diff --git a/apps/server/src/provider/Drivers/claudeUsage.test.ts b/apps/server/src/provider/Drivers/claudeUsage.test.ts index 1db2197a53e55..baf587efa1a7d 100644 --- a/apps/server/src/provider/Drivers/claudeUsage.test.ts +++ b/apps/server/src/provider/Drivers/claudeUsage.test.ts @@ -22,6 +22,7 @@ function claudeLine(overrides: { usage: { input_tokens: 2, cache_creation_input_tokens: 66818, + cache_creation: { ephemeral_5m_input_tokens: 818, ephemeral_1h_input_tokens: 66000 }, cache_read_input_tokens: 1000, output_tokens: overrides.outputTokens ?? 286, ...(overrides.speed === undefined ? {} : { speed: overrides.speed }), @@ -44,6 +45,7 @@ describe("parseClaudeLine", () => { outputTokens: 286, reasoningTokens: 0, }); + expect(record?.oneHourCacheCreationTokens).toBe(66000); expect(record?.dedupeKey).toBe("msg_1:"); expect(record?.speed).toBe("standard"); }); diff --git a/apps/server/src/provider/Drivers/claudeUsage.ts b/apps/server/src/provider/Drivers/claudeUsage.ts index c5ff71a5c839e..38d115c845f65 100644 --- a/apps/server/src/provider/Drivers/claudeUsage.ts +++ b/apps/server/src/provider/Drivers/claudeUsage.ts @@ -56,6 +56,15 @@ function parseClaudeRecord(parsed: unknown): UsageRecord | null { messageId === null && requestId === null ? null : `${messageId ?? ""}:${requestId ?? ""}`; const cost = record["costUSD"]; + const cacheCreationTokens = tokenCount(usageRecord["cache_creation_input_tokens"]); + const cacheCreation = usageRecord["cache_creation"]; + const oneHourCacheCreationTokens = + typeof cacheCreation === "object" && cacheCreation !== null + ? Math.min( + tokenCount((cacheCreation as Record)["ephemeral_1h_input_tokens"]), + cacheCreationTokens, + ) + : 0; return { provider: "claude", @@ -65,11 +74,12 @@ function parseClaudeRecord(parsed: unknown): UsageRecord | null { totals: { uncachedInputTokens: tokenCount(usageRecord["input_tokens"]), cachedInputTokens: tokenCount(usageRecord["cache_read_input_tokens"]), - cacheCreationTokens: tokenCount(usageRecord["cache_creation_input_tokens"]), + cacheCreationTokens, outputTokens: tokenCount(usageRecord["output_tokens"]), // Anthropic folds thinking tokens into output and does not break them out. reasoningTokens: 0, }, + ...(oneHourCacheCreationTokens > 0 ? { oneHourCacheCreationTokens } : {}), reportedCostUsd: typeof cost === "number" && Number.isFinite(cost) ? cost : null, speed: usageRecord["speed"] === "fast" ? "fast" : "standard", dedupeKey, diff --git a/apps/server/src/usage/UsageService.test.ts b/apps/server/src/usage/UsageService.test.ts index 75de5ec25a4ec..f5350cb82b706 100644 --- a/apps/server/src/usage/UsageService.test.ts +++ b/apps/server/src/usage/UsageService.test.ts @@ -1971,7 +1971,7 @@ describe("UsageService", () => { 60 + 20 * (version === 4 ? 1 : 2), ); yield* upgraded.awaitPersisted; - assert.strictEqual(SCAN_CACHE_FILE_NAME, "usage-scan-cache-v8.json"); + assert.strictEqual(SCAN_CACHE_FILE_NAME, "usage-scan-cache-v9.json"); assert.strictEqual( yield* Effect.promise(() => NodeFSP.readFile(legacyPath, "utf8")), legacy, diff --git a/apps/server/src/usage/usageAggregation.test.ts b/apps/server/src/usage/usageAggregation.test.ts index 08b3f3ae1b8c8..e959fb044eba4 100644 --- a/apps/server/src/usage/usageAggregation.test.ts +++ b/apps/server/src/usage/usageAggregation.test.ts @@ -12,11 +12,13 @@ const rates: RateTable = new Map([ outputCostPerToken: 5e-5, cacheReadCostPerToken: 1e-6, cacheCreationCostPerToken: 1.25e-5, + oneHourCacheCreationCostPerToken: 2e-5, fast: { inputCostPerToken: 2e-5, outputCostPerToken: 1e-4, cacheReadCostPerToken: 2e-6, cacheCreationCostPerToken: 2.5e-5, + oneHourCacheCreationCostPerToken: 4e-5, }, ultrafast: null, }, diff --git a/apps/server/src/usage/usagePricing.test.ts b/apps/server/src/usage/usagePricing.test.ts index 3b8b62d48a6ee..5270b41dcab16 100644 --- a/apps/server/src/usage/usagePricing.test.ts +++ b/apps/server/src/usage/usagePricing.test.ts @@ -171,6 +171,31 @@ describe("usage pricing", () => { }); }); + it("prices one-hour cache writes at the one-hour write rate", () => { + const table = parseRateTable({ + "claude-opus-5-5": { + ...rate(4e-6, 4e-7), + cache_creation_input_token_cost: 5e-6, + cache_creation_input_token_cost_above_1hr: 8e-6, + input_cost_per_token_priority: 8e-6, + output_cost_per_token_priority: 40e-6, + cache_creation_input_token_cost_priority: 12e-6, + }, + }); + const writes = (speed: UsageSpeed) => ({ + ...record("claude-opus-5-5", null, speed), + oneHourCacheCreationTokens: 600_000, + }); + + expect(priceUsage(table, writes("standard"))).toMatchObject({ + categoryCostUsd: { cacheWrite: expect.closeTo(0.4 * 5 + 0.6 * 8) }, + }); + // The priority tier publishes no one-hour rate, so it keeps the standard 8/5 premium. + expect(priceUsage(table, writes("fast"))).toMatchObject({ + categoryCostUsd: { cacheWrite: expect.closeTo(0.4 * 12 + 0.6 * 12 * 1.6) }, + }); + }); + it("prices Codex priority and ultrafast requests at their published tier rates", () => { const table = parseRateTable({ "gpt-6-astra": { diff --git a/apps/server/src/usage/usagePricing.ts b/apps/server/src/usage/usagePricing.ts index 50d839f8175c4..2c47b6e11817b 100644 --- a/apps/server/src/usage/usagePricing.ts +++ b/apps/server/src/usage/usagePricing.ts @@ -2,7 +2,6 @@ import type { UsageCategoryCost, UsageCostSource, UsageModelPriceOverride, - UsageTokenTotals, } from "@t3tools/contracts"; import type { UsageRecord, UsageSpeed } from "@t3tools/provider-core/server/usage"; @@ -12,6 +11,7 @@ export interface TokenRates { readonly outputCostPerToken: number; readonly cacheReadCostPerToken: number; readonly cacheCreationCostPerToken: number; + readonly oneHourCacheCreationCostPerToken: number; } /** @@ -38,6 +38,8 @@ export function createOverrideRateTable( (prices.cacheReadCostPerMillionTokens ?? prices.inputCostPerMillionTokens) / 1_000_000, cacheCreationCostPerToken: (prices.cacheWriteCostPerMillionTokens ?? prices.inputCostPerMillionTokens) / 1_000_000, + oneHourCacheCreationCostPerToken: + (prices.cacheWriteCostPerMillionTokens ?? prices.inputCostPerMillionTokens) / 1_000_000, fast: null, ultrafast: null, }, @@ -68,14 +70,20 @@ function readTokenRates( (standard !== undefined && standard.inputCostPerToken > 0 ? (standard[field] / standard.inputCostPerToken) * input : input); + const cacheCreation = cacheRate("cache_creation_input_token_cost", "cacheCreationCostPerToken"); + // Without its own one-hour rate, a tier keeps the standard tier's one-hour premium over writes. + const oneHourCacheCreation = + finiteNumber(entry[`cache_creation_input_token_cost_above_1hr${suffix}`]) ?? + (standard !== undefined && standard.cacheCreationCostPerToken > 0 + ? (standard.oneHourCacheCreationCostPerToken / standard.cacheCreationCostPerToken) * + cacheCreation + : cacheCreation); return { inputCostPerToken: input, outputCostPerToken: output, cacheReadCostPerToken: cacheRate("cache_read_input_token_cost", "cacheReadCostPerToken"), - cacheCreationCostPerToken: cacheRate( - "cache_creation_input_token_cost", - "cacheCreationCostPerToken", - ), + cacheCreationCostPerToken: cacheCreation, + oneHourCacheCreationCostPerToken: oneHourCacheCreation, }; } @@ -85,6 +93,7 @@ function scaleTokenRates(rates: TokenRates, multiple: number): TokenRates { outputCostPerToken: rates.outputCostPerToken * multiple, cacheReadCostPerToken: rates.cacheReadCostPerToken * multiple, cacheCreationCostPerToken: rates.cacheCreationCostPerToken * multiple, + oneHourCacheCreationCostPerToken: rates.oneHourCacheCreationCostPerToken * multiple, }; } @@ -146,7 +155,8 @@ function sameTokenRates(a: TokenRates | null, b: TokenRates | null): boolean { a.inputCostPerToken === b.inputCostPerToken && a.outputCostPerToken === b.outputCostPerToken && a.cacheReadCostPerToken === b.cacheReadCostPerToken && - a.cacheCreationCostPerToken === b.cacheCreationCostPerToken + a.cacheCreationCostPerToken === b.cacheCreationCostPerToken && + a.oneHourCacheCreationCostPerToken === b.oneHourCacheCreationCostPerToken ); } @@ -224,7 +234,7 @@ function resolveRate(table: RateTable, model: string): ModelRate | null { export type PricedRecord = Pick< UsageRecord, - "model" | "rateModel" | "totals" | "speed" | "reportedCostUsd" + "model" | "rateModel" | "totals" | "oneHourCacheCreationTokens" | "speed" | "reportedCostUsd" >; export interface PricedUsage { @@ -234,11 +244,15 @@ export interface PricedUsage { readonly speedPremiumUsd: number; } -function costByCategory(totals: UsageTokenTotals, rates: TokenRates): UsageCategoryCost { +function costByCategory(record: PricedRecord, rates: TokenRates): UsageCategoryCost { + const { totals } = record; + const oneHourWrites = record.oneHourCacheCreationTokens ?? 0; return { input: totals.uncachedInputTokens * rates.inputCostPerToken, cacheRead: totals.cachedInputTokens * rates.cacheReadCostPerToken, - cacheWrite: totals.cacheCreationTokens * rates.cacheCreationCostPerToken, + cacheWrite: + (totals.cacheCreationTokens - oneHourWrites) * rates.cacheCreationCostPerToken + + oneHourWrites * rates.oneHourCacheCreationCostPerToken, output: totals.outputTokens * rates.outputCostPerToken, }; } @@ -256,7 +270,7 @@ export function priceUsage( record: PricedRecord, overrides?: RateTable, ): PricedUsage { - const { model, totals, reportedCostUsd } = record; + const { model, reportedCostUsd } = record; const override = overrides?.get(model.trim()); const reported = override === undefined && reportedCostUsd !== null && Number.isFinite(reportedCostUsd) @@ -273,11 +287,11 @@ export function priceUsage( return reported === null ? unsplit(0, "unpriced") : unsplit(reported, "providerReported"); } - const listCost = costByCategory(totals, ratesAt(rate, record.speed)); + const listCost = costByCategory(record, ratesAt(rate, record.speed)); const listCostUsd = sumCategories(listCost); if (reported !== null && listCostUsd <= 0) return unsplit(reported, "providerReported"); const premiumUsd = - record.speed === "standard" ? 0 : listCostUsd - sumCategories(costByCategory(totals, rate)); + record.speed === "standard" ? 0 : listCostUsd - sumCategories(costByCategory(record, rate)); const scale = reported === null ? 1 : reported / listCostUsd; return { costUsd: reported ?? listCostUsd, diff --git a/apps/server/src/usage/usageScanCache.test.ts b/apps/server/src/usage/usageScanCache.test.ts index 231454d7c1758..56ef71dbfc805 100644 --- a/apps/server/src/usage/usageScanCache.test.ts +++ b/apps/server/src/usage/usageScanCache.test.ts @@ -64,7 +64,15 @@ describe("scan cache round trip", () => { [ "/a.jsonl", 100, - [record(), record({ dedupeKey: "msg_2:", model: "claude-opus-5-5", speed: "fast" })], + [ + record(), + record({ + dedupeKey: "msg_2:", + model: "claude-opus-5-5", + speed: "fast", + oneHourCacheCreationTokens: 6, + }), + ], ], [ "/b.jsonl", @@ -162,7 +170,7 @@ describe("scan cache round trip", () => { expect(restored).toEqual(original); }); - it.each([5, 6, 7])( + it.each([5, 6, 7, 8])( "retains v%s history and compatible positions but invalidates obsolete state", (version) => { const original = cacheWith([["/gone.jsonl", 100, [record({ speed: "fast" })]]]); @@ -213,12 +221,20 @@ describe("scan cache round trip", () => { }; const restored = decodeScanCache(JSON.parse(JSON.stringify(legacy)), TEST_FORMAT_MAP); - expect(restored.get("/gone.jsonl")).toEqual(original.get("/gone.jsonl")); - expect(restored.get("/gone-codex.jsonl")).toEqual({ - ...original.get("/gone-codex.jsonl"), + expect(restored.get("/gone.jsonl")).toEqual({ + ...original.get("/gone.jsonl"), size: -1, position: { resumeOffset: 0, guardLength: 0, guardHash: 0, state: null }, }); + expect(restored.get("/gone-codex.jsonl")).toEqual( + version >= 8 + ? original.get("/gone-codex.jsonl") + : { + ...original.get("/gone-codex.jsonl"), + size: -1, + position: { resumeOffset: 0, guardLength: 0, guardHash: 0, state: null }, + }, + ); expect(restored.get("/gone-pi.jsonl")).toEqual( version >= 6 ? original.get("/gone-pi.jsonl") diff --git a/apps/server/src/usage/usageScanCache.ts b/apps/server/src/usage/usageScanCache.ts index 2cffb633debf3..a6062090c9186 100644 --- a/apps/server/src/usage/usageScanCache.ts +++ b/apps/server/src/usage/usageScanCache.ts @@ -9,7 +9,7 @@ import * as Schema from "effect/Schema"; import { CLAUDE_SYNTHETIC_MODEL } from "../provider/Drivers/claudeUsage.ts"; import { GUARD_LENGTH, type TranscriptParsePosition } from "./usageTranscriptReader.ts"; -const USAGE_SCAN_CACHE_VERSION = 8 as const; +const USAGE_SCAN_CACHE_VERSION = 9 as const; const SPEED_COMPATIBLE_SINCE_VERSION = 4; /** @@ -17,8 +17,9 @@ const SPEED_COMPATIBLE_SINCE_VERSION = 4; * replace it, dropping saved usage for deleted transcripts. Separate files keep * both. */ -export const SCAN_CACHE_FILE_NAME = "usage-scan-cache-v8.json"; +export const SCAN_CACHE_FILE_NAME = "usage-scan-cache-v9.json"; export const LEGACY_SCAN_CACHE_FILE_NAMES = [ + "usage-scan-cache-v8.json", "usage-scan-cache-v7.json", "usage-scan-cache-v6.json", "usage-scan-cache-v5.json", @@ -58,7 +59,8 @@ type SerializedRecord = readonly [ reportedCostUsd: number | null, speed: number, rateModelIndex?: number | null, - unresolvedServiceTierTurnId?: string, + unresolvedServiceTierTurnId?: string | null, + oneHourCacheCreationTokens?: number, ]; interface SerializedFile { @@ -117,9 +119,12 @@ function serializeFile(entry: CachedFile, tables: InternTables): SerializedFile record.rateModel === undefined ? null : intern(tables.models, tables.modelIndex, record.rateModel), - ...(record.unresolvedServiceTierTurnId === undefined - ? ([] as const) - : ([record.unresolvedServiceTierTurnId] as const)), + // Optional trailing slots are written only when set, so most rows stay short. + ...(record.oneHourCacheCreationTokens !== undefined + ? ([record.unresolvedServiceTierTurnId ?? null, record.oneHourCacheCreationTokens] as const) + : record.unresolvedServiceTierTurnId !== undefined + ? ([record.unresolvedServiceTierTurnId] as const) + : ([] as const)), ]; return { s: entry.size, @@ -232,6 +237,7 @@ export function decodeScanCache( speedIndex, rateModelIndex, unresolvedServiceTierTurnId, + oneHourCacheCreation, ] = row as SerializedRecord; const speed = typeof speedIndex === "number" ? SPEEDS[speedIndex] : undefined; @@ -241,8 +247,8 @@ export function decodeScanCache( typeof timestampMs !== "number" || !Number.isFinite(timestampMs) || model === undefined || - (unresolvedServiceTierTurnId !== undefined && - typeof unresolvedServiceTierTurnId !== "string") || + (unresolvedServiceTierTurnId != null && typeof unresolvedServiceTierTurnId !== "string") || + (oneHourCacheCreation !== undefined && !Number.isFinite(oneHourCacheCreation)) || (rateModelIndex != null && rateModel === undefined) || !Number.isFinite(uncached) || !Number.isFinite(cached) || @@ -273,9 +279,12 @@ export function decodeScanCache( }, reportedCostUsd: typeof reportedCostUsd === "number" ? reportedCostUsd : null, speed, - ...(version >= 7 && unresolvedServiceTierTurnId !== undefined + ...(version >= 7 && typeof unresolvedServiceTierTurnId === "string" ? { unresolvedServiceTierTurnId } : {}), + ...(oneHourCacheCreation !== undefined && oneHourCacheCreation > 0 + ? { oneHourCacheCreationTokens: oneHourCacheCreation } + : {}), dedupeKey: typeof dedupeKey === "string" ? dedupeKey : null, }); } @@ -312,8 +321,10 @@ export function decodeScanCache( } // Keep history for deleted rollouts, but re-parse incompatible reducer state. // Older Codex entries lack provenance state (v7) or misread compaction and resume counters (v8). + // Older Claude entries lack the one-hour cache-write split (v9). const legacy = - format.state !== undefined && (version < 6 || (entry.p === "codex" && version < 8)); + (format.state !== undefined && (version < 6 || (entry.p === "codex" && version < 8))) || + (entry.p === "claude" && version < 9); // A corrupt state disqualifies the entry: resuming with it would attach // appended usage to the wrong model or replay fork-copied history. if (!legacy && !isValidState.get(entry.p)?.(entry.cs)) continue; diff --git a/packages/provider-core/src/server/usage.ts b/packages/provider-core/src/server/usage.ts index 47fb344c7c6bd..1f6ca61556c21 100644 --- a/packages/provider-core/src/server/usage.ts +++ b/packages/provider-core/src/server/usage.ts @@ -18,6 +18,8 @@ export interface UsageRecord { readonly rateModel?: string; readonly sessionId: string; readonly totals: UsageTokenTotals; + /** Part of `totals.cacheCreationTokens` written with Anthropic's one-hour TTL, billed higher. */ + readonly oneHourCacheCreationTokens?: number; readonly reportedCostUsd: number | null; readonly speed: UsageSpeed; readonly unresolvedServiceTierTurnId?: string;