@trazum/core 2.7.3 → 2.7.5
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/dist/anthropic-cost.d.ts +22 -5
- package/dist/anthropic-cost.d.ts.map +1 -1
- package/dist/anthropic-cost.js +33 -10
- package/dist/anthropic-cost.js.map +1 -1
- package/dist/anthropic-usage.d.ts +15 -2
- package/dist/anthropic-usage.d.ts.map +1 -1
- package/dist/anthropic-usage.js +41 -5
- package/dist/anthropic-usage.js.map +1 -1
- package/dist/batch-results.d.ts +77 -0
- package/dist/batch-results.d.ts.map +1 -0
- package/dist/batch-results.js +145 -0
- package/dist/batch-results.js.map +1 -0
- package/dist/billing.d.ts +10 -0
- package/dist/billing.d.ts.map +1 -1
- package/dist/billing.js +15 -0
- package/dist/billing.js.map +1 -1
- package/dist/claude-code.d.ts +4 -0
- package/dist/claude-code.d.ts.map +1 -1
- package/dist/claude-code.js +14 -1
- package/dist/claude-code.js.map +1 -1
- package/dist/connector.d.ts +19 -1
- package/dist/connector.d.ts.map +1 -1
- package/dist/connector.js +46 -6
- package/dist/connector.js.map +1 -1
- package/dist/csv.d.ts +1 -1
- package/dist/csv.d.ts.map +1 -1
- package/dist/csv.js +11 -0
- package/dist/csv.js.map +1 -1
- package/dist/gateway.d.ts +2 -0
- package/dist/gateway.d.ts.map +1 -1
- package/dist/gateway.js +16 -0
- package/dist/gateway.js.map +1 -1
- package/dist/index.d.ts +2 -0
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +1 -0
- package/dist/index.js.map +1 -1
- package/dist/openai-usage.d.ts +5 -1
- package/dist/openai-usage.d.ts.map +1 -1
- package/dist/openai-usage.js +20 -4
- package/dist/openai-usage.js.map +1 -1
- package/dist/pricing-overlay.d.ts +1 -1
- package/dist/pricing-overlay.d.ts.map +1 -1
- package/dist/pricing-overlay.js +11 -0
- package/dist/pricing-overlay.js.map +1 -1
- package/dist/pricing.d.ts +8 -0
- package/dist/pricing.d.ts.map +1 -1
- package/dist/pricing.js +74 -17
- package/dist/pricing.js.map +1 -1
- package/dist/receipt.d.ts +38 -4
- package/dist/receipt.d.ts.map +1 -1
- package/dist/receipt.js +19 -1
- package/dist/receipt.js.map +1 -1
- package/dist/repeats.d.ts.map +1 -1
- package/dist/repeats.js +3 -2
- package/dist/repeats.js.map +1 -1
- package/dist/rollup.d.ts.map +1 -1
- package/dist/rollup.js +4 -0
- package/dist/rollup.js.map +1 -1
- package/dist/session-cost.d.ts.map +1 -1
- package/dist/session-cost.js +4 -2
- package/dist/session-cost.js.map +1 -1
- package/dist/store.d.ts +5 -0
- package/dist/store.d.ts.map +1 -1
- package/dist/store.js +4 -0
- package/dist/store.js.map +1 -1
- package/dist/truncation-retry.d.ts.map +1 -1
- package/dist/truncation-retry.js +3 -2
- package/dist/truncation-retry.js.map +1 -1
- package/dist/types.d.ts +13 -0
- package/dist/types.d.ts.map +1 -1
- package/dist/usage.d.ts +42 -0
- package/dist/usage.d.ts.map +1 -1
- package/dist/usage.js +42 -3
- package/dist/usage.js.map +1 -1
- package/package.json +1 -1
- package/src/anthropic-cost.ts +46 -12
- package/src/anthropic-usage.ts +54 -7
- package/src/batch-results.ts +174 -0
- package/src/billing.ts +16 -0
- package/src/claude-code.ts +16 -1
- package/src/connector.ts +68 -8
- package/src/csv.ts +11 -0
- package/src/gateway.ts +16 -0
- package/src/index.ts +2 -0
- package/src/openai-usage.ts +24 -5
- package/src/pricing-overlay.ts +15 -0
- package/src/pricing.ts +83 -17
- package/src/receipt.ts +53 -6
- package/src/repeats.ts +3 -2
- package/src/rollup.ts +4 -0
- package/src/session-cost.ts +4 -2
- package/src/store.ts +9 -0
- package/src/truncation-retry.ts +3 -2
- package/src/types.ts +13 -0
- package/src/usage.ts +84 -3
|
@@ -0,0 +1,174 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* A provider's Batch API results file, read as a usage log.
|
|
3
|
+
*
|
|
4
|
+
* Both providers return a batch's results as JSON Lines, one per request, and
|
|
5
|
+
* bill every one that ran at half the standard price. Neither shape is a usage
|
|
6
|
+
* line: the model and usage sit one or two objects down, so `parseUsageLine`
|
|
7
|
+
* read a whole results file as nothing at all. This converter takes each line
|
|
8
|
+
* that was billed, keeps its model, usage and end reason, marks it `batch`, and
|
|
9
|
+
* drops everything else, the generated text first among it.
|
|
10
|
+
*
|
|
11
|
+
* ## The two shapes, each from its provider's own page
|
|
12
|
+
*
|
|
13
|
+
* **Anthropic** (platform.claude.com/docs/en/build-with-claude/batch-processing,
|
|
14
|
+
* read 2026-11-05): `{"custom_id", "result": {"type", "message"}}`. `type` is
|
|
15
|
+
* `succeeded`, `errored`, `canceled` or `expired`, and the page says of the
|
|
16
|
+
* last three "You will not be billed for these requests". A succeeded result's
|
|
17
|
+
* `message` is a Messages response, `usage` and all.
|
|
18
|
+
*
|
|
19
|
+
* **OpenAI** (developers.openai.com/api/docs/guides/batch, read 2026-11-05):
|
|
20
|
+
* `{"id", "custom_id", "response": {"status_code", "request_id", "body"},
|
|
21
|
+
* "error"}`. `body` is the endpoint's own response, a Responses or Chat
|
|
22
|
+
* Completions object. An expired request is written with `"response": null`
|
|
23
|
+
* and an `error`; the page says "You will be charged for tokens consumed from
|
|
24
|
+
* any completed requests", so a line is billed only when it carries a 200 and
|
|
25
|
+
* a body with usage.
|
|
26
|
+
*
|
|
27
|
+
* ## What it does not invent
|
|
28
|
+
*
|
|
29
|
+
* **No time.** An Anthropic result carries none; it is counted as undated, the
|
|
30
|
+
* way `bill` already names an undated call, rather than stamped with the day
|
|
31
|
+
* the file was read. An OpenAI body's `created` is Unix seconds and is kept.
|
|
32
|
+
*
|
|
33
|
+
* **No price.** The record says `batch: true` and the model; the batch ratio is
|
|
34
|
+
* the catalogue's, applied where every other call is priced.
|
|
35
|
+
*
|
|
36
|
+
* **No content.** The message's `content`, the body's `output` and `choices`,
|
|
37
|
+
* the `custom_id` (which a caller may have filled with anything) and every id
|
|
38
|
+
* are dropped. A test plants a marker in each and greps the output.
|
|
39
|
+
*/
|
|
40
|
+
|
|
41
|
+
/** One converted record, in the shape `parseUsageLine` reads. */
|
|
42
|
+
export interface BatchResultRecord {
|
|
43
|
+
model: string;
|
|
44
|
+
ts?: string;
|
|
45
|
+
label?: string;
|
|
46
|
+
/** Always true: every line here ran on the Batch API. */
|
|
47
|
+
batch: true;
|
|
48
|
+
stop_reason?: string;
|
|
49
|
+
usage: Record<string, unknown>;
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
export interface BatchResultConversion {
|
|
53
|
+
records: BatchResultRecord[];
|
|
54
|
+
/** Which provider's shape the file was, or null when no line was either. */
|
|
55
|
+
provider: 'anthropic' | 'openai' | null;
|
|
56
|
+
/** Non-blank lines read. */
|
|
57
|
+
lines: number;
|
|
58
|
+
/** Results the provider says it did not bill: errored, canceled, expired, or not a 200. */
|
|
59
|
+
notBilled: number;
|
|
60
|
+
/** Billed results with no model or usage to price them by. */
|
|
61
|
+
unpriceable: number;
|
|
62
|
+
/** Lines that were not JSON, or neither shape. */
|
|
63
|
+
unparseable: number;
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
const asObject = (value: unknown): Record<string, unknown> | null =>
|
|
67
|
+
typeof value === 'object' && value !== null && !Array.isArray(value)
|
|
68
|
+
? (value as Record<string, unknown>)
|
|
69
|
+
: null;
|
|
70
|
+
|
|
71
|
+
/** Usage as the response gave it, minus nothing it bills by and plus nothing. */
|
|
72
|
+
const usageOf = (value: unknown): Record<string, unknown> | null => asObject(value);
|
|
73
|
+
|
|
74
|
+
/**
|
|
75
|
+
* @param label the project these results belong to, chosen by the operator.
|
|
76
|
+
*/
|
|
77
|
+
export function batchResultRecords(text: string, options: { label?: string } = {}): BatchResultConversion {
|
|
78
|
+
const records: BatchResultRecord[] = [];
|
|
79
|
+
let provider: BatchResultConversion['provider'] = null;
|
|
80
|
+
let lines = 0;
|
|
81
|
+
let notBilled = 0;
|
|
82
|
+
let unpriceable = 0;
|
|
83
|
+
let unparseable = 0;
|
|
84
|
+
|
|
85
|
+
for (const raw of text.split('\n')) {
|
|
86
|
+
if (raw.trim() === '') continue;
|
|
87
|
+
lines += 1;
|
|
88
|
+
let parsed: unknown;
|
|
89
|
+
try {
|
|
90
|
+
parsed = JSON.parse(raw);
|
|
91
|
+
} catch {
|
|
92
|
+
unparseable += 1;
|
|
93
|
+
continue;
|
|
94
|
+
}
|
|
95
|
+
const line = asObject(parsed);
|
|
96
|
+
if (line === null || typeof line.custom_id !== 'string') {
|
|
97
|
+
unparseable += 1;
|
|
98
|
+
continue;
|
|
99
|
+
}
|
|
100
|
+
|
|
101
|
+
const result = asObject(line.result);
|
|
102
|
+
if (result !== null && typeof result.type === 'string') {
|
|
103
|
+
provider ??= 'anthropic';
|
|
104
|
+
if (result.type !== 'succeeded') {
|
|
105
|
+
notBilled += 1;
|
|
106
|
+
continue;
|
|
107
|
+
}
|
|
108
|
+
const message = asObject(result.message);
|
|
109
|
+
const usage = usageOf(message?.usage);
|
|
110
|
+
if (message === null || typeof message.model !== 'string' || usage === null) {
|
|
111
|
+
unpriceable += 1;
|
|
112
|
+
continue;
|
|
113
|
+
}
|
|
114
|
+
records.push({
|
|
115
|
+
model: message.model,
|
|
116
|
+
...(options.label === undefined ? {} : { label: options.label }),
|
|
117
|
+
batch: true,
|
|
118
|
+
...(typeof message.stop_reason === 'string' ? { stop_reason: message.stop_reason } : {}),
|
|
119
|
+
usage,
|
|
120
|
+
});
|
|
121
|
+
continue;
|
|
122
|
+
}
|
|
123
|
+
|
|
124
|
+
if ('response' in line) {
|
|
125
|
+
provider ??= 'openai';
|
|
126
|
+
const response = asObject(line.response);
|
|
127
|
+
if (response === null || response.status_code !== 200) {
|
|
128
|
+
notBilled += 1;
|
|
129
|
+
continue;
|
|
130
|
+
}
|
|
131
|
+
const body = asObject(response.body);
|
|
132
|
+
const usage = usageOf(body?.usage);
|
|
133
|
+
if (body === null || typeof body.model !== 'string' || usage === null) {
|
|
134
|
+
unpriceable += 1;
|
|
135
|
+
continue;
|
|
136
|
+
}
|
|
137
|
+
const created =
|
|
138
|
+
typeof body.created_at === 'number' ? body.created_at : typeof body.created === 'number' ? body.created : null;
|
|
139
|
+
records.push({
|
|
140
|
+
model: body.model,
|
|
141
|
+
...(created !== null && Number.isFinite(created) ? { ts: new Date(created * 1000).toISOString() } : {}),
|
|
142
|
+
...(options.label === undefined ? {} : { label: options.label }),
|
|
143
|
+
batch: true,
|
|
144
|
+
usage,
|
|
145
|
+
});
|
|
146
|
+
continue;
|
|
147
|
+
}
|
|
148
|
+
|
|
149
|
+
unparseable += 1;
|
|
150
|
+
}
|
|
151
|
+
|
|
152
|
+
return { records, provider, lines, notBilled, unpriceable, unparseable };
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
/**
|
|
156
|
+
* Whether this text is a Batch API results file.
|
|
157
|
+
*
|
|
158
|
+
* Every line carries a `custom_id`, which no other log here does, beside
|
|
159
|
+
* either Anthropic's `result` with a `type` or OpenAI's `response` with a
|
|
160
|
+
* `status_code`. The first line decides; one that is neither is not this.
|
|
161
|
+
*/
|
|
162
|
+
export function looksLikeBatchResults(text: string, prefixBytes = 8192): boolean {
|
|
163
|
+
const first = text.slice(0, prefixBytes).split('\n').find((line) => line.trim() !== '');
|
|
164
|
+
if (first === undefined || !first.includes('"custom_id"')) return false;
|
|
165
|
+
try {
|
|
166
|
+
const line = asObject(JSON.parse(first));
|
|
167
|
+
if (line === null || typeof line.custom_id !== 'string') return false;
|
|
168
|
+
const result = asObject(line.result);
|
|
169
|
+
if (result !== null && typeof result.type === 'string') return true;
|
|
170
|
+
return 'response' in line && ('error' in line || asObject(line.response)?.status_code !== undefined);
|
|
171
|
+
} catch {
|
|
172
|
+
return false;
|
|
173
|
+
}
|
|
174
|
+
}
|
package/src/billing.ts
CHANGED
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
import type { PromptSize } from './pricing.js';
|
|
2
|
+
import type { ModelPricing } from './types.js';
|
|
2
3
|
import type { UsageRecord } from './usage.js';
|
|
3
4
|
|
|
4
5
|
/*
|
|
@@ -37,6 +38,21 @@ export function promptSizeOf(record: UsageRecord): PromptSize {
|
|
|
37
38
|
...(record.speed === undefined ? {} : { speed: record.speed }),
|
|
38
39
|
...(record.serviceTier === undefined ? {} : { serviceTier: record.serviceTier }),
|
|
39
40
|
...(record.inferenceGeo === undefined ? {} : { inferenceGeo: record.inferenceGeo }),
|
|
41
|
+
...(record.batch === true ? { batch: true } : {}),
|
|
40
42
|
};
|
|
41
43
|
return record.requests === undefined ? { inputTokens: prompt, ...how } : { inputTokensAtMost: prompt, ...how };
|
|
42
44
|
}
|
|
45
|
+
|
|
46
|
+
/**
|
|
47
|
+
* What a record's web searches cost: per search, at the row's fee, flat.
|
|
48
|
+
*
|
|
49
|
+
* Here beside `billedAt` and `promptSizeOf` for the same reason they are: the
|
|
50
|
+
* total and every tracker under it price a call, and a finding that costs a
|
|
51
|
+
* call without its searches disagrees with the bill it sits under. Zero for a
|
|
52
|
+
* row with no fee, whose searches the total names as unpriced instead.
|
|
53
|
+
*/
|
|
54
|
+
export function webSearchUsdOf(record: UsageRecord, model: ModelPricing): number {
|
|
55
|
+
const searches = record.webSearchRequests ?? 0;
|
|
56
|
+
if (searches <= 0 || model.webSearchPerThousandUsd === undefined) return 0;
|
|
57
|
+
return (searches / 1_000) * model.webSearchPerThousandUsd;
|
|
58
|
+
}
|
package/src/claude-code.ts
CHANGED
|
@@ -78,6 +78,8 @@ export interface ClaudeCodeRecord {
|
|
|
78
78
|
speed?: string;
|
|
79
79
|
/** The response's `usage.inference_geo`, kept because US-only inference bills 1.1x on Claude 4.6 and later. */
|
|
80
80
|
inference_geo?: string;
|
|
81
|
+
/** The searches the call ran, kept because each is billed at $10 per 1,000 on top of the tokens. */
|
|
82
|
+
server_tool_use?: { web_search_requests: number };
|
|
81
83
|
};
|
|
82
84
|
}
|
|
83
85
|
|
|
@@ -291,6 +293,18 @@ export function claudeCodeRecords(
|
|
|
291
293
|
...(typeof usage.inference_geo === 'string' && usage.inference_geo !== ''
|
|
292
294
|
? { inference_geo: usage.inference_geo }
|
|
293
295
|
: {}),
|
|
296
|
+
/*
|
|
297
|
+
* And the searches: Claude Code's web search is the API's server-side
|
|
298
|
+
* tool, billed at $10 per 1,000 searches on top of the tokens it
|
|
299
|
+
* reads, and the response counts them in
|
|
300
|
+
* `usage.server_tool_use.web_search_requests`. Rebuilding the usage
|
|
301
|
+
* without it billed every searching session as if it had searched
|
|
302
|
+
* for nothing. Only the count crosses; what was searched for is in
|
|
303
|
+
* the content, which never does.
|
|
304
|
+
*/
|
|
305
|
+
...((asCount(asObject(usage.server_tool_use)?.web_search_requests) ?? 0) > 0
|
|
306
|
+
? { server_tool_use: { web_search_requests: asCount(asObject(usage.server_tool_use)?.web_search_requests)! } }
|
|
307
|
+
: {}),
|
|
294
308
|
},
|
|
295
309
|
};
|
|
296
310
|
|
|
@@ -326,7 +340,8 @@ export function claudeCodeRecords(
|
|
|
326
340
|
record.usage.input_tokens >= before.input_tokens &&
|
|
327
341
|
record.usage.output_tokens >= before.output_tokens &&
|
|
328
342
|
(record.usage.cache_read_input_tokens ?? 0) >= (before.cache_read_input_tokens ?? 0) &&
|
|
329
|
-
(record.usage.cache_creation_input_tokens ?? 0) >= (before.cache_creation_input_tokens ?? 0)
|
|
343
|
+
(record.usage.cache_creation_input_tokens ?? 0) >= (before.cache_creation_input_tokens ?? 0) &&
|
|
344
|
+
(record.usage.server_tool_use?.web_search_requests ?? 0) >= (before.server_tool_use?.web_search_requests ?? 0);
|
|
330
345
|
if (grewOnly) out.streamed += 1;
|
|
331
346
|
else out.disagreements += 1;
|
|
332
347
|
seenUsage.set(requestId, record.usage);
|
package/src/connector.ts
CHANGED
|
@@ -178,6 +178,12 @@ export interface UsageBucket {
|
|
|
178
178
|
/** False when the provider reported writes without saying which TTL. */
|
|
179
179
|
writeTtlKnown: boolean;
|
|
180
180
|
outputTokens: number;
|
|
181
|
+
/**
|
|
182
|
+
* Web searches the bucket's requests ran, billed per search on top of the
|
|
183
|
+
* tokens: Anthropic's `server_tool_use.web_search_requests`. Absent is none,
|
|
184
|
+
* which is also what a pull stored before 2.7.5 reads as.
|
|
185
|
+
*/
|
|
186
|
+
webSearchRequests?: number;
|
|
181
187
|
/** Whatever the provider grouped by beyond the model — workspace, key, tier. */
|
|
182
188
|
group: Record<string, string>;
|
|
183
189
|
}
|
|
@@ -198,7 +204,9 @@ export interface PullGap {
|
|
|
198
204
|
| 'unreadable-entry'
|
|
199
205
|
| 'unreadable-field'
|
|
200
206
|
/** Usage at a tier no catalogue rate prices, left out and named. */
|
|
201
|
-
| 'unpriced-tier'
|
|
207
|
+
| 'unpriced-tier'
|
|
208
|
+
/** Web searches on a model whose row publishes no per-search fee, left out and named. */
|
|
209
|
+
| 'unpriced-web-search';
|
|
202
210
|
detail: string;
|
|
203
211
|
}
|
|
204
212
|
|
|
@@ -241,6 +249,9 @@ function collect(buckets: UsageBucket[]): UsageBucket[] {
|
|
|
241
249
|
seen.cacheWrite5mTokens += bucket.cacheWrite5mTokens;
|
|
242
250
|
seen.cacheWrite1hTokens += bucket.cacheWrite1hTokens;
|
|
243
251
|
seen.outputTokens += bucket.outputTokens;
|
|
252
|
+
if ((bucket.webSearchRequests ?? 0) > 0) {
|
|
253
|
+
seen.webSearchRequests = (seen.webSearchRequests ?? 0) + bucket.webSearchRequests!;
|
|
254
|
+
}
|
|
244
255
|
seen.writeTtlKnown = seen.writeTtlKnown && bucket.writeTtlKnown;
|
|
245
256
|
// A count merged with an absent count is still absent: adding a number to
|
|
246
257
|
// "unknown" produces a number that describes only part of the traffic.
|
|
@@ -308,6 +319,8 @@ export function normalizeAnthropicUsage(payload: unknown): ConnectorPull {
|
|
|
308
319
|
const write1h = num(creation.ephemeral_1h_input_tokens) ?? 0;
|
|
309
320
|
const flatWrite = num(r.cache_creation_input_tokens) ?? 0;
|
|
310
321
|
const ttlKnown = !(flatWrite > 0 && write5m === 0 && write1h === 0);
|
|
322
|
+
/* Billed per search, $10 per 1,000 on Claude: the one count here that is not tokens. */
|
|
323
|
+
const searches = num((r.server_tool_use as Record<string, unknown> | undefined)?.web_search_requests) ?? 0;
|
|
311
324
|
buckets.push({
|
|
312
325
|
fromMs,
|
|
313
326
|
toMs,
|
|
@@ -320,6 +333,7 @@ export function normalizeAnthropicUsage(payload: unknown): ConnectorPull {
|
|
|
320
333
|
cacheWrite1hTokens: ttlKnown ? write1h : 0,
|
|
321
334
|
writeTtlKnown: ttlKnown,
|
|
322
335
|
outputTokens: num(r.output_tokens) ?? 0,
|
|
336
|
+
...(searches > 0 ? { webSearchRequests: searches } : {}),
|
|
323
337
|
group: groupOf(r, ['workspace_id', 'api_key_id', 'service_tier', 'context_window', 'speed', 'inference_geo']),
|
|
324
338
|
});
|
|
325
339
|
}
|
|
@@ -461,6 +475,13 @@ export interface BucketedSlice {
|
|
|
461
475
|
tierCeilingBuckets: number;
|
|
462
476
|
tierCeilingUsd: number;
|
|
463
477
|
tierFloorUsd: number;
|
|
478
|
+
/**
|
|
479
|
+
* Web searches, billed per search at the model's fee, flat: the Batch API,
|
|
480
|
+
* fast mode and US-only inference move token rates, not this. Part of
|
|
481
|
+
* `totalUsd`.
|
|
482
|
+
*/
|
|
483
|
+
webSearchRequests: number;
|
|
484
|
+
webSearchUsd: number;
|
|
464
485
|
}
|
|
465
486
|
|
|
466
487
|
export interface BucketedReport {
|
|
@@ -476,6 +497,9 @@ export interface BucketedReport {
|
|
|
476
497
|
cacheReadTokens: number;
|
|
477
498
|
cacheWriteTokens: number;
|
|
478
499
|
outputTokens: number;
|
|
500
|
+
/** Web searches billed, and what they came to: already inside `totalUsd`. */
|
|
501
|
+
webSearchRequests: number;
|
|
502
|
+
webSearchUsd: number;
|
|
479
503
|
};
|
|
480
504
|
byModel: BucketedSlice[];
|
|
481
505
|
/** Spend per UTC day, oldest first — the shape a total hides. */
|
|
@@ -521,6 +545,8 @@ export function bucketedProfile(
|
|
|
521
545
|
+ bucket.outputTokens;
|
|
522
546
|
leftOut.set(key, seen);
|
|
523
547
|
};
|
|
548
|
+
/* Searches on a model with no per-search fee, per model, so the gap names it. */
|
|
549
|
+
const searchesLeftOut = new Map<string, number>();
|
|
524
550
|
|
|
525
551
|
for (const bucket of pull.buckets) {
|
|
526
552
|
const model = catalogue.byId.get(bucket.model);
|
|
@@ -540,10 +566,14 @@ export function bucketedProfile(
|
|
|
540
566
|
* them reads as standard, as it always did.
|
|
541
567
|
*/
|
|
542
568
|
const group = bucket.group;
|
|
543
|
-
const how: { speed?: string; serviceTier?: string; inferenceGeo?: string } = {};
|
|
569
|
+
const how: { speed?: string; serviceTier?: string; inferenceGeo?: string; batch?: boolean } = {};
|
|
570
|
+
/* The Batch API at the row's batch ratio, where the row has one. */
|
|
571
|
+
const batchRate = multipliersFor(model).batch !== null;
|
|
544
572
|
if (pull.provider === 'anthropic') {
|
|
545
573
|
const tier = group.service_tier;
|
|
546
|
-
if (tier
|
|
574
|
+
if (tier === 'batch' && batchRate) {
|
|
575
|
+
how.batch = true;
|
|
576
|
+
} else if (tier !== undefined && tier !== 'standard') {
|
|
547
577
|
leave(bucket, tier);
|
|
548
578
|
continue;
|
|
549
579
|
}
|
|
@@ -558,11 +588,14 @@ export function bucketedProfile(
|
|
|
558
588
|
if (group.inference_geo !== undefined) how.inferenceGeo = group.inference_geo;
|
|
559
589
|
} else if (pull.provider === 'openai') {
|
|
560
590
|
if (group.batch === 'true') {
|
|
561
|
-
|
|
562
|
-
|
|
591
|
+
if (!batchRate) {
|
|
592
|
+
leave(bucket, 'batch');
|
|
593
|
+
continue;
|
|
594
|
+
}
|
|
595
|
+
how.batch = true;
|
|
563
596
|
}
|
|
564
597
|
const tier = group.service_tier;
|
|
565
|
-
if (tier !== undefined && tier !== 'default') {
|
|
598
|
+
if (tier !== undefined && tier !== 'default' && !(how.batch === true && tier === 'batch')) {
|
|
566
599
|
const named = serviceTierOf(tier);
|
|
567
600
|
if (named === undefined || model.serviceTiers?.[named] === undefined) {
|
|
568
601
|
leave(bucket, tier);
|
|
@@ -606,6 +639,17 @@ export function bucketedProfile(
|
|
|
606
639
|
const ifAssumed1h = bucket.writeTtlKnown
|
|
607
640
|
? cacheWriteUsd
|
|
608
641
|
: per(writeTokens, inputPerMTok * rates.cacheWrite1h);
|
|
642
|
+
/*
|
|
643
|
+
* Searches at the row's fee, flat, read off the model and not out of
|
|
644
|
+
* `effectivePricing`: none of the multipliers `how` carries reaches a
|
|
645
|
+
* search. A model with no fee leaves them out, named below.
|
|
646
|
+
*/
|
|
647
|
+
const searches = bucket.webSearchRequests ?? 0;
|
|
648
|
+
const fee = model.webSearchPerThousandUsd;
|
|
649
|
+
const searchUsd = searches > 0 && fee !== undefined ? (searches / 1_000) * fee : 0;
|
|
650
|
+
if (searches > 0 && fee === undefined) {
|
|
651
|
+
searchesLeftOut.set(bucket.model, (searchesLeftOut.get(bucket.model) ?? 0) + searches);
|
|
652
|
+
}
|
|
609
653
|
|
|
610
654
|
const slice = slices.get(bucket.model) ?? {
|
|
611
655
|
model: bucket.model,
|
|
@@ -625,6 +669,8 @@ export function bucketedProfile(
|
|
|
625
669
|
tierCeilingBuckets: 0,
|
|
626
670
|
tierCeilingUsd: 0,
|
|
627
671
|
tierFloorUsd: 0,
|
|
672
|
+
webSearchRequests: 0,
|
|
673
|
+
webSearchUsd: 0,
|
|
628
674
|
};
|
|
629
675
|
slice.calls = slice.calls === null || bucket.calls === null ? null : slice.calls + bucket.calls;
|
|
630
676
|
slice.inputTokens += bucket.inputTokens;
|
|
@@ -635,7 +681,11 @@ export function bucketedProfile(
|
|
|
635
681
|
slice.cacheReadUsd += cacheReadUsd;
|
|
636
682
|
slice.cacheWriteUsd += cacheWriteUsd;
|
|
637
683
|
slice.outputUsd += outputUsd;
|
|
638
|
-
slice.totalUsd += inputUsd + cacheReadUsd + cacheWriteUsd + outputUsd;
|
|
684
|
+
slice.totalUsd += inputUsd + cacheReadUsd + cacheWriteUsd + outputUsd + searchUsd;
|
|
685
|
+
if (fee !== undefined) {
|
|
686
|
+
slice.webSearchRequests += searches;
|
|
687
|
+
slice.webSearchUsd += searchUsd;
|
|
688
|
+
}
|
|
639
689
|
slice.cachedTokensAtInputRateUsd += atInputRate;
|
|
640
690
|
slice.cacheWriteUsdIfAssumed1h += ifAssumed1h;
|
|
641
691
|
slice.writeTtlKnown = slice.writeTtlKnown && bucket.writeTtlKnown;
|
|
@@ -648,7 +698,7 @@ export function bucketedProfile(
|
|
|
648
698
|
|
|
649
699
|
const day = new Date(bucket.fromMs).toISOString().slice(0, 10);
|
|
650
700
|
const entry = days.get(day) ?? { usd: 0, calls: bucket.calls === null ? null : 0 };
|
|
651
|
-
entry.usd += inputUsd + cacheReadUsd + cacheWriteUsd + outputUsd;
|
|
701
|
+
entry.usd += inputUsd + cacheReadUsd + cacheWriteUsd + outputUsd + searchUsd;
|
|
652
702
|
entry.calls = entry.calls === null || bucket.calls === null ? null : entry.calls + bucket.calls;
|
|
653
703
|
days.set(day, entry);
|
|
654
704
|
}
|
|
@@ -668,6 +718,8 @@ export function bucketedProfile(
|
|
|
668
718
|
cacheReadTokens: byModel.reduce((sum, s) => sum + s.cacheReadTokens, 0),
|
|
669
719
|
cacheWriteTokens: byModel.reduce((sum, s) => sum + s.cacheWriteTokens, 0),
|
|
670
720
|
outputTokens: byModel.reduce((sum, s) => sum + s.outputTokens, 0),
|
|
721
|
+
webSearchRequests: byModel.reduce((sum, s) => sum + s.webSearchRequests, 0),
|
|
722
|
+
webSearchUsd: byModel.reduce((sum, s) => sum + s.webSearchUsd, 0),
|
|
671
723
|
},
|
|
672
724
|
byModel,
|
|
673
725
|
byDay: [...days.entries()]
|
|
@@ -686,6 +738,14 @@ export function bucketedProfile(
|
|
|
686
738
|
+ `and ${left.tokens.toLocaleString('en-US')} token(s) were left out rather than priced at the standard rate`,
|
|
687
739
|
}),
|
|
688
740
|
),
|
|
741
|
+
...[...searchesLeftOut.entries()].map(
|
|
742
|
+
([model, searches]): PullGap => ({
|
|
743
|
+
kind: 'unpriced-web-search',
|
|
744
|
+
detail:
|
|
745
|
+
`${searches.toLocaleString('en-US')} web search(es) on ${model}, whose row publishes no per-search fee, `
|
|
746
|
+
+ 'were left out rather than billed at another provider\'s',
|
|
747
|
+
}),
|
|
748
|
+
),
|
|
689
749
|
],
|
|
690
750
|
unavailable: pull.unavailable,
|
|
691
751
|
};
|
package/src/csv.ts
CHANGED
|
@@ -45,6 +45,13 @@ export const PROFILE_CSV_COLUMNS = [
|
|
|
45
45
|
'cache_write_usd',
|
|
46
46
|
'output_usd',
|
|
47
47
|
'total_usd',
|
|
48
|
+
/*
|
|
49
|
+
* Appended, never inserted, so a reader that addresses columns by position
|
|
50
|
+
* keeps reading the same twelve. Web searches are billed per search, so
|
|
51
|
+
* `total_usd` is the four token columns plus `web_search_usd`.
|
|
52
|
+
*/
|
|
53
|
+
'web_search_requests',
|
|
54
|
+
'web_search_usd',
|
|
48
55
|
] as const;
|
|
49
56
|
|
|
50
57
|
/**
|
|
@@ -164,6 +171,8 @@ export function profileToCsv(report: UsageProfileReport, options: ProfileCsvOpti
|
|
|
164
171
|
usd(breakdown.cacheWriteUsd),
|
|
165
172
|
usd(breakdown.outputUsd),
|
|
166
173
|
usd(breakdown.totalUsd),
|
|
174
|
+
String(breakdown.webSearchRequests),
|
|
175
|
+
usd(breakdown.webSearchUsd),
|
|
167
176
|
].join(','),
|
|
168
177
|
);
|
|
169
178
|
}
|
|
@@ -193,6 +202,8 @@ export function profileToCsv(report: UsageProfileReport, options: ProfileCsvOpti
|
|
|
193
202
|
'',
|
|
194
203
|
'',
|
|
195
204
|
'',
|
|
205
|
+
String(row.breakdown.webSearchRequests),
|
|
206
|
+
'',
|
|
196
207
|
].join(','),
|
|
197
208
|
);
|
|
198
209
|
}
|
package/src/gateway.ts
CHANGED
|
@@ -523,8 +523,18 @@ export interface GatewayUsage {
|
|
|
523
523
|
serviceTier?: string;
|
|
524
524
|
/** Anthropic's `usage.inference_geo`, the geography that ran: "us" bills 1.1x on Claude 4.6 and later. */
|
|
525
525
|
inferenceGeo?: string;
|
|
526
|
+
/** Anthropic's `usage.server_tool_use.web_search_requests`: each search bills $10 per 1,000 on top. */
|
|
527
|
+
webSearchRequests?: number;
|
|
526
528
|
}
|
|
527
529
|
|
|
530
|
+
/** The searches a usage object counts, when it counts a readable number of them. */
|
|
531
|
+
const webSearchesIn = (usage: Record<string, unknown>): number | undefined => {
|
|
532
|
+
const tools = usage.server_tool_use;
|
|
533
|
+
if (typeof tools !== 'object' || tools === null) return undefined;
|
|
534
|
+
const searches = (tools as Record<string, unknown>).web_search_requests;
|
|
535
|
+
return typeof searches === 'number' && Number.isFinite(searches) && searches >= 0 ? searches : undefined;
|
|
536
|
+
};
|
|
537
|
+
|
|
528
538
|
/**
|
|
529
539
|
* OpenAI's cache writes, inside `prompt_tokens` like its cached reads.
|
|
530
540
|
*
|
|
@@ -693,6 +703,8 @@ export function streamingUsageReader(
|
|
|
693
703
|
usage.outputTokens = num(start.output_tokens);
|
|
694
704
|
if (typeof start.speed === 'string' && start.speed !== '') usage.speed = start.speed;
|
|
695
705
|
if (typeof start.inference_geo === 'string' && start.inference_geo !== '') usage.inferenceGeo = start.inference_geo;
|
|
706
|
+
const searches = webSearchesIn(start);
|
|
707
|
+
if (searches !== undefined) usage.webSearchRequests = searches;
|
|
696
708
|
seen = true;
|
|
697
709
|
}
|
|
698
710
|
}
|
|
@@ -711,6 +723,9 @@ export function streamingUsageReader(
|
|
|
711
723
|
if (typeof d.input_tokens === 'number') usage.inputTokens = num(d.input_tokens);
|
|
712
724
|
if (typeof d.speed === 'string' && d.speed !== '') usage.speed = d.speed;
|
|
713
725
|
if (typeof d.inference_geo === 'string' && d.inference_geo !== '') usage.inferenceGeo = d.inference_geo;
|
|
726
|
+
// Cumulative like the output count: the last delta's searches are the call's.
|
|
727
|
+
const searches = webSearchesIn(d);
|
|
728
|
+
if (searches !== undefined) usage.webSearchRequests = searches;
|
|
714
729
|
seen = true;
|
|
715
730
|
}
|
|
716
731
|
return;
|
|
@@ -814,6 +829,7 @@ export function usageFromResponse(
|
|
|
814
829
|
...cacheWriteBreakdown(u),
|
|
815
830
|
...(typeof u.speed === 'string' && u.speed !== '' ? { speed: u.speed } : {}),
|
|
816
831
|
...(typeof u.inference_geo === 'string' && u.inference_geo !== '' ? { inferenceGeo: u.inference_geo } : {}),
|
|
832
|
+
...(webSearchesIn(u) === undefined ? {} : { webSearchRequests: webSearchesIn(u) }),
|
|
817
833
|
};
|
|
818
834
|
}
|
|
819
835
|
if (shape === 'openai') {
|
package/src/index.ts
CHANGED
|
@@ -242,6 +242,8 @@ export type { AnthropicCostReading, BilledReading, Reconciliation } from './anth
|
|
|
242
242
|
export { openaiCostReport, looksLikeOpenaiCost } from './openai-cost.js';
|
|
243
243
|
export type { OpenaiCostReading } from './openai-cost.js';
|
|
244
244
|
export { openaiUsageRecords, looksLikeOpenaiUsage } from './openai-usage.js';
|
|
245
|
+
export { batchResultRecords, looksLikeBatchResults } from './batch-results.js';
|
|
246
|
+
export type { BatchResultConversion, BatchResultRecord } from './batch-results.js';
|
|
245
247
|
export { openrouterActivityRecords, looksLikeOpenrouterActivity } from './openrouter-activity.js';
|
|
246
248
|
export type {
|
|
247
249
|
OpenrouterActivityConversion,
|
package/src/openai-usage.ts
CHANGED
|
@@ -94,7 +94,7 @@
|
|
|
94
94
|
* absence and no second reading to derive.
|
|
95
95
|
*/
|
|
96
96
|
|
|
97
|
-
import { BUNDLED_CATALOGUE, serviceTierOf } from './pricing.js';
|
|
97
|
+
import { BUNDLED_CATALOGUE, multipliersFor, serviceTierOf } from './pricing.js';
|
|
98
98
|
import type { PricingCatalogue } from './pricing.js';
|
|
99
99
|
|
|
100
100
|
/** One project, and the label the operator gives it. Matched exactly. */
|
|
@@ -121,6 +121,8 @@ export interface OpenaiUsageRecord {
|
|
|
121
121
|
* the default tier.
|
|
122
122
|
*/
|
|
123
123
|
service_tier?: string;
|
|
124
|
+
/** True when the row ran on the Batch API: billed at the row's batch ratio. */
|
|
125
|
+
batch?: true;
|
|
124
126
|
usage: {
|
|
125
127
|
prompt_tokens: number;
|
|
126
128
|
completion_tokens: number;
|
|
@@ -143,8 +145,10 @@ export interface OpenaiUsageConversion {
|
|
|
143
145
|
requests: number;
|
|
144
146
|
/** Results with `model: null`, which nothing can price. */
|
|
145
147
|
unnamedModel: number;
|
|
146
|
-
/** Results with `batch: true
|
|
148
|
+
/** Results with `batch: true` on a model with no batch rate, left out. */
|
|
147
149
|
batch: number;
|
|
150
|
+
/** Results with `batch: true` that became records priced at the row's batch ratio. */
|
|
151
|
+
batchRows: number;
|
|
148
152
|
/** Whether any result said whether it was batch at all. */
|
|
149
153
|
batchNamed: boolean;
|
|
150
154
|
/**
|
|
@@ -229,6 +233,7 @@ export function openaiUsageRecords(
|
|
|
229
233
|
requests: 0,
|
|
230
234
|
unnamedModel: 0,
|
|
231
235
|
batch: 0,
|
|
236
|
+
batchRows: 0,
|
|
232
237
|
batchNamed: false,
|
|
233
238
|
nonDefaultTier: 0,
|
|
234
239
|
tiersRefused: [],
|
|
@@ -267,6 +272,7 @@ export function openaiUsageRecords(
|
|
|
267
272
|
let unnamedModel = 0;
|
|
268
273
|
let batch = 0;
|
|
269
274
|
let batchNamed = false;
|
|
275
|
+
let batchRows = 0;
|
|
270
276
|
let nonDefaultTier = 0;
|
|
271
277
|
let serviceTierRows = 0;
|
|
272
278
|
let tierNamed = false;
|
|
@@ -295,18 +301,28 @@ export function openaiUsageRecords(
|
|
|
295
301
|
if (result === null) continue;
|
|
296
302
|
requests += count(result.num_model_requests);
|
|
297
303
|
|
|
304
|
+
/*
|
|
305
|
+
A batch job is billed at the row's batch ratio, half the standard rate,
|
|
306
|
+
which the pricing page publishes as its own table and `price-drift.mjs`
|
|
307
|
+
checks weekly. A batch row on a model with no batch rate is left out.
|
|
308
|
+
*/
|
|
309
|
+
let batchRow = false;
|
|
298
310
|
if (typeof result.batch === 'boolean') {
|
|
299
311
|
batchNamed = true;
|
|
300
312
|
if (result.batch) {
|
|
301
|
-
|
|
302
|
-
|
|
313
|
+
const row = typeof result.model === 'string' ? catalogue.byId.get(result.model) : undefined;
|
|
314
|
+
if (row === undefined || multipliersFor(row).batch === null) {
|
|
315
|
+
batch += 1;
|
|
316
|
+
continue;
|
|
317
|
+
}
|
|
318
|
+
batchRow = true;
|
|
303
319
|
}
|
|
304
320
|
}
|
|
305
321
|
|
|
306
322
|
let serviceTier: string | undefined;
|
|
307
323
|
if (typeof result.service_tier === 'string') {
|
|
308
324
|
tierNamed = true;
|
|
309
|
-
if (result.service_tier !== 'default') {
|
|
325
|
+
if (result.service_tier !== 'default' && !(batchRow && result.service_tier === 'batch')) {
|
|
310
326
|
/*
|
|
311
327
|
Priced only from a table that exists for this model at this tier.
|
|
312
328
|
The response path falls back to the standard rate for a tier with
|
|
@@ -390,6 +406,7 @@ export function openaiUsageRecords(
|
|
|
390
406
|
...(label === undefined ? {} : { label }),
|
|
391
407
|
...requestsField(result.num_model_requests),
|
|
392
408
|
...(serviceTier === undefined ? {} : { service_tier: serviceTier }),
|
|
409
|
+
...(batchRow ? { batch: true as const } : {}),
|
|
393
410
|
usage: {
|
|
394
411
|
prompt_tokens: prompt,
|
|
395
412
|
completion_tokens: completion,
|
|
@@ -404,6 +421,7 @@ export function openaiUsageRecords(
|
|
|
404
421
|
},
|
|
405
422
|
});
|
|
406
423
|
rows += 1;
|
|
424
|
+
if (batchRow) batchRows += 1;
|
|
407
425
|
if (serviceTier !== undefined) {
|
|
408
426
|
serviceTierRows += 1;
|
|
409
427
|
tiersPriced.add(serviceTier);
|
|
@@ -419,6 +437,7 @@ export function openaiUsageRecords(
|
|
|
419
437
|
unnamedModel,
|
|
420
438
|
batch,
|
|
421
439
|
batchNamed,
|
|
440
|
+
batchRows,
|
|
422
441
|
nonDefaultTier,
|
|
423
442
|
tiersRefused: [...tiersRefused],
|
|
424
443
|
serviceTierRows,
|
package/src/pricing-overlay.ts
CHANGED
|
@@ -36,6 +36,7 @@ export const PRICING_MODEL_KEYS = [
|
|
|
36
36
|
'promo',
|
|
37
37
|
'multipliers',
|
|
38
38
|
'inferenceGeoMultipliers',
|
|
39
|
+
'webSearchPerThousandUsd',
|
|
39
40
|
] as const;
|
|
40
41
|
|
|
41
42
|
/** The geographies a provider prices differently, as `InferenceGeo` names them. */
|
|
@@ -281,6 +282,20 @@ function parseModel(
|
|
|
281
282
|
model.inferenceGeoMultipliers = {};
|
|
282
283
|
if (geo.us !== undefined) model.inferenceGeoMultipliers.us = positiveNumber(geo.us, `${label}.us`, source);
|
|
283
284
|
}
|
|
285
|
+
/*
|
|
286
|
+
* A per-search fee, in dollars per 1,000 searches, as the catalogue carries
|
|
287
|
+
* it on every Claude row. Accepted for the same two reasons: a copied row
|
|
288
|
+
* stays valid, and a contract price for searches can be stated. Zero and
|
|
289
|
+
* negatives are refused like every other rate; a row with no searches to
|
|
290
|
+
* price simply leaves the key out.
|
|
291
|
+
*/
|
|
292
|
+
if (raw.webSearchPerThousandUsd !== undefined) {
|
|
293
|
+
model.webSearchPerThousandUsd = positiveNumber(
|
|
294
|
+
raw.webSearchPerThousandUsd,
|
|
295
|
+
`models.${id}.webSearchPerThousandUsd`,
|
|
296
|
+
source,
|
|
297
|
+
);
|
|
298
|
+
}
|
|
284
299
|
|
|
285
300
|
return model;
|
|
286
301
|
}
|