fix(console): openai usage normalization and tier threshold config (#47342)

Co-authored-by: Jack <jack@anoma.ly>

packages/console/app/src/routes/zen/util/handler.ts

@@ -1023,7 +1023,8 @@ export async function handler(
       modelInfo.costPeak && isPeakPricing(new Date())
         ? modelInfo.costPeak
         : modelInfo.cost200K &&
-            inputTokens + (cacheReadTokens ?? 0) + (cacheWrite5mTokens ?? 0) + (cacheWrite1hTokens ?? 0) > 200_000
+            inputTokens + (cacheReadTokens ?? 0) + (cacheWrite5mTokens ?? 0) + (cacheWrite1hTokens ?? 0) >
+              modelInfo.cost200K.threshold
           ? modelInfo.cost200K
           : modelInfo.cost
 

packages/console/app/src/routes/zen/util/provider/openai.ts

@@ -51,7 +51,10 @@ export const openaiHelper: ProviderHelper = ({ workspaceID }) => ({
     const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? undefined
     const cacheWriteTokens = usage.input_tokens_details?.cache_write_tokens ?? undefined
     return {
-      inputTokens: inputTokens - (cacheReadTokens ?? 0),
+      // OpenAI's input_tokens includes both cached_tokens and cache_write_tokens;
+      // each is billed separately at its own rate. Clamp to zero so a provider
+      // reporting overlapping detail fields cannot drive input cost negative.
+      inputTokens: Math.max(0, inputTokens - (cacheReadTokens ?? 0) - (cacheWriteTokens ?? 0)),
       outputTokens,
       reasoningTokens,
       cacheReadTokens,

packages/console/app/test/providerUsage.test.ts

@@ -26,7 +26,7 @@ describe("provider usage extraction", () => {
 
     expect(providers.google.normalizeUsage(usage)).toEqual({
       inputTokens: 6,
-      outputTokens: 3,
+      outputTokens: 5,
       reasoningTokens: 2,
       cacheReadTokens: 4,
       cacheWrite5mTokens: undefined,
@@ -42,7 +42,7 @@ describe("provider usage extraction", () => {
 
     expect(providers.google.normalizeUsage(usageParser.retrieve())).toEqual({
       inputTokens: 6,
-      outputTokens: 3,
+      outputTokens: 5,
       reasoningTokens: 2,
       cacheReadTokens: 4,
       cacheWrite5mTokens: undefined,
@@ -73,7 +73,23 @@ describe("provider usage extraction", () => {
     )
 
     expect(providers.openai.normalizeUsage(usageParser.retrieve())).toEqual({
-      inputTokens: 6,
+      inputTokens: 3,
+      outputTokens: 2,
+      reasoningTokens: undefined,
+      cacheReadTokens: 4,
+      cacheWrite5mTokens: 3,
+      cacheWrite1hTokens: undefined,
+    })
+  })
+
+  test("clamps input tokens when detail fields overlap", () => {
+    const usageParser = providers.openai.createUsageParser()
+    usageParser.parse(
+      'event: response.completed\ndata: {"response":{"usage":{"input_tokens":5,"input_tokens_details":{"cached_tokens":4,"cache_write_tokens":3},"output_tokens":2}}}',
+    )
+
+    expect(providers.openai.normalizeUsage(usageParser.retrieve())).toEqual({
+      inputTokens: 0,
       outputTokens: 2,
       reasoningTokens: undefined,
       cacheReadTokens: 4,

packages/console/core/src/model.ts

@@ -19,11 +19,17 @@ export namespace ZenData {
     cacheWrite1h: z.number().optional(),
   })
 
+  // Long-context tier. The flip threshold defaults to 200_000 for backward
+  // compatibility with existing ZEN_MODELS secrets.
+  const ModelCostTierSchema = ModelCostSchema.extend({
+    threshold: z.number().default(200_000),
+  })
+
   const ModelSchema = z.object({
     name: z.string(),
     cost: ModelCostSchema,
     costMultiplier: z.number().default(1),
-    cost200K: ModelCostSchema.optional(),
+    cost200K: ModelCostTierSchema.optional(),
     costPeak: ModelCostSchema.optional(),
     allowAnonymous: z.boolean().optional(),
     byokProvider: z.enum(["openai", "anthropic", "google"]).optional(),

packages/console/core/test/model.test.ts

@@ -0,0 +1,49 @@
+import { describe, expect, test } from "bun:test"
+import { ZenData } from "../src/model"
+
+const base = {
+  zenModels: {
+    "gpt-5.6-sol": {
+      name: "GPT-5.6 Sol",
+      cost: { input: 2, output: 10 },
+      costMultiplier: 1,
+      providers: [{ id: "openai", model: "gpt-5.6-sol" }],
+    },
+  },
+  liteModels: {},
+  providers: { openai: { api: "https://api.openai.com/v1", apiKey: "test" } },
+}
+
+const entry = (data: ReturnType<typeof ZenData.validate>) => {
+  const value = data.zenModels["gpt-5.6-sol"]
+  return Array.isArray(value) ? value[0] : value
+}
+
+describe("ZenData cost200K threshold", () => {
+  test("defaults to 200_000 when not configured", () => {
+    const data = ZenData.validate({
+      ...base,
+      zenModels: {
+        "gpt-5.6-sol": {
+          ...base.zenModels["gpt-5.6-sol"],
+          // The secret arrives as parsed JSON in production, so untyped here too.
+          cost200K: JSON.parse('{"input":4,"output":15}'),
+        },
+      },
+    })
+    expect(entry(data).cost200K?.threshold).toBe(200_000)
+  })
+
+  test("accepts an explicit 272_000 threshold", () => {
+    const data = ZenData.validate({
+      ...base,
+      zenModels: {
+        "gpt-5.6-sol": {
+          ...base.zenModels["gpt-5.6-sol"],
+          cost200K: { input: 4, output: 15, threshold: 272_000 },
+        },
+      },
+    })
+    expect(entry(data).cost200K?.threshold).toBe(272_000)
+  })
+})