Skip to content

Commit e289456

Browse files
fix(console): openai usage normalization and tier threshold config (#47342)
Co-authored-by: Jack <[email protected]>
1 parent 70b4ca8 commit e289456

5 files changed

Lines changed: 81 additions & 6 deletions

File tree

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

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1023,7 +1023,8 @@ export async function handler(
10231023
modelInfo.costPeak && isPeakPricing(new Date())
10241024
? modelInfo.costPeak
10251025
: modelInfo.cost200K &&
1026-
inputTokens + (cacheReadTokens ?? 0) + (cacheWrite5mTokens ?? 0) + (cacheWrite1hTokens ?? 0) > 200_000
1026+
inputTokens + (cacheReadTokens ?? 0) + (cacheWrite5mTokens ?? 0) + (cacheWrite1hTokens ?? 0) >
1027+
modelInfo.cost200K.threshold
10271028
? modelInfo.cost200K
10281029
: modelInfo.cost
10291030

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

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -51,7 +51,10 @@ export const openaiHelper: ProviderHelper = ({ workspaceID }) => ({
5151
const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? undefined
5252
const cacheWriteTokens = usage.input_tokens_details?.cache_write_tokens ?? undefined
5353
return {
54-
inputTokens: inputTokens - (cacheReadTokens ?? 0),
54+
// OpenAI's input_tokens includes both cached_tokens and cache_write_tokens;
55+
// each is billed separately at its own rate. Clamp to zero so a provider
56+
// reporting overlapping detail fields cannot drive input cost negative.
57+
inputTokens: Math.max(0, inputTokens - (cacheReadTokens ?? 0) - (cacheWriteTokens ?? 0)),
5558
outputTokens,
5659
reasoningTokens,
5760
cacheReadTokens,

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

Lines changed: 19 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -26,7 +26,7 @@ describe("provider usage extraction", () => {
2626

2727
expect(providers.google.normalizeUsage(usage)).toEqual({
2828
inputTokens: 6,
29-
outputTokens: 3,
29+
outputTokens: 5,
3030
reasoningTokens: 2,
3131
cacheReadTokens: 4,
3232
cacheWrite5mTokens: undefined,
@@ -42,7 +42,7 @@ describe("provider usage extraction", () => {
4242

4343
expect(providers.google.normalizeUsage(usageParser.retrieve())).toEqual({
4444
inputTokens: 6,
45-
outputTokens: 3,
45+
outputTokens: 5,
4646
reasoningTokens: 2,
4747
cacheReadTokens: 4,
4848
cacheWrite5mTokens: undefined,
@@ -73,7 +73,23 @@ describe("provider usage extraction", () => {
7373
)
7474

7575
expect(providers.openai.normalizeUsage(usageParser.retrieve())).toEqual({
76-
inputTokens: 6,
76+
inputTokens: 3,
77+
outputTokens: 2,
78+
reasoningTokens: undefined,
79+
cacheReadTokens: 4,
80+
cacheWrite5mTokens: 3,
81+
cacheWrite1hTokens: undefined,
82+
})
83+
})
84+
85+
test("clamps input tokens when detail fields overlap", () => {
86+
const usageParser = providers.openai.createUsageParser()
87+
usageParser.parse(
88+
'event: response.completed\ndata: {"response":{"usage":{"input_tokens":5,"input_tokens_details":{"cached_tokens":4,"cache_write_tokens":3},"output_tokens":2}}}',
89+
)
90+
91+
expect(providers.openai.normalizeUsage(usageParser.retrieve())).toEqual({
92+
inputTokens: 0,
7793
outputTokens: 2,
7894
reasoningTokens: undefined,
7995
cacheReadTokens: 4,

‎packages/console/core/src/model.ts‎

Lines changed: 7 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -19,11 +19,17 @@ export namespace ZenData {
1919
cacheWrite1h: z.number().optional(),
2020
})
2121

22+
// Long-context tier. The flip threshold defaults to 200_000 for backward
23+
// compatibility with existing ZEN_MODELS secrets.
24+
const ModelCostTierSchema = ModelCostSchema.extend({
25+
threshold: z.number().default(200_000),
26+
})
27+
2228
const ModelSchema = z.object({
2329
name: z.string(),
2430
cost: ModelCostSchema,
2531
costMultiplier: z.number().default(1),
26-
cost200K: ModelCostSchema.optional(),
32+
cost200K: ModelCostTierSchema.optional(),
2733
costPeak: ModelCostSchema.optional(),
2834
allowAnonymous: z.boolean().optional(),
2935
byokProvider: z.enum(["openai", "anthropic", "google"]).optional(),
Lines changed: 49 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,49 @@
1+
import { describe, expect, test } from "bun:test"
2+
import { ZenData } from "../src/model"
3+
4+
const base = {
5+
zenModels: {
6+
"gpt-5.6-sol": {
7+
name: "GPT-5.6 Sol",
8+
cost: { input: 2, output: 10 },
9+
costMultiplier: 1,
10+
providers: [{ id: "openai", model: "gpt-5.6-sol" }],
11+
},
12+
},
13+
liteModels: {},
14+
providers: { openai: { api: "https://api.openai.com/v1", apiKey: "test" } },
15+
}
16+
17+
const entry = (data: ReturnType<typeof ZenData.validate>) => {
18+
const value = data.zenModels["gpt-5.6-sol"]
19+
return Array.isArray(value) ? value[0] : value
20+
}
21+
22+
describe("ZenData cost200K threshold", () => {
23+
test("defaults to 200_000 when not configured", () => {
24+
const data = ZenData.validate({
25+
...base,
26+
zenModels: {
27+
"gpt-5.6-sol": {
28+
...base.zenModels["gpt-5.6-sol"],
29+
// The secret arrives as parsed JSON in production, so untyped here too.
30+
cost200K: JSON.parse('{"input":4,"output":15}'),
31+
},
32+
},
33+
})
34+
expect(entry(data).cost200K?.threshold).toBe(200_000)
35+
})
36+
37+
test("accepts an explicit 272_000 threshold", () => {
38+
const data = ZenData.validate({
39+
...base,
40+
zenModels: {
41+
"gpt-5.6-sol": {
42+
...base.zenModels["gpt-5.6-sol"],
43+
cost200K: { input: 4, output: 15, threshold: 272_000 },
44+
},
45+
},
46+
})
47+
expect(entry(data).cost200K?.threshold).toBe(272_000)
48+
})
49+
})

0 commit comments

Comments
 (0)