plugin.ui = {
     });
     return out;
   }
+  // recentUnpriced counts unpriced requests over the last n days. When the
+  // result is zero THAT is the useful signal: every pricing rule added since
+  // then is holding. Null when there is no day data at all (a state file
+  // written before the buckets existed) so the UI can say "unknown" rather than
+  // claim a clean record it cannot support.
+  function recentUnpriced(st, n) {
+    var byDay = st && st.unpriced_by_day;
+    if (!byDay) return null;
+    var cutoff = Date.now() - n * 86400000;
+    var sum = 0, seen = false;
+    Object.keys(byDay).forEach(function (d) {
+      var t = Date.parse(d + "T00:00:00Z");
+      if (isNaN(t)) return;
+      seen = true;
+      if (t >= cutoff) sum += (byDay[d].reqs || 0);
+    });
+    return seen ? sum : null;
+  }
+
   // rescale returns the state as the selected period presents it. Degraded and
   // unpriced counters are lifetime-only (there is no day-keyed tally for them)
   // so they are returned as null and the UI hides them instead of printing a
   // lifetime number beside a per-window cost — the exact mismatch the period
   // view exists to remove.
   function rescale(st) {
-    if (PERIOD === "all" || !st.by_day) return st;
+    if (PERIOD === "all" || !st.by_day) {
+      if (st) st.unpriced_recent = recentUnpriced(st, 7);
+      return st;
+    }
     var days = daysInWindow(PERIOD, utcToday());
     var total = zeroBucket();
     var byDay = {};
@@ -758,6 +860,10 @@ plugin.ui = {
       by_day: byDay,
       degraded_reqs: null,
       unpriced_reqs: null,
+      // The lifetime unpriced total is null here for the same reason, but the
+      // day buckets are not — they are what lets the card report "0 in the last
+      // 7 days" instead of leaving a settled debt indistinguishable from a leak.
+      unpriced_recent: recentUnpriced(st, 7),
     };
     return out;
   }
@@ -838,6 +944,15 @@ plugin.ui = {
       // print a lifetime number next to a per-window cost.
       view.degraded_reqs != null ? [T.degraded, fmtInt(view.degraded_reqs)] : null,
       view.unpriced_reqs != null ? [T.unpriced, fmtInt(view.unpriced_reqs)] : null,
+      // "Unpriced (7d)" answers what the lifetime number cannot: is it STILL
+      // happening? A settled debt and a live leak print the same lifetime total.
+      view.unpriced_recent != null
+        ? [
+            T.unpriced7d,
+            fmtInt(view.unpriced_recent),
+            view.unpriced_recent === 0 ? T.unpricedClean : null,
+          ]
+        : null,
       [T.prompt, fmtInt(t.prompt_tokens || 0)],
       [T.completion, fmtInt(t.completion_tokens || 0)],
       [T.failures, t.failures || 0],
@@ -949,7 +1064,11 @@ plugin.ui = {
       monthly: "Monthly fee", peakMult: "Peak multiplier",
       peakHours: "Peak hours (UTC)", peakDays: "Peak days (UTC 1=Mon)",
       models: "Per-model prices", addModel: "+ model",
-      mFree: "free", mToken: "token", mSub: "subscription", mUnpriced: "unpriced",
+
+      reasoning: "think $/M",
+
+      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",
@@ -971,7 +1090,11 @@ plugin.ui = {
       monthly: "月费", peakMult: "峰段倍数",
       peakHours: "峰段小时 (UTC)", peakDays: "峰段星期 (UTC 1=周一)",
       models: "按模型价格", addModel: "+ 模型",
-      mFree: "free 免费", mToken: "token 按量",
+
+      reasoning: "思考 $/M",
+
+      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",
@@ -1069,6 +1192,25 @@ plugin.ui = {
         "<input class='m-name' value='" + esc(m) + "' style='flex:1 1 90px;min-width:0'>" +
         "<input class='m-p' value='" + esc(t.prompt == null ? "" : t.prompt) +
           "' placeholder='" + esc(T.prompt) + "' style='flex:0 1 62px;min-width:0'>" +
+          // Left BLANK by default on purpose: empty means "bill thinking at the
+          // completion rate", which is what OpenAI, DeepSeek and Anthropic do.
+          // A 0 here would mean "thinking is free" and would under-bill quietly.
+          "<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'>&times;</button></div>";
@@ -1145,6 +1287,39 @@ plugin.ui = {
           return;
         }
         models[n] = { prompt: pv, completion: cv };
+
+          // Optional thinking-token rate, read from the same row. Blank means
+          // "bill reasoning at the completion rate" — what OpenAI, DeepSeek and
+          // Anthropic all do. A 0 would mean "thinking is free", so an empty
+          // field must stay ABSENT from the saved rule rather than becoming 0.
+          var rel = r.querySelector(".m-r");
+          var rtxt = rel ? String(rel.value || "").trim() : "";
+          if (rtxt !== "") {
+            if (isNaN(parseFloat(rtxt))) {
+              bad = n + " (" + T.reasoning + ")";
+              return;
+            }
+            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;
@@ -1335,6 +1510,11 @@ plugin.ui = {
           "<input class='m-name' placeholder='model' style='flex:1;min-width:80px'>" +
           "<input class='m-p' placeholder='" + esc(T.prompt) + "' style='width:96px'>" +
           "<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'>&times;</button>";
         box.appendChild(row);
       };