tokenusage-cli 0.1.0 → 0.2.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +39 -19
- package/LICENSE +22 -22
- package/README.md +94 -94
- package/THIRD_PARTY_NOTICES.md +140 -140
- package/bin/tokenusage.js +69 -69
- package/package.json +59 -59
- package/src/cli.js +146 -134
- package/src/commands/device-login.js +205 -205
- package/src/commands/diagnostics.js +38 -38
- package/src/commands/doctor.js +110 -110
- package/src/commands/init.js +2484 -2436
- package/src/commands/pause.js +77 -0
- package/src/commands/sessions.js +29 -29
- package/src/commands/status.js +1758 -1747
- package/src/commands/sync.js +8048 -7934
- package/src/commands/uninstall.js +322 -312
- package/src/lib/account-usage-dedup.js +267 -267
- package/src/lib/antigravity-paths.js +48 -48
- package/src/lib/ark-agent-plan-limits.js +438 -438
- package/src/lib/ark-coding-plan-limits.js +537 -537
- package/src/lib/async-queue-reader.js +138 -138
- package/src/lib/atomic-json.js +27 -27
- package/src/lib/bearer-token.js +11 -11
- package/src/lib/categorizer-utils.js +232 -232
- package/src/lib/claude-categorizer.js +1343 -1343
- package/src/lib/claude-config.js +279 -279
- package/src/lib/claude-user-identity.js +56 -56
- package/src/lib/cli-ui.js +179 -179
- package/src/lib/cloud-account.js +541 -541
- package/src/lib/cloud-device-token.js +141 -141
- package/src/lib/cloud-sync-prefs.js +14 -14
- package/src/lib/cloud-usage-reset.js +115 -115
- package/src/lib/codex-config.js +595 -595
- package/src/lib/codex-context-breakdown.js +806 -806
- package/src/lib/codex-model-attribution.js +128 -128
- package/src/lib/codex-rollout-parser.js +1471 -1471
- package/src/lib/codex-service-tier.js +49 -49
- package/src/lib/codex-token-refresh.js +140 -140
- package/src/lib/codex-token-usage.js +195 -195
- package/src/lib/codex-usage-record.js +195 -195
- package/src/lib/collection-pause.js +77 -0
- package/src/lib/command-runner.js +313 -313
- package/src/lib/commandcode-limits.js +317 -317
- package/src/lib/context-health.js +124 -124
- package/src/lib/cursor-config.js +605 -605
- package/src/lib/cursor-store.js +1918 -1918
- package/src/lib/debug-flags.js +7 -7
- package/src/lib/devin-limits.js +249 -249
- package/src/lib/devin-usage.js +212 -212
- package/src/lib/diagnostics.js +383 -383
- package/src/lib/doctor.js +385 -385
- package/src/lib/dumate.js +95 -95
- package/src/lib/fs.js +455 -455
- package/src/lib/function-url.js +14 -14
- package/src/lib/gemini-config.js +291 -291
- package/src/lib/git-outcomes.js +243 -243
- package/src/lib/grok-context-breakdown.js +890 -890
- package/src/lib/grok-hook.js +360 -360
- package/src/lib/grok-limits.js +865 -865
- package/src/lib/grok-usage.js +98 -98
- package/src/lib/heatmap-compact.js +114 -114
- package/src/lib/ingest-sanitize.js +41 -41
- package/src/lib/init-flow.js +42 -42
- package/src/lib/install-resolver.js +170 -170
- package/src/lib/jsonl-lines.js +134 -134
- package/src/lib/kimi-model-repair.js +173 -173
- package/src/lib/kimi-profile.js +61 -61
- package/src/lib/kimi-session.js +216 -216
- package/src/lib/local-api.js +3588 -3588
- package/src/lib/machine-id.js +238 -238
- package/src/lib/multi-install-parser.js +113 -113
- package/src/lib/notify-command.js +38 -38
- package/src/lib/omp-hook.js +589 -589
- package/src/lib/openclaw-cli.js +104 -104
- package/src/lib/openclaw-hook.js +375 -375
- package/src/lib/openclaw-session-plugin.js +560 -560
- package/src/lib/opencode-config.js +121 -121
- package/src/lib/opencode-go-limits.js +807 -807
- package/src/lib/outcomes-engine.js +253 -253
- package/src/lib/passive-mode.js +219 -219
- package/src/lib/pet-packages.js +722 -722
- package/src/lib/pricing/curated-overrides.json +822 -822
- package/src/lib/pricing/index.js +379 -379
- package/src/lib/pricing/litellm-fetcher.js +172 -172
- package/src/lib/pricing/matcher.js +456 -456
- package/src/lib/pricing/seed-snapshot.json +1 -1
- package/src/lib/progress.js +76 -76
- package/src/lib/project-usage-purge.js +117 -117
- package/src/lib/prompt.js +24 -24
- package/src/lib/provider-status.js +104 -104
- package/src/lib/proxy-env.js +489 -489
- package/src/lib/proxy-settings.js +163 -163
- package/src/lib/qoder-cn-cli-limits.js +119 -119
- package/src/lib/qoder-limits.js +880 -880
- package/src/lib/quiet-warnings.js +26 -26
- package/src/lib/rollout.js +26005 -25966
- package/src/lib/runtime-config.js +9 -4
- package/src/lib/scan-roots.js +277 -277
- package/src/lib/scheduled-task.js +145 -0
- package/src/lib/sentry.js +231 -231
- package/src/lib/session-analytics.js +2972 -2972
- package/src/lib/session-performance.js +237 -237
- package/src/lib/session-upload.js +379 -379
- package/src/lib/skill-usage.js +267 -267
- package/src/lib/skills-manager.js +1591 -1591
- package/src/lib/source-metadata.js +48 -48
- package/src/lib/sqlite-reader.js +215 -215
- package/src/lib/star-cta.js +73 -73
- package/src/lib/subscription-manager.js +293 -293
- package/src/lib/subscriptions.js +508 -508
- package/src/lib/sync-report.js +138 -0
- package/src/lib/telemetry.js +193 -193
- package/src/lib/tracker-paths.js +16 -16
- package/src/lib/trae-cn-config.js +548 -548
- package/src/lib/trae-db.js +127 -127
- package/src/lib/trae-snapshot.js +230 -230
- package/src/lib/trae-sqlite.js +143 -143
- package/src/lib/trae-usage.js +121 -121
- package/src/lib/upload-queue.js +134 -134
- package/src/lib/upload-throttle.js +171 -171
- package/src/lib/usage-limits.js +4552 -4552
- package/src/lib/wrapped-aggregator.js +236 -236
- package/src/lib/wsl-probe.js +281 -281
- package/src/lib/zcode-limits.js +1172 -1172
- package/src/lib/zcode-reset-credits.js +66 -66
- package/test/audit-fixes.test.js +200 -0
- package/test/cli-smoke.test.js +53 -53
- package/test/collection-controls.test.js +181 -0
- package/test/ingest-sanitize.test.js +162 -162
- package/test/quiet-warnings.test.js +53 -53
- package/test/session-upload.test.js +427 -427
package/src/lib/pricing/index.js
CHANGED
|
@@ -1,379 +1,379 @@
|
|
|
1
|
-
// Public pricing API. Replaces the hard-coded MODEL_PRICING table that used
|
|
2
|
-
// to live in src/lib/local-api.js. Keeps the same synchronous shape so all
|
|
3
|
-
// existing callers (computeRowCost, /api/functions/* handlers, tests) work
|
|
4
|
-
// unchanged after `await ensurePricingLoaded()` is awaited once at startup.
|
|
5
|
-
|
|
6
|
-
const fs = require("node:fs");
|
|
7
|
-
const path = require("node:path");
|
|
8
|
-
const os = require("node:os");
|
|
9
|
-
|
|
10
|
-
const curatedOverrides = require("./curated-overrides.json");
|
|
11
|
-
const {
|
|
12
|
-
lookupPricing,
|
|
13
|
-
buildLitellmPerMillionMap,
|
|
14
|
-
} = require("./matcher");
|
|
15
|
-
const { loadLitellmData } = require("./litellm-fetcher");
|
|
16
|
-
|
|
17
|
-
const ZERO_PRICING = { input: 0, output: 0, cache_read: 0, cache_write: 0 };
|
|
18
|
-
const PI_SUBSCRIPTION_SOURCES = new Set([
|
|
19
|
-
"pi-github-copilot",
|
|
20
|
-
"pi-copilot",
|
|
21
|
-
"prime-agent-github-copilot",
|
|
22
|
-
"prime-agent-copilot",
|
|
23
|
-
]);
|
|
24
|
-
// LM Studio's server logs describe inference served by the local developer
|
|
25
|
-
// server (including LM Link). Secure Cloud usage has a separate billing path
|
|
26
|
-
// and is not present in these logs.
|
|
27
|
-
const LOCAL_INFERENCE_SOURCES = new Set(["lmstudio"]);
|
|
28
|
-
// OpenAI long-context pricing depends on a single request's raw input,
|
|
29
|
-
// including cached tokens, never a session/day aggregate. Astra supports
|
|
30
|
-
// a larger context window; only observed request subsets receive the premium.
|
|
31
|
-
const OPENAI_LONG_CONTEXT_INPUT_THRESHOLD = 272_000;
|
|
32
|
-
const SOURCES_WITH_AUTHORITATIVE_COST = new Set(["grok", "cline"]);
|
|
33
|
-
const SEED_SNAPSHOT_PATH = path.resolve(__dirname, "seed-snapshot.json");
|
|
34
|
-
const DEEPSEEK_TIME_PRICED_MODELS = [
|
|
35
|
-
"deepseek-v4-flash",
|
|
36
|
-
"deepseek-v4.1-flash",
|
|
37
|
-
"deepseek-flash",
|
|
38
|
-
"deepseek-v4-pro",
|
|
39
|
-
];
|
|
40
|
-
// Sync seed load. Done at require-time so callers that haven't awaited
|
|
41
|
-
// ensurePricingLoaded() (e.g. tests, vite mock startup, edge functions) still
|
|
42
|
-
// get LiteLLM-backed pricing instead of all-zero. ensurePricingLoaded() will
|
|
43
|
-
// later upgrade this to fresh disk cache or upstream data.
|
|
44
|
-
function loadSeedSync() {
|
|
45
|
-
try {
|
|
46
|
-
const raw = fs.readFileSync(SEED_SNAPSHOT_PATH, "utf8");
|
|
47
|
-
const parsed = JSON.parse(raw);
|
|
48
|
-
delete parsed._meta;
|
|
49
|
-
return parsed;
|
|
50
|
-
} catch (e) {
|
|
51
|
-
return {};
|
|
52
|
-
}
|
|
53
|
-
}
|
|
54
|
-
|
|
55
|
-
const seedRaw = loadSeedSync();
|
|
56
|
-
|
|
57
|
-
const state = {
|
|
58
|
-
loaded: false,
|
|
59
|
-
loadingPromise: null,
|
|
60
|
-
revision: 0,
|
|
61
|
-
litellmRawMap: seedRaw, // raw per-token; field shape from LiteLLM JSON
|
|
62
|
-
litellmPerMillionMap: buildLitellmPerMillionMap(seedRaw), // USD/MTok
|
|
63
|
-
source: Object.keys(seedRaw).length ? "seed-snapshot:sync" : null,
|
|
64
|
-
// negativeCache prevents re-walking the LiteLLM map for models we've already
|
|
65
|
-
// determined are unknown. Cleared on every reload.
|
|
66
|
-
negativeCache: new Set(),
|
|
67
|
-
};
|
|
68
|
-
|
|
69
|
-
function defaultCachePath() {
|
|
70
|
-
return path.join(os.homedir(), ".tokenusage", "cache", "pricing.json");
|
|
71
|
-
}
|
|
72
|
-
|
|
73
|
-
async function ensurePricingLoaded(opts = {}) {
|
|
74
|
-
if (state.loaded) return state;
|
|
75
|
-
if (state.loadingPromise) return state.loadingPromise;
|
|
76
|
-
|
|
77
|
-
state.loadingPromise = (async () => {
|
|
78
|
-
try {
|
|
79
|
-
const cachePath = opts.cachePath || defaultCachePath();
|
|
80
|
-
const { data, source } = await loadLitellmData({ ...opts, cachePath });
|
|
81
|
-
state.litellmRawMap = data || {};
|
|
82
|
-
state.litellmPerMillionMap = buildLitellmPerMillionMap(state.litellmRawMap);
|
|
83
|
-
state.source = source;
|
|
84
|
-
state.loaded = true;
|
|
85
|
-
state.revision += 1;
|
|
86
|
-
state.negativeCache.clear();
|
|
87
|
-
return state;
|
|
88
|
-
} finally {
|
|
89
|
-
state.loadingPromise = null;
|
|
90
|
-
}
|
|
91
|
-
})();
|
|
92
|
-
|
|
93
|
-
return state.loadingPromise;
|
|
94
|
-
}
|
|
95
|
-
|
|
96
|
-
// For tests: drop loaded state so a fresh call can re-load. Seeds with the
|
|
97
|
-
// bundled snapshot so getModelPricing() still works without ensurePricingLoaded.
|
|
98
|
-
function resetPricingForTests() {
|
|
99
|
-
state.loaded = false;
|
|
100
|
-
state.loadingPromise = null;
|
|
101
|
-
state.litellmRawMap = seedRaw;
|
|
102
|
-
state.litellmPerMillionMap = buildLitellmPerMillionMap(seedRaw);
|
|
103
|
-
state.source = Object.keys(seedRaw).length ? "seed-snapshot:sync" : null;
|
|
104
|
-
state.revision += 1;
|
|
105
|
-
state.negativeCache.clear();
|
|
106
|
-
}
|
|
107
|
-
|
|
108
|
-
function getPricingRevision() {
|
|
109
|
-
return state.revision;
|
|
110
|
-
}
|
|
111
|
-
|
|
112
|
-
function getModelPricing(model, opts = {}) {
|
|
113
|
-
if (!model) return ZERO_PRICING;
|
|
114
|
-
let lookupSource = null;
|
|
115
|
-
if (typeof opts === "string") {
|
|
116
|
-
lookupSource = opts.toLowerCase();
|
|
117
|
-
} else if (typeof opts.source === "string") {
|
|
118
|
-
lookupSource = opts.source.toLowerCase();
|
|
119
|
-
}
|
|
120
|
-
const cacheKey = lookupSource ? `${lookupSource}\0${model}` : model;
|
|
121
|
-
if (state.negativeCache.has(cacheKey)) return ZERO_PRICING;
|
|
122
|
-
|
|
123
|
-
const result = lookupPricing(model, {
|
|
124
|
-
curated: curatedOverrides,
|
|
125
|
-
litellm: state.litellmPerMillionMap,
|
|
126
|
-
source: lookupSource,
|
|
127
|
-
});
|
|
128
|
-
if (result.hit) return result.value;
|
|
129
|
-
|
|
130
|
-
state.negativeCache.add(cacheKey);
|
|
131
|
-
return ZERO_PRICING;
|
|
132
|
-
}
|
|
133
|
-
|
|
134
|
-
// Callers hit this once per queue row, and a miss walks the whole LiteLLM map,
|
|
135
|
-
// so results are memoized per pricing revision. Copies go out so a caller that
|
|
136
|
-
// mutates its object cannot poison the cache.
|
|
137
|
-
const pricingInfoCache = { revision: -1, entries: new Map() };
|
|
138
|
-
|
|
139
|
-
// Preserve matcher provenance: a known zero rate is free, a missing model is
|
|
140
|
-
// unpriced. Both return zero dollars, but they mean different things in the UI.
|
|
141
|
-
function getModelPricingInfo(model, opts = {}) {
|
|
142
|
-
const source = typeof opts === "string" ? opts : opts.source;
|
|
143
|
-
const sourceKey = typeof source === "string" ? source.toLowerCase() : null;
|
|
144
|
-
if (pricingInfoCache.revision !== state.revision) {
|
|
145
|
-
pricingInfoCache.revision = state.revision;
|
|
146
|
-
pricingInfoCache.entries.clear();
|
|
147
|
-
}
|
|
148
|
-
const cacheKey = `${sourceKey || ""}\0${model || ""}`;
|
|
149
|
-
let info = pricingInfoCache.entries.get(cacheKey);
|
|
150
|
-
if (!info) {
|
|
151
|
-
info = computeModelPricingInfo(model, opts, sourceKey);
|
|
152
|
-
pricingInfoCache.entries.set(cacheKey, info);
|
|
153
|
-
}
|
|
154
|
-
return { ...info };
|
|
155
|
-
}
|
|
156
|
-
|
|
157
|
-
function computeModelPricingInfo(model, opts, sourceKey) {
|
|
158
|
-
if (LOCAL_INFERENCE_SOURCES.has(sourceKey) || PI_SUBSCRIPTION_SOURCES.has(sourceKey)) {
|
|
159
|
-
return { status: "free", source: LOCAL_INFERENCE_SOURCES.has(sourceKey) ? "local_inference" : "subscription", ...ZERO_PRICING };
|
|
160
|
-
}
|
|
161
|
-
const result = lookupPricing(model, {
|
|
162
|
-
curated: curatedOverrides,
|
|
163
|
-
litellm: state.litellmPerMillionMap,
|
|
164
|
-
source: sourceKey,
|
|
165
|
-
});
|
|
166
|
-
const rates = getModelPricing(model, opts);
|
|
167
|
-
return {
|
|
168
|
-
status: !result.hit ? "unpriced" : [rates.input, rates.output, rates.cache_read, rates.cache_write]
|
|
169
|
-
.some((value) => Number(value) > 0) ? "priced" : "free",
|
|
170
|
-
source: result.hit ? result.source : null,
|
|
171
|
-
input: Number(rates.input) || 0,
|
|
172
|
-
output: Number(rates.output) || 0,
|
|
173
|
-
cache_read: Number(rates.cache_read) || 0,
|
|
174
|
-
cache_write: Number(rates.cache_write) || 0,
|
|
175
|
-
};
|
|
176
|
-
}
|
|
177
|
-
|
|
178
|
-
function isDeepSeekTimePricedModel(model) {
|
|
179
|
-
const lower = String(model || "").toLowerCase();
|
|
180
|
-
return DEEPSEEK_TIME_PRICED_MODELS.some((name) => lower.includes(name));
|
|
181
|
-
}
|
|
182
|
-
|
|
183
|
-
// From 00:00 Beijing time on 2026-08-23 — this instant — DeepSeek bills whole
|
|
184
|
-
// Beijing weekends off-peak, peak hours included.
|
|
185
|
-
// https://api-docs.deepseek.com/quick_start/pricing/
|
|
186
|
-
const DEEPSEEK_WEEKEND_OFF_PEAK_FROM_MS = Date.UTC(2026, 7, 22, 16, 0, 0);
|
|
187
|
-
|
|
188
|
-
// The weekend is bounded in Beijing time, so it runs 16:00Z Friday to 16:00Z
|
|
189
|
-
// Sunday. Shifting the instant by +08:00 before reading the weekday is what puts
|
|
190
|
-
// both of those edges in the right place; getUTCDay() on the raw instant marks a
|
|
191
|
-
// different 48 hours. China has had no daylight saving since 1991, so the fixed
|
|
192
|
-
// offset is exact.
|
|
193
|
-
function isBeijingWeekendOffPeak(timestamp) {
|
|
194
|
-
if (timestamp < DEEPSEEK_WEEKEND_OFF_PEAK_FROM_MS) return false;
|
|
195
|
-
const beijingDay = new Date(timestamp + 8 * 60 * 60 * 1000).getUTCDay();
|
|
196
|
-
return beijingDay === 0 || beijingDay === 6;
|
|
197
|
-
}
|
|
198
|
-
|
|
199
|
-
function isDeepSeekOffPeak(row) {
|
|
200
|
-
if (row?.pricing_tier === "off_peak") return true;
|
|
201
|
-
if (row?.pricing_tier === "peak") return false;
|
|
202
|
-
const timestamp = Date.parse(String(row?.hour_start || ""));
|
|
203
|
-
if (!Number.isFinite(timestamp)) return false;
|
|
204
|
-
if (isBeijingWeekendOffPeak(timestamp)) return true;
|
|
205
|
-
const hour = new Date(timestamp).getUTCHours();
|
|
206
|
-
return !((hour >= 1 && hour < 4) || (hour >= 6 && hour < 10));
|
|
207
|
-
}
|
|
208
|
-
|
|
209
|
-
function getRowPricing(row) {
|
|
210
|
-
const pricing = getModelPricing(row?.model, { source: row?.source });
|
|
211
|
-
// AStudio uses iFlytek MaaS fixed prices and does not inherit DeepSeek public API
|
|
212
|
-
// time-based discounts.
|
|
213
|
-
if (String(row?.source || "").toLowerCase() === "acode") return pricing;
|
|
214
|
-
if (!isDeepSeekTimePricedModel(row?.model) || !isDeepSeekOffPeak(row)) return pricing;
|
|
215
|
-
return {
|
|
216
|
-
...pricing,
|
|
217
|
-
input: (pricing.input || 0) * 0.5,
|
|
218
|
-
output: (pricing.output || 0) * 0.5,
|
|
219
|
-
cache_read: (pricing.cache_read || 0) * 0.5,
|
|
220
|
-
cache_write: (pricing.cache_write || 0) * 0.5,
|
|
221
|
-
};
|
|
222
|
-
}
|
|
223
|
-
|
|
224
|
-
// Same formula and Codex/every-code reasoning-folding rule as the previous
|
|
225
|
-
// computeRowCost in src/lib/local-api.js. Moved here so vite mock + local
|
|
226
|
-
// server share one source of truth.
|
|
227
|
-
function computeRowCost(row) {
|
|
228
|
-
if (LOCAL_INFERENCE_SOURCES.has(String(row?.source || "").toLowerCase())) return 0;
|
|
229
|
-
// Pi can route a turn through a subscription-backed Copilot provider. Pi's
|
|
230
|
-
// usage record reports a zero marginal cost for those turns; do not
|
|
231
|
-
// reinterpret the Claude model name as an Anthropic API bill.
|
|
232
|
-
if (PI_SUBSCRIPTION_SOURCES.has(String(row?.source || "").toLowerCase())) return 0;
|
|
233
|
-
// Some providers (currently Grok) persist an exact server-reported cost on
|
|
234
|
-
// the usage bucket. Prefer it when positive; zero remains the legacy
|
|
235
|
-
// "unreported" sentinel and falls through to model pricing.
|
|
236
|
-
const reportedCost = Number(row?.total_cost_usd);
|
|
237
|
-
if (
|
|
238
|
-
SOURCES_WITH_AUTHORITATIVE_COST.has(row?.source) &&
|
|
239
|
-
Number.isFinite(reportedCost) &&
|
|
240
|
-
reportedCost > 0
|
|
241
|
-
) return reportedCost;
|
|
242
|
-
const pricing = getRowPricing(row);
|
|
243
|
-
// OmO, like Codex, reports reasoning as a subset of `output` (its own
|
|
244
|
-
// usage.cost bills no separate reasoning component), so charging it again
|
|
245
|
-
// here would double-bill every reasoning token. Cline is the same: its
|
|
246
|
-
// `outputTokens` is the AI SDK total (text + reasoning), so
|
|
247
|
-
// reasoning_output_tokens is a subset marker, never a second line item.
|
|
248
|
-
const reasoningIncludedInOutput =
|
|
249
|
-
row.source === "codex" ||
|
|
250
|
-
row.source === "acode" ||
|
|
251
|
-
row.source === "every-code" ||
|
|
252
|
-
row.source === "omo" ||
|
|
253
|
-
row.source === "cline" ||
|
|
254
|
-
row.source === "muse";
|
|
255
|
-
const reasoningCost = reasoningIncludedInOutput
|
|
256
|
-
? 0
|
|
257
|
-
: (row.reasoning_output_tokens || 0) * (pricing.output || 0);
|
|
258
|
-
const baseCost = (
|
|
259
|
-
((row.input_tokens || 0) * (pricing.input || 0) +
|
|
260
|
-
(row.output_tokens || 0) * (pricing.output || 0) +
|
|
261
|
-
(row.cached_input_tokens || 0) * (pricing.cache_read || 0) +
|
|
262
|
-
(row.cache_creation_input_tokens || 0) * (pricing.cache_write || 0) +
|
|
263
|
-
reasoningCost) /
|
|
264
|
-
1_000_000
|
|
265
|
-
);
|
|
266
|
-
|
|
267
|
-
const model = String(row?.model || "").toLowerCase();
|
|
268
|
-
const usesOpenAILongContextTier =
|
|
269
|
-
model === "gpt-5.6" || model.includes("gpt-5.6-sol") || model.includes("gpt-6-astra") ||
|
|
270
|
-
model.includes("gpt-6-sol") || model.includes("gpt-6.1-sol");
|
|
271
|
-
// Only models whose curated entry carries an explicit Fast/priority rate get
|
|
272
|
-
// the tier premium. The other OpenAI tiers have no published priority price,
|
|
273
|
-
// and inventing one is worse than reporting Standard.
|
|
274
|
-
const priorityMultiplier = Number(pricing.priority_multiplier);
|
|
275
|
-
const hasPriorityRates = Number.isFinite(priorityMultiplier) && priorityMultiplier > 1;
|
|
276
|
-
if (!usesOpenAILongContextTier && !hasPriorityRates) return baseCost;
|
|
277
|
-
|
|
278
|
-
const bounded = (value, total) => Math.min(
|
|
279
|
-
Math.max(0, Number(value) || 0),
|
|
280
|
-
Math.max(0, Number(total) || 0),
|
|
281
|
-
);
|
|
282
|
-
// Both premium families below read "observed subset" columns: how much of the
|
|
283
|
-
// row's own token columns came from requests with a given property. They are
|
|
284
|
-
// annotations, never addends — total_tokens stays the sum of the base columns
|
|
285
|
-
// whether or not they are present, and a legacy row that omits them prices
|
|
286
|
-
// exactly as it did before.
|
|
287
|
-
//
|
|
288
|
-
// The parser records only usage from requests whose cache-inclusive input
|
|
289
|
-
// exceeded 272K. OpenAI prices the whole such request at 2x input and 1.5x
|
|
290
|
-
// output, so add the premium over the standard-rate base exactly once.
|
|
291
|
-
const longInput = bounded(row.long_context_input_tokens, row.input_tokens);
|
|
292
|
-
const longCached = bounded(row.long_context_cached_input_tokens, row.cached_input_tokens);
|
|
293
|
-
const longCacheWrite = bounded(
|
|
294
|
-
row.long_context_cache_creation_input_tokens,
|
|
295
|
-
row.cache_creation_input_tokens,
|
|
296
|
-
);
|
|
297
|
-
const longOutput = bounded(row.long_context_output_tokens, row.output_tokens);
|
|
298
|
-
const longReasoning = reasoningIncludedInOutput
|
|
299
|
-
? 0
|
|
300
|
-
: bounded(row.long_context_reasoning_output_tokens, row.reasoning_output_tokens);
|
|
301
|
-
const longContextPremium = usesOpenAILongContextTier ? (
|
|
302
|
-
longInput * (pricing.input || 0) +
|
|
303
|
-
longCached * (pricing.cache_read || 0) +
|
|
304
|
-
longCacheWrite * (pricing.cache_write || 0) +
|
|
305
|
-
0.5 * longOutput * (pricing.output || 0) +
|
|
306
|
-
0.5 * longReasoning * (pricing.output || 0)
|
|
307
|
-
) / 1_000_000 : 0;
|
|
308
|
-
|
|
309
|
-
if (!hasPriorityRates) return baseCost + longContextPremium;
|
|
310
|
-
|
|
311
|
-
// Astra Fast is a flat multiple of whichever context tier the request was on
|
|
312
|
-
// (short 10/1/12.5/50 -> 20/2/25/100; long 20/2/25/75 -> 40/4/50/150), so the
|
|
313
|
-
// tier multiplier and the long-context premium are orthogonal: a request that
|
|
314
|
-
// is both bills at the LONG rate times the multiplier, not standard times the
|
|
315
|
-
// multiplier. That needs the intersection, which is why the parser records
|
|
316
|
-
// priority_long_context_* alongside priority_*; without it (queue rows, which
|
|
317
|
-
// carry no context-length subset at all) the overlap is 0 and the premium
|
|
318
|
-
// falls back to the short-context multiple, matching the row's own base.
|
|
319
|
-
const priorityInput = bounded(row.priority_input_tokens, row.input_tokens);
|
|
320
|
-
const priorityCached = bounded(row.priority_cached_input_tokens, row.cached_input_tokens);
|
|
321
|
-
const priorityCacheWrite = bounded(
|
|
322
|
-
row.priority_cache_creation_input_tokens,
|
|
323
|
-
row.cache_creation_input_tokens,
|
|
324
|
-
);
|
|
325
|
-
const priorityOutput = bounded(row.priority_output_tokens, row.output_tokens);
|
|
326
|
-
const priorityReasoning = reasoningIncludedInOutput
|
|
327
|
-
? 0
|
|
328
|
-
: bounded(row.priority_reasoning_output_tokens, row.reasoning_output_tokens);
|
|
329
|
-
const priorityBase = (
|
|
330
|
-
priorityInput * (pricing.input || 0) +
|
|
331
|
-
priorityCached * (pricing.cache_read || 0) +
|
|
332
|
-
priorityCacheWrite * (pricing.cache_write || 0) +
|
|
333
|
-
priorityOutput * (pricing.output || 0) +
|
|
334
|
-
priorityReasoning * (pricing.output || 0)
|
|
335
|
-
) / 1_000_000;
|
|
336
|
-
|
|
337
|
-
const priorityLong = usesOpenAILongContextTier ? (
|
|
338
|
-
bounded(row.priority_long_context_input_tokens, Math.min(priorityInput, longInput)) *
|
|
339
|
-
(pricing.input || 0) +
|
|
340
|
-
bounded(row.priority_long_context_cached_input_tokens, Math.min(priorityCached, longCached)) *
|
|
341
|
-
(pricing.cache_read || 0) +
|
|
342
|
-
bounded(
|
|
343
|
-
row.priority_long_context_cache_creation_input_tokens,
|
|
344
|
-
Math.min(priorityCacheWrite, longCacheWrite),
|
|
345
|
-
) * (pricing.cache_write || 0) +
|
|
346
|
-
0.5 * bounded(row.priority_long_context_output_tokens, Math.min(priorityOutput, longOutput)) *
|
|
347
|
-
(pricing.output || 0) +
|
|
348
|
-
0.5 * bounded(
|
|
349
|
-
row.priority_long_context_reasoning_output_tokens,
|
|
350
|
-
Math.min(priorityReasoning, longReasoning),
|
|
351
|
-
) * (pricing.output || 0)
|
|
352
|
-
) / 1_000_000 : 0;
|
|
353
|
-
|
|
354
|
-
const priorityPremium = (priorityMultiplier - 1) * (priorityBase + priorityLong);
|
|
355
|
-
return baseCost + longContextPremium + priorityPremium;
|
|
356
|
-
}
|
|
357
|
-
|
|
358
|
-
// Backwards-compatible MODEL_PRICING export. Test at
|
|
359
|
-
// test/model-breakdown.test.js:236 reads `localApi.MODEL_PRICING["kiro-agent"]`
|
|
360
|
-
// and expects { input, output, cache_read, cache_write } shape. We expose the
|
|
361
|
-
// CURATED.exact map (which contains the kiro entries by design); LiteLLM
|
|
362
|
-
// entries are NOT included here because they're keyed dynamically and the old
|
|
363
|
-
// table was authoritative for what is now CURATED.
|
|
364
|
-
const MODEL_PRICING = curatedOverrides.exact;
|
|
365
|
-
|
|
366
|
-
module.exports = {
|
|
367
|
-
ensurePricingLoaded,
|
|
368
|
-
getPricingRevision,
|
|
369
|
-
getModelPricing,
|
|
370
|
-
getModelPricingInfo,
|
|
371
|
-
getRowPricing,
|
|
372
|
-
computeRowCost,
|
|
373
|
-
resetPricingForTests,
|
|
374
|
-
MODEL_PRICING,
|
|
375
|
-
ZERO_PRICING,
|
|
376
|
-
OPENAI_LONG_CONTEXT_INPUT_THRESHOLD,
|
|
377
|
-
// Internal hooks for tests.
|
|
378
|
-
__getStateForTests: () => state,
|
|
379
|
-
};
|
|
1
|
+
// Public pricing API. Replaces the hard-coded MODEL_PRICING table that used
|
|
2
|
+
// to live in src/lib/local-api.js. Keeps the same synchronous shape so all
|
|
3
|
+
// existing callers (computeRowCost, /api/functions/* handlers, tests) work
|
|
4
|
+
// unchanged after `await ensurePricingLoaded()` is awaited once at startup.
|
|
5
|
+
|
|
6
|
+
const fs = require("node:fs");
|
|
7
|
+
const path = require("node:path");
|
|
8
|
+
const os = require("node:os");
|
|
9
|
+
|
|
10
|
+
const curatedOverrides = require("./curated-overrides.json");
|
|
11
|
+
const {
|
|
12
|
+
lookupPricing,
|
|
13
|
+
buildLitellmPerMillionMap,
|
|
14
|
+
} = require("./matcher");
|
|
15
|
+
const { loadLitellmData } = require("./litellm-fetcher");
|
|
16
|
+
|
|
17
|
+
const ZERO_PRICING = { input: 0, output: 0, cache_read: 0, cache_write: 0 };
|
|
18
|
+
const PI_SUBSCRIPTION_SOURCES = new Set([
|
|
19
|
+
"pi-github-copilot",
|
|
20
|
+
"pi-copilot",
|
|
21
|
+
"prime-agent-github-copilot",
|
|
22
|
+
"prime-agent-copilot",
|
|
23
|
+
]);
|
|
24
|
+
// LM Studio's server logs describe inference served by the local developer
|
|
25
|
+
// server (including LM Link). Secure Cloud usage has a separate billing path
|
|
26
|
+
// and is not present in these logs.
|
|
27
|
+
const LOCAL_INFERENCE_SOURCES = new Set(["lmstudio"]);
|
|
28
|
+
// OpenAI long-context pricing depends on a single request's raw input,
|
|
29
|
+
// including cached tokens, never a session/day aggregate. Astra supports
|
|
30
|
+
// a larger context window; only observed request subsets receive the premium.
|
|
31
|
+
const OPENAI_LONG_CONTEXT_INPUT_THRESHOLD = 272_000;
|
|
32
|
+
const SOURCES_WITH_AUTHORITATIVE_COST = new Set(["grok", "cline"]);
|
|
33
|
+
const SEED_SNAPSHOT_PATH = path.resolve(__dirname, "seed-snapshot.json");
|
|
34
|
+
const DEEPSEEK_TIME_PRICED_MODELS = [
|
|
35
|
+
"deepseek-v4-flash",
|
|
36
|
+
"deepseek-v4.1-flash",
|
|
37
|
+
"deepseek-flash",
|
|
38
|
+
"deepseek-v4-pro",
|
|
39
|
+
];
|
|
40
|
+
// Sync seed load. Done at require-time so callers that haven't awaited
|
|
41
|
+
// ensurePricingLoaded() (e.g. tests, vite mock startup, edge functions) still
|
|
42
|
+
// get LiteLLM-backed pricing instead of all-zero. ensurePricingLoaded() will
|
|
43
|
+
// later upgrade this to fresh disk cache or upstream data.
|
|
44
|
+
function loadSeedSync() {
|
|
45
|
+
try {
|
|
46
|
+
const raw = fs.readFileSync(SEED_SNAPSHOT_PATH, "utf8");
|
|
47
|
+
const parsed = JSON.parse(raw);
|
|
48
|
+
delete parsed._meta;
|
|
49
|
+
return parsed;
|
|
50
|
+
} catch (e) {
|
|
51
|
+
return {};
|
|
52
|
+
}
|
|
53
|
+
}
|
|
54
|
+
|
|
55
|
+
const seedRaw = loadSeedSync();
|
|
56
|
+
|
|
57
|
+
const state = {
|
|
58
|
+
loaded: false,
|
|
59
|
+
loadingPromise: null,
|
|
60
|
+
revision: 0,
|
|
61
|
+
litellmRawMap: seedRaw, // raw per-token; field shape from LiteLLM JSON
|
|
62
|
+
litellmPerMillionMap: buildLitellmPerMillionMap(seedRaw), // USD/MTok
|
|
63
|
+
source: Object.keys(seedRaw).length ? "seed-snapshot:sync" : null,
|
|
64
|
+
// negativeCache prevents re-walking the LiteLLM map for models we've already
|
|
65
|
+
// determined are unknown. Cleared on every reload.
|
|
66
|
+
negativeCache: new Set(),
|
|
67
|
+
};
|
|
68
|
+
|
|
69
|
+
function defaultCachePath() {
|
|
70
|
+
return path.join(os.homedir(), ".tokenusage", "cache", "pricing.json");
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
async function ensurePricingLoaded(opts = {}) {
|
|
74
|
+
if (state.loaded) return state;
|
|
75
|
+
if (state.loadingPromise) return state.loadingPromise;
|
|
76
|
+
|
|
77
|
+
state.loadingPromise = (async () => {
|
|
78
|
+
try {
|
|
79
|
+
const cachePath = opts.cachePath || defaultCachePath();
|
|
80
|
+
const { data, source } = await loadLitellmData({ ...opts, cachePath });
|
|
81
|
+
state.litellmRawMap = data || {};
|
|
82
|
+
state.litellmPerMillionMap = buildLitellmPerMillionMap(state.litellmRawMap);
|
|
83
|
+
state.source = source;
|
|
84
|
+
state.loaded = true;
|
|
85
|
+
state.revision += 1;
|
|
86
|
+
state.negativeCache.clear();
|
|
87
|
+
return state;
|
|
88
|
+
} finally {
|
|
89
|
+
state.loadingPromise = null;
|
|
90
|
+
}
|
|
91
|
+
})();
|
|
92
|
+
|
|
93
|
+
return state.loadingPromise;
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
// For tests: drop loaded state so a fresh call can re-load. Seeds with the
|
|
97
|
+
// bundled snapshot so getModelPricing() still works without ensurePricingLoaded.
|
|
98
|
+
function resetPricingForTests() {
|
|
99
|
+
state.loaded = false;
|
|
100
|
+
state.loadingPromise = null;
|
|
101
|
+
state.litellmRawMap = seedRaw;
|
|
102
|
+
state.litellmPerMillionMap = buildLitellmPerMillionMap(seedRaw);
|
|
103
|
+
state.source = Object.keys(seedRaw).length ? "seed-snapshot:sync" : null;
|
|
104
|
+
state.revision += 1;
|
|
105
|
+
state.negativeCache.clear();
|
|
106
|
+
}
|
|
107
|
+
|
|
108
|
+
function getPricingRevision() {
|
|
109
|
+
return state.revision;
|
|
110
|
+
}
|
|
111
|
+
|
|
112
|
+
function getModelPricing(model, opts = {}) {
|
|
113
|
+
if (!model) return ZERO_PRICING;
|
|
114
|
+
let lookupSource = null;
|
|
115
|
+
if (typeof opts === "string") {
|
|
116
|
+
lookupSource = opts.toLowerCase();
|
|
117
|
+
} else if (typeof opts.source === "string") {
|
|
118
|
+
lookupSource = opts.source.toLowerCase();
|
|
119
|
+
}
|
|
120
|
+
const cacheKey = lookupSource ? `${lookupSource}\0${model}` : model;
|
|
121
|
+
if (state.negativeCache.has(cacheKey)) return ZERO_PRICING;
|
|
122
|
+
|
|
123
|
+
const result = lookupPricing(model, {
|
|
124
|
+
curated: curatedOverrides,
|
|
125
|
+
litellm: state.litellmPerMillionMap,
|
|
126
|
+
source: lookupSource,
|
|
127
|
+
});
|
|
128
|
+
if (result.hit) return result.value;
|
|
129
|
+
|
|
130
|
+
state.negativeCache.add(cacheKey);
|
|
131
|
+
return ZERO_PRICING;
|
|
132
|
+
}
|
|
133
|
+
|
|
134
|
+
// Callers hit this once per queue row, and a miss walks the whole LiteLLM map,
|
|
135
|
+
// so results are memoized per pricing revision. Copies go out so a caller that
|
|
136
|
+
// mutates its object cannot poison the cache.
|
|
137
|
+
const pricingInfoCache = { revision: -1, entries: new Map() };
|
|
138
|
+
|
|
139
|
+
// Preserve matcher provenance: a known zero rate is free, a missing model is
|
|
140
|
+
// unpriced. Both return zero dollars, but they mean different things in the UI.
|
|
141
|
+
function getModelPricingInfo(model, opts = {}) {
|
|
142
|
+
const source = typeof opts === "string" ? opts : opts.source;
|
|
143
|
+
const sourceKey = typeof source === "string" ? source.toLowerCase() : null;
|
|
144
|
+
if (pricingInfoCache.revision !== state.revision) {
|
|
145
|
+
pricingInfoCache.revision = state.revision;
|
|
146
|
+
pricingInfoCache.entries.clear();
|
|
147
|
+
}
|
|
148
|
+
const cacheKey = `${sourceKey || ""}\0${model || ""}`;
|
|
149
|
+
let info = pricingInfoCache.entries.get(cacheKey);
|
|
150
|
+
if (!info) {
|
|
151
|
+
info = computeModelPricingInfo(model, opts, sourceKey);
|
|
152
|
+
pricingInfoCache.entries.set(cacheKey, info);
|
|
153
|
+
}
|
|
154
|
+
return { ...info };
|
|
155
|
+
}
|
|
156
|
+
|
|
157
|
+
function computeModelPricingInfo(model, opts, sourceKey) {
|
|
158
|
+
if (LOCAL_INFERENCE_SOURCES.has(sourceKey) || PI_SUBSCRIPTION_SOURCES.has(sourceKey)) {
|
|
159
|
+
return { status: "free", source: LOCAL_INFERENCE_SOURCES.has(sourceKey) ? "local_inference" : "subscription", ...ZERO_PRICING };
|
|
160
|
+
}
|
|
161
|
+
const result = lookupPricing(model, {
|
|
162
|
+
curated: curatedOverrides,
|
|
163
|
+
litellm: state.litellmPerMillionMap,
|
|
164
|
+
source: sourceKey,
|
|
165
|
+
});
|
|
166
|
+
const rates = getModelPricing(model, opts);
|
|
167
|
+
return {
|
|
168
|
+
status: !result.hit ? "unpriced" : [rates.input, rates.output, rates.cache_read, rates.cache_write]
|
|
169
|
+
.some((value) => Number(value) > 0) ? "priced" : "free",
|
|
170
|
+
source: result.hit ? result.source : null,
|
|
171
|
+
input: Number(rates.input) || 0,
|
|
172
|
+
output: Number(rates.output) || 0,
|
|
173
|
+
cache_read: Number(rates.cache_read) || 0,
|
|
174
|
+
cache_write: Number(rates.cache_write) || 0,
|
|
175
|
+
};
|
|
176
|
+
}
|
|
177
|
+
|
|
178
|
+
function isDeepSeekTimePricedModel(model) {
|
|
179
|
+
const lower = String(model || "").toLowerCase();
|
|
180
|
+
return DEEPSEEK_TIME_PRICED_MODELS.some((name) => lower.includes(name));
|
|
181
|
+
}
|
|
182
|
+
|
|
183
|
+
// From 00:00 Beijing time on 2026-08-23 — this instant — DeepSeek bills whole
|
|
184
|
+
// Beijing weekends off-peak, peak hours included.
|
|
185
|
+
// https://api-docs.deepseek.com/quick_start/pricing/
|
|
186
|
+
const DEEPSEEK_WEEKEND_OFF_PEAK_FROM_MS = Date.UTC(2026, 7, 22, 16, 0, 0);
|
|
187
|
+
|
|
188
|
+
// The weekend is bounded in Beijing time, so it runs 16:00Z Friday to 16:00Z
|
|
189
|
+
// Sunday. Shifting the instant by +08:00 before reading the weekday is what puts
|
|
190
|
+
// both of those edges in the right place; getUTCDay() on the raw instant marks a
|
|
191
|
+
// different 48 hours. China has had no daylight saving since 1991, so the fixed
|
|
192
|
+
// offset is exact.
|
|
193
|
+
function isBeijingWeekendOffPeak(timestamp) {
|
|
194
|
+
if (timestamp < DEEPSEEK_WEEKEND_OFF_PEAK_FROM_MS) return false;
|
|
195
|
+
const beijingDay = new Date(timestamp + 8 * 60 * 60 * 1000).getUTCDay();
|
|
196
|
+
return beijingDay === 0 || beijingDay === 6;
|
|
197
|
+
}
|
|
198
|
+
|
|
199
|
+
function isDeepSeekOffPeak(row) {
|
|
200
|
+
if (row?.pricing_tier === "off_peak") return true;
|
|
201
|
+
if (row?.pricing_tier === "peak") return false;
|
|
202
|
+
const timestamp = Date.parse(String(row?.hour_start || ""));
|
|
203
|
+
if (!Number.isFinite(timestamp)) return false;
|
|
204
|
+
if (isBeijingWeekendOffPeak(timestamp)) return true;
|
|
205
|
+
const hour = new Date(timestamp).getUTCHours();
|
|
206
|
+
return !((hour >= 1 && hour < 4) || (hour >= 6 && hour < 10));
|
|
207
|
+
}
|
|
208
|
+
|
|
209
|
+
function getRowPricing(row) {
|
|
210
|
+
const pricing = getModelPricing(row?.model, { source: row?.source });
|
|
211
|
+
// AStudio uses iFlytek MaaS fixed prices and does not inherit DeepSeek public API
|
|
212
|
+
// time-based discounts.
|
|
213
|
+
if (String(row?.source || "").toLowerCase() === "acode") return pricing;
|
|
214
|
+
if (!isDeepSeekTimePricedModel(row?.model) || !isDeepSeekOffPeak(row)) return pricing;
|
|
215
|
+
return {
|
|
216
|
+
...pricing,
|
|
217
|
+
input: (pricing.input || 0) * 0.5,
|
|
218
|
+
output: (pricing.output || 0) * 0.5,
|
|
219
|
+
cache_read: (pricing.cache_read || 0) * 0.5,
|
|
220
|
+
cache_write: (pricing.cache_write || 0) * 0.5,
|
|
221
|
+
};
|
|
222
|
+
}
|
|
223
|
+
|
|
224
|
+
// Same formula and Codex/every-code reasoning-folding rule as the previous
|
|
225
|
+
// computeRowCost in src/lib/local-api.js. Moved here so vite mock + local
|
|
226
|
+
// server share one source of truth.
|
|
227
|
+
function computeRowCost(row) {
|
|
228
|
+
if (LOCAL_INFERENCE_SOURCES.has(String(row?.source || "").toLowerCase())) return 0;
|
|
229
|
+
// Pi can route a turn through a subscription-backed Copilot provider. Pi's
|
|
230
|
+
// usage record reports a zero marginal cost for those turns; do not
|
|
231
|
+
// reinterpret the Claude model name as an Anthropic API bill.
|
|
232
|
+
if (PI_SUBSCRIPTION_SOURCES.has(String(row?.source || "").toLowerCase())) return 0;
|
|
233
|
+
// Some providers (currently Grok) persist an exact server-reported cost on
|
|
234
|
+
// the usage bucket. Prefer it when positive; zero remains the legacy
|
|
235
|
+
// "unreported" sentinel and falls through to model pricing.
|
|
236
|
+
const reportedCost = Number(row?.total_cost_usd);
|
|
237
|
+
if (
|
|
238
|
+
SOURCES_WITH_AUTHORITATIVE_COST.has(row?.source) &&
|
|
239
|
+
Number.isFinite(reportedCost) &&
|
|
240
|
+
reportedCost > 0
|
|
241
|
+
) return reportedCost;
|
|
242
|
+
const pricing = getRowPricing(row);
|
|
243
|
+
// OmO, like Codex, reports reasoning as a subset of `output` (its own
|
|
244
|
+
// usage.cost bills no separate reasoning component), so charging it again
|
|
245
|
+
// here would double-bill every reasoning token. Cline is the same: its
|
|
246
|
+
// `outputTokens` is the AI SDK total (text + reasoning), so
|
|
247
|
+
// reasoning_output_tokens is a subset marker, never a second line item.
|
|
248
|
+
const reasoningIncludedInOutput =
|
|
249
|
+
row.source === "codex" ||
|
|
250
|
+
row.source === "acode" ||
|
|
251
|
+
row.source === "every-code" ||
|
|
252
|
+
row.source === "omo" ||
|
|
253
|
+
row.source === "cline" ||
|
|
254
|
+
row.source === "muse";
|
|
255
|
+
const reasoningCost = reasoningIncludedInOutput
|
|
256
|
+
? 0
|
|
257
|
+
: (row.reasoning_output_tokens || 0) * (pricing.output || 0);
|
|
258
|
+
const baseCost = (
|
|
259
|
+
((row.input_tokens || 0) * (pricing.input || 0) +
|
|
260
|
+
(row.output_tokens || 0) * (pricing.output || 0) +
|
|
261
|
+
(row.cached_input_tokens || 0) * (pricing.cache_read || 0) +
|
|
262
|
+
(row.cache_creation_input_tokens || 0) * (pricing.cache_write || 0) +
|
|
263
|
+
reasoningCost) /
|
|
264
|
+
1_000_000
|
|
265
|
+
);
|
|
266
|
+
|
|
267
|
+
const model = String(row?.model || "").toLowerCase();
|
|
268
|
+
const usesOpenAILongContextTier =
|
|
269
|
+
model === "gpt-5.6" || model.includes("gpt-5.6-sol") || model.includes("gpt-6-astra") ||
|
|
270
|
+
model.includes("gpt-6-sol") || model.includes("gpt-6.1-sol");
|
|
271
|
+
// Only models whose curated entry carries an explicit Fast/priority rate get
|
|
272
|
+
// the tier premium. The other OpenAI tiers have no published priority price,
|
|
273
|
+
// and inventing one is worse than reporting Standard.
|
|
274
|
+
const priorityMultiplier = Number(pricing.priority_multiplier);
|
|
275
|
+
const hasPriorityRates = Number.isFinite(priorityMultiplier) && priorityMultiplier > 1;
|
|
276
|
+
if (!usesOpenAILongContextTier && !hasPriorityRates) return baseCost;
|
|
277
|
+
|
|
278
|
+
const bounded = (value, total) => Math.min(
|
|
279
|
+
Math.max(0, Number(value) || 0),
|
|
280
|
+
Math.max(0, Number(total) || 0),
|
|
281
|
+
);
|
|
282
|
+
// Both premium families below read "observed subset" columns: how much of the
|
|
283
|
+
// row's own token columns came from requests with a given property. They are
|
|
284
|
+
// annotations, never addends — total_tokens stays the sum of the base columns
|
|
285
|
+
// whether or not they are present, and a legacy row that omits them prices
|
|
286
|
+
// exactly as it did before.
|
|
287
|
+
//
|
|
288
|
+
// The parser records only usage from requests whose cache-inclusive input
|
|
289
|
+
// exceeded 272K. OpenAI prices the whole such request at 2x input and 1.5x
|
|
290
|
+
// output, so add the premium over the standard-rate base exactly once.
|
|
291
|
+
const longInput = bounded(row.long_context_input_tokens, row.input_tokens);
|
|
292
|
+
const longCached = bounded(row.long_context_cached_input_tokens, row.cached_input_tokens);
|
|
293
|
+
const longCacheWrite = bounded(
|
|
294
|
+
row.long_context_cache_creation_input_tokens,
|
|
295
|
+
row.cache_creation_input_tokens,
|
|
296
|
+
);
|
|
297
|
+
const longOutput = bounded(row.long_context_output_tokens, row.output_tokens);
|
|
298
|
+
const longReasoning = reasoningIncludedInOutput
|
|
299
|
+
? 0
|
|
300
|
+
: bounded(row.long_context_reasoning_output_tokens, row.reasoning_output_tokens);
|
|
301
|
+
const longContextPremium = usesOpenAILongContextTier ? (
|
|
302
|
+
longInput * (pricing.input || 0) +
|
|
303
|
+
longCached * (pricing.cache_read || 0) +
|
|
304
|
+
longCacheWrite * (pricing.cache_write || 0) +
|
|
305
|
+
0.5 * longOutput * (pricing.output || 0) +
|
|
306
|
+
0.5 * longReasoning * (pricing.output || 0)
|
|
307
|
+
) / 1_000_000 : 0;
|
|
308
|
+
|
|
309
|
+
if (!hasPriorityRates) return baseCost + longContextPremium;
|
|
310
|
+
|
|
311
|
+
// Astra Fast is a flat multiple of whichever context tier the request was on
|
|
312
|
+
// (short 10/1/12.5/50 -> 20/2/25/100; long 20/2/25/75 -> 40/4/50/150), so the
|
|
313
|
+
// tier multiplier and the long-context premium are orthogonal: a request that
|
|
314
|
+
// is both bills at the LONG rate times the multiplier, not standard times the
|
|
315
|
+
// multiplier. That needs the intersection, which is why the parser records
|
|
316
|
+
// priority_long_context_* alongside priority_*; without it (queue rows, which
|
|
317
|
+
// carry no context-length subset at all) the overlap is 0 and the premium
|
|
318
|
+
// falls back to the short-context multiple, matching the row's own base.
|
|
319
|
+
const priorityInput = bounded(row.priority_input_tokens, row.input_tokens);
|
|
320
|
+
const priorityCached = bounded(row.priority_cached_input_tokens, row.cached_input_tokens);
|
|
321
|
+
const priorityCacheWrite = bounded(
|
|
322
|
+
row.priority_cache_creation_input_tokens,
|
|
323
|
+
row.cache_creation_input_tokens,
|
|
324
|
+
);
|
|
325
|
+
const priorityOutput = bounded(row.priority_output_tokens, row.output_tokens);
|
|
326
|
+
const priorityReasoning = reasoningIncludedInOutput
|
|
327
|
+
? 0
|
|
328
|
+
: bounded(row.priority_reasoning_output_tokens, row.reasoning_output_tokens);
|
|
329
|
+
const priorityBase = (
|
|
330
|
+
priorityInput * (pricing.input || 0) +
|
|
331
|
+
priorityCached * (pricing.cache_read || 0) +
|
|
332
|
+
priorityCacheWrite * (pricing.cache_write || 0) +
|
|
333
|
+
priorityOutput * (pricing.output || 0) +
|
|
334
|
+
priorityReasoning * (pricing.output || 0)
|
|
335
|
+
) / 1_000_000;
|
|
336
|
+
|
|
337
|
+
const priorityLong = usesOpenAILongContextTier ? (
|
|
338
|
+
bounded(row.priority_long_context_input_tokens, Math.min(priorityInput, longInput)) *
|
|
339
|
+
(pricing.input || 0) +
|
|
340
|
+
bounded(row.priority_long_context_cached_input_tokens, Math.min(priorityCached, longCached)) *
|
|
341
|
+
(pricing.cache_read || 0) +
|
|
342
|
+
bounded(
|
|
343
|
+
row.priority_long_context_cache_creation_input_tokens,
|
|
344
|
+
Math.min(priorityCacheWrite, longCacheWrite),
|
|
345
|
+
) * (pricing.cache_write || 0) +
|
|
346
|
+
0.5 * bounded(row.priority_long_context_output_tokens, Math.min(priorityOutput, longOutput)) *
|
|
347
|
+
(pricing.output || 0) +
|
|
348
|
+
0.5 * bounded(
|
|
349
|
+
row.priority_long_context_reasoning_output_tokens,
|
|
350
|
+
Math.min(priorityReasoning, longReasoning),
|
|
351
|
+
) * (pricing.output || 0)
|
|
352
|
+
) / 1_000_000 : 0;
|
|
353
|
+
|
|
354
|
+
const priorityPremium = (priorityMultiplier - 1) * (priorityBase + priorityLong);
|
|
355
|
+
return baseCost + longContextPremium + priorityPremium;
|
|
356
|
+
}
|
|
357
|
+
|
|
358
|
+
// Backwards-compatible MODEL_PRICING export. Test at
|
|
359
|
+
// test/model-breakdown.test.js:236 reads `localApi.MODEL_PRICING["kiro-agent"]`
|
|
360
|
+
// and expects { input, output, cache_read, cache_write } shape. We expose the
|
|
361
|
+
// CURATED.exact map (which contains the kiro entries by design); LiteLLM
|
|
362
|
+
// entries are NOT included here because they're keyed dynamically and the old
|
|
363
|
+
// table was authoritative for what is now CURATED.
|
|
364
|
+
const MODEL_PRICING = curatedOverrides.exact;
|
|
365
|
+
|
|
366
|
+
module.exports = {
|
|
367
|
+
ensurePricingLoaded,
|
|
368
|
+
getPricingRevision,
|
|
369
|
+
getModelPricing,
|
|
370
|
+
getModelPricingInfo,
|
|
371
|
+
getRowPricing,
|
|
372
|
+
computeRowCost,
|
|
373
|
+
resetPricingForTests,
|
|
374
|
+
MODEL_PRICING,
|
|
375
|
+
ZERO_PRICING,
|
|
376
|
+
OPENAI_LONG_CONTEXT_INPUT_THRESHOLD,
|
|
377
|
+
// Internal hooks for tests.
|
|
378
|
+
__getStateForTests: () => state,
|
|
379
|
+
};
|