@trazum/core 2.7.1 → 2.7.3

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.
Files changed (58) hide show
  1. package/dist/anthropic-usage.d.ts +58 -0
  2. package/dist/anthropic-usage.d.ts.map +1 -1
  3. package/dist/anthropic-usage.js +91 -0
  4. package/dist/anthropic-usage.js.map +1 -1
  5. package/dist/billing.d.ts.map +1 -1
  6. package/dist/billing.js +6 -2
  7. package/dist/billing.js.map +1 -1
  8. package/dist/claude-code.d.ts +2 -0
  9. package/dist/claude-code.d.ts.map +1 -1
  10. package/dist/claude-code.js +9 -0
  11. package/dist/claude-code.js.map +1 -1
  12. package/dist/connector.d.ts +3 -1
  13. package/dist/connector.d.ts.map +1 -1
  14. package/dist/connector.js +64 -5
  15. package/dist/connector.js.map +1 -1
  16. package/dist/gateway.d.ts +18 -0
  17. package/dist/gateway.d.ts.map +1 -1
  18. package/dist/gateway.js +20 -1
  19. package/dist/gateway.js.map +1 -1
  20. package/dist/index.d.ts +1 -1
  21. package/dist/index.d.ts.map +1 -1
  22. package/dist/index.js +1 -1
  23. package/dist/index.js.map +1 -1
  24. package/dist/openai-usage.d.ts +38 -6
  25. package/dist/openai-usage.d.ts.map +1 -1
  26. package/dist/openai-usage.js +44 -8
  27. package/dist/openai-usage.js.map +1 -1
  28. package/dist/pricing-overlay.d.ts +1 -1
  29. package/dist/pricing-overlay.d.ts.map +1 -1
  30. package/dist/pricing-overlay.js +21 -0
  31. package/dist/pricing-overlay.js.map +1 -1
  32. package/dist/pricing.d.ts +37 -1
  33. package/dist/pricing.d.ts.map +1 -1
  34. package/dist/pricing.js +528 -9
  35. package/dist/pricing.js.map +1 -1
  36. package/dist/store.d.ts +4 -0
  37. package/dist/store.d.ts.map +1 -1
  38. package/dist/store.js +22 -1
  39. package/dist/store.js.map +1 -1
  40. package/dist/types.d.ts +54 -0
  41. package/dist/types.d.ts.map +1 -1
  42. package/dist/usage.d.ts +16 -0
  43. package/dist/usage.d.ts.map +1 -1
  44. package/dist/usage.js +17 -0
  45. package/dist/usage.js.map +1 -1
  46. package/package.json +1 -1
  47. package/src/anthropic-usage.ts +134 -1
  48. package/src/billing.ts +6 -2
  49. package/src/claude-code.ts +11 -0
  50. package/src/connector.ts +69 -6
  51. package/src/gateway.ts +35 -2
  52. package/src/index.ts +2 -0
  53. package/src/openai-usage.ts +72 -10
  54. package/src/pricing-overlay.ts +21 -0
  55. package/src/pricing.ts +552 -10
  56. package/src/store.ts +24 -3
  57. package/src/types.ts +57 -0
  58. package/src/usage.ts +33 -0
package/dist/pricing.js CHANGED
@@ -39,10 +39,12 @@ export const PROVIDER_REVIEWED = Object.freeze({
39
39
  * which bills Opus 5.5 at $8/$40 and Opus 5 and 4.8 at $10/$50 and which
40
40
  * every response records in `usage.speed`, so fast calls had been priced at
41
41
  * half their cost; and Claude Sonnet 4.5, deprecated but still served until
42
- * 2026-11-30, which read as unpriced. Both are here now. The retired Opus
43
- * 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are on the pricing page too and are
44
- * not priced here yet: a row needs a context window, which their own pages
45
- * would give, and those were not read.
42
+ * 2026-11-30, which read as unpriced. Both are here now. So are the
43
+ * retired Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5, still on the pricing
44
+ * page and in every log from before their retirement: their own model pages
45
+ * are gone, and the context-windows page gives each the 200k window every
46
+ * Claude model outside its 1M list has. Claude Sonnet 3.7 and Haiku 3 are
47
+ * retired and absent from the pricing page, so they stay unpriced.
46
48
  */
47
49
  anthropic: '2026-11-05',
48
50
  /*
@@ -248,6 +250,17 @@ const NO_PUBLISHED_CACHE_MINIMUM = 'Caches automatically, but no minimum prefix
248
250
  * catalogue's short and long write, whatever each provider's durations.
249
251
  */
250
252
  const OPENAI_CACHE_TTL = Object.freeze({ writeMs: 30 * 60_000, longerMs: 12 * 60 * 60_000 });
253
+ /**
254
+ * OpenAI's service-tier tables, read from developers.openai.com/api/docs/pricing.md
255
+ * on 2026-11-05: the Flex, Fast and Ultrafast tables beside the Standard one,
256
+ * each model's row as printed, and its "Long context" columns where the table
257
+ * publishes them ("Short context: <=272K input tokens. Long context: >272K").
258
+ * A row whose long columns read "-" gets no long band here: GPT-5.5 and 5.4 in
259
+ * Fast, GPT-5.5 pro in Flex. See `ModelPricing.serviceTiers`.
260
+ */
261
+ const openAiLongBand = (inputPerMTok, outputPerMTok) => [
262
+ { id: 'over-272k', when: { kind: 'input-tokens-above', tokens: 272_000 }, inputPerMTok, outputPerMTok },
263
+ ];
251
264
  const OPENAI_CURRENT_NOTE = 'Priced per request by prompt length: over 272,000 input tokens, 2x input and cache and 1.5x output for'
252
265
  + ' the whole request. Cache writes cost 1.25x input for the default 30-minute entry and 2.5x for a'
253
266
  + ' 12-hour one (Responses API).';
@@ -257,11 +270,26 @@ const OPENAI_EARLIER_LONG_NOTE = 'Priced per request by prompt length: over 272,
257
270
  + OPENAI_EARLIER_NOTE;
258
271
  const OPENAI_EARLIER_NO_CACHE_NOTE = 'OpenAI publishes no cached-input price for this model, so caching is not priced.';
259
272
  const OPENAI_DEPRECATED_SUFFIX = ' Deprecated by OpenAI: priced for the logs that hold it, never recommended.';
273
+ const OPENAI_SHUT_DOWN_SUFFIX = ' Shut down by OpenAI: priced for the logs that hold it, never recommended.';
274
+ const OPENAI_CODEX_NOTE = 'Caching is automatic; a write costs what input costs. Responses API only, and no batch. OpenAI says the minimum for'
275
+ + ' models before GPT-5.6 varies by request settings, so 1,024 here is the nearest published floor.';
260
276
  const OPENAI_GPT5_NOTE = 'Caching is automatic, with no cache_control to set. OpenAI says the minimum for models before GPT-5.6 varies by'
261
277
  + ' request settings, so 1,024 here is the nearest published floor. The 2025-08-07 snapshot shuts down on'
262
278
  + ' 2026-12-11; OpenAI names the gpt-5.6 models as replacements.';
263
279
  const MISTRAL_CACHE_NOTE = 'Cached input costs 10% of the input price (mistral.ai/cloud/inference, read 2026-10-11). How caching starts and'
264
280
  + ' its minimum are not published anywhere Trazum can read, so no caching advice is given for it.';
281
+ /**
282
+ * US-only inference on Claude 4.6 and later.
283
+ *
284
+ * platform.claude.com/docs/en/about-claude/pricing, "Data residency pricing",
285
+ * read 2026-11-05: "For Claude 4.6 and later models, specifying US-only
286
+ * inference through the `inference_geo` parameter incurs a 1.1x multiplier on
287
+ * all token pricing categories, including input tokens, output tokens, cache
288
+ * writes, and cache reads." It stacks with fast mode and the Batch API
289
+ * discount, the same page says. Earlier models reject the parameter, so
290
+ * they carry no multiplier. `price-drift.mjs` reads the sentence weekly.
291
+ */
292
+ const US_ONLY_INFERENCE = Object.freeze({ us: 1.1 });
265
293
  export const MODELS = [
266
294
  {
267
295
  /*
@@ -273,6 +301,7 @@ export const MODELS = [
273
301
  */
274
302
  id: 'claude-fable-5-1',
275
303
  provider: 'anthropic',
304
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
276
305
  displayName: 'Claude Fable 5.1',
277
306
  inputPerMTok: 10,
278
307
  outputPerMTok: 50,
@@ -291,6 +320,7 @@ export const MODELS = [
291
320
  */
292
321
  id: 'claude-mythos-5-1',
293
322
  provider: 'anthropic',
323
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
294
324
  displayName: 'Claude Mythos 5.1',
295
325
  recommendable: false,
296
326
  inputPerMTok: 10,
@@ -306,6 +336,7 @@ export const MODELS = [
306
336
  {
307
337
  id: 'claude-fable-5',
308
338
  provider: 'anthropic',
339
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
309
340
  displayName: 'Claude Fable 5',
310
341
  inputPerMTok: 10,
311
342
  outputPerMTok: 50,
@@ -318,6 +349,7 @@ export const MODELS = [
318
349
  {
319
350
  id: 'claude-mythos-5',
320
351
  provider: 'anthropic',
352
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
321
353
  displayName: 'Claude Mythos 5',
322
354
  recommendable: false,
323
355
  inputPerMTok: 10,
@@ -338,6 +370,7 @@ export const MODELS = [
338
370
  */
339
371
  id: 'claude-opus-5-5',
340
372
  provider: 'anthropic',
373
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
341
374
  displayName: 'Claude Opus 5.5',
342
375
  inputPerMTok: 4,
343
376
  outputPerMTok: 20,
@@ -360,6 +393,7 @@ export const MODELS = [
360
393
  {
361
394
  id: 'claude-opus-5',
362
395
  provider: 'anthropic',
396
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
363
397
  displayName: 'Claude Opus 5',
364
398
  inputPerMTok: 5,
365
399
  outputPerMTok: 25,
@@ -374,6 +408,7 @@ export const MODELS = [
374
408
  {
375
409
  id: 'claude-opus-4-8',
376
410
  provider: 'anthropic',
411
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
377
412
  displayName: 'Claude Opus 4.8',
378
413
  inputPerMTok: 5,
379
414
  outputPerMTok: 25,
@@ -392,6 +427,7 @@ export const MODELS = [
392
427
  {
393
428
  id: 'claude-opus-4-7',
394
429
  provider: 'anthropic',
430
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
395
431
  displayName: 'Claude Opus 4.7',
396
432
  inputPerMTok: 5,
397
433
  outputPerMTok: 25,
@@ -403,6 +439,7 @@ export const MODELS = [
403
439
  {
404
440
  id: 'claude-opus-4-6',
405
441
  provider: 'anthropic',
442
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
406
443
  displayName: 'Claude Opus 4.6',
407
444
  inputPerMTok: 5,
408
445
  outputPerMTok: 25,
@@ -432,6 +469,51 @@ export const MODELS = [
432
469
  tier: 'opus',
433
470
  notes: '200K context window and a 64K output cap.',
434
471
  },
472
+ {
473
+ /*
474
+ * Retired on the Claude API on 2026-08-05 (the deprecations page, read
475
+ * 2026-11-05) and still served on Claude Models on Google Cloud (Google
476
+ * managed) and on Bedrock. Priced because a year of logs carries it and,
477
+ * unpriced, every one of those calls read as a gap. The pricing page, read
478
+ * the same day: $15/$75, writes $18.75 and $30, cache hits $1.50, batch
479
+ * $7.50/$37.50, which are the default multipliers. Its model page is gone;
480
+ * the context-windows page gives every Claude model outside its 1M list a
481
+ * 200k window, and the prompt-caching page a 1,024-token cache minimum.
482
+ * The dated id is the one the deprecations page names and what a log
483
+ * carries; no shorter alias is on any page read, so none is declared.
484
+ */
485
+ id: 'claude-opus-4-1-20250805',
486
+ provider: 'anthropic',
487
+ displayName: 'Claude Opus 4.1',
488
+ recommendable: false,
489
+ inputPerMTok: 15,
490
+ outputPerMTok: 75,
491
+ contextWindow: 200_000,
492
+ cacheMinTokens: 1024,
493
+ capability: 'large',
494
+ tier: 'opus',
495
+ notes: 'Retired on the Claude API on 2026-08-05; still served on Google Cloud (Google managed) and Bedrock.'
496
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
497
+ },
498
+ {
499
+ /*
500
+ * Retired on the Claude API on 2026-06-15; still served on Claude Models
501
+ * on Google Cloud (Google managed). The same pages and the same figures as
502
+ * Opus 4.1 above: $15/$75, 200k context, a 1,024-token cache minimum.
503
+ */
504
+ id: 'claude-opus-4-20250514',
505
+ provider: 'anthropic',
506
+ displayName: 'Claude Opus 4',
507
+ recommendable: false,
508
+ inputPerMTok: 15,
509
+ outputPerMTok: 75,
510
+ contextWindow: 200_000,
511
+ cacheMinTokens: 1024,
512
+ capability: 'large',
513
+ tier: 'opus',
514
+ notes: 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed).'
515
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
516
+ },
435
517
  {
436
518
  /*
437
519
  * The pricing page and the Sonnet 5.5 model page, read 2026-10-10. The
@@ -440,6 +522,7 @@ export const MODELS = [
440
522
  */
441
523
  id: 'claude-sonnet-5-5',
442
524
  provider: 'anthropic',
525
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
443
526
  displayName: 'Claude Sonnet 5.5',
444
527
  inputPerMTok: 2,
445
528
  outputPerMTok: 10,
@@ -453,6 +536,7 @@ export const MODELS = [
453
536
  {
454
537
  id: 'claude-sonnet-5',
455
538
  provider: 'anthropic',
539
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
456
540
  displayName: 'Claude Sonnet 5',
457
541
  inputPerMTok: 2,
458
542
  outputPerMTok: 10,
@@ -466,6 +550,7 @@ export const MODELS = [
466
550
  {
467
551
  id: 'claude-sonnet-4-6',
468
552
  provider: 'anthropic',
553
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
469
554
  displayName: 'Claude Sonnet 4.6',
470
555
  inputPerMTok: 3,
471
556
  outputPerMTok: 15,
@@ -497,6 +582,29 @@ export const MODELS = [
497
582
  tier: 'sonnet',
498
583
  notes: 'Deprecated; retires 2026-11-30. Recommended replacement: Claude Sonnet 5.5.',
499
584
  },
585
+ {
586
+ /*
587
+ * Retired on the Claude API on 2026-06-15 (the deprecations page, read
588
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
589
+ * managed) and on Bedrock. The pricing page, read the same day: $3/$15,
590
+ * writes $3.75 and $6, cache hits $0.30, batch $1.50/$7.50. 200k context
591
+ * from the context-windows page, a 1,024-token cache minimum from the
592
+ * prompt-caching page. The pricing page publishes one rate for it and no
593
+ * long-context band, so every call is priced at that rate.
594
+ */
595
+ id: 'claude-sonnet-4-20250514',
596
+ provider: 'anthropic',
597
+ displayName: 'Claude Sonnet 4',
598
+ recommendable: false,
599
+ inputPerMTok: 3,
600
+ outputPerMTok: 15,
601
+ contextWindow: 200_000,
602
+ cacheMinTokens: 1024,
603
+ capability: 'mid',
604
+ tier: 'sonnet',
605
+ notes: 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed) and Bedrock.'
606
+ + ' Anthropic names Claude Sonnet 5.5 as the replacement.',
607
+ },
500
608
  {
501
609
  /*
502
610
  * The pricing page, the models overview and the prompt-caching page, read
@@ -516,6 +624,7 @@ export const MODELS = [
516
624
  */
517
625
  id: 'claude-haiku-5-5',
518
626
  provider: 'anthropic',
627
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
519
628
  displayName: 'Claude Haiku 5.5',
520
629
  inputPerMTok: 0.1,
521
630
  outputPerMTok: 0.5,
@@ -551,6 +660,28 @@ export const MODELS = [
551
660
  tier: 'haiku',
552
661
  notes: '200K context window and a 64K output cap, smaller than the rest.',
553
662
  },
663
+ {
664
+ /*
665
+ * Retired on the Claude API on 2026-02-19 (the deprecations page, read
666
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
667
+ * managed). The pricing page, read the same day: $0.80/$4, writes $1 and
668
+ * $1.60, cache hits $0.08, batch $0.40/$2, the default multipliers. 200k
669
+ * context from the context-windows page, and the prompt-caching page's
670
+ * 2,048-token cache minimum.
671
+ */
672
+ id: 'claude-3-5-haiku-20241022',
673
+ provider: 'anthropic',
674
+ displayName: 'Claude Haiku 3.5',
675
+ recommendable: false,
676
+ inputPerMTok: 0.8,
677
+ outputPerMTok: 4,
678
+ contextWindow: 200_000,
679
+ cacheMinTokens: 2048,
680
+ capability: 'small',
681
+ tier: 'haiku',
682
+ notes: 'Retired on the Claude API on 2026-02-19; still served on Google Cloud (Google managed).'
683
+ + ' Anthropic names Claude Haiku 4.5 as the replacement.',
684
+ },
554
685
  // ------------------------------------------------------------------------
555
686
  // OpenAI
556
687
  //
@@ -610,6 +741,11 @@ export const MODELS = [
610
741
  outputPerMTok: 75,
611
742
  },
612
743
  ],
744
+ serviceTiers: {
745
+ flex: { inputPerMTok: 5, outputPerMTok: 25, tiers: openAiLongBand(10, 37.5) },
746
+ fast: { inputPerMTok: 20, outputPerMTok: 100, tiers: openAiLongBand(40, 150) },
747
+ ultrafast: { inputPerMTok: 60, outputPerMTok: 300, tiers: openAiLongBand(120, 450) },
748
+ },
613
749
  capability: 'frontier',
614
750
  tier: 'frontier',
615
751
  notes: OPENAI_CURRENT_NOTE,
@@ -633,6 +769,11 @@ export const MODELS = [
633
769
  outputPerMTok: 15,
634
770
  },
635
771
  ],
772
+ serviceTiers: {
773
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
774
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
775
+ ultrafast: { inputPerMTok: 12, outputPerMTok: 60, tiers: openAiLongBand(24, 90) },
776
+ },
636
777
  capability: 'large',
637
778
  tier: 'opus',
638
779
  notes: OPENAI_CURRENT_NOTE,
@@ -662,6 +803,10 @@ export const MODELS = [
662
803
  outputPerMTok: 7.5,
663
804
  },
664
805
  ],
806
+ serviceTiers: {
807
+ flex: { inputPerMTok: 0.5, outputPerMTok: 2.5, tiers: openAiLongBand(1, 3.75) },
808
+ fast: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
809
+ },
665
810
  capability: 'mid',
666
811
  tier: 'sonnet',
667
812
  notes: OPENAI_CURRENT_NOTE,
@@ -685,6 +830,10 @@ export const MODELS = [
685
830
  outputPerMTok: 15,
686
831
  },
687
832
  ],
833
+ serviceTiers: {
834
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
835
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
836
+ },
688
837
  capability: 'large',
689
838
  tier: 'opus',
690
839
  notes: OPENAI_CURRENT_NOTE,
@@ -708,12 +857,30 @@ export const MODELS = [
708
857
  outputPerMTok: 0.75,
709
858
  },
710
859
  ],
860
+ serviceTiers: {
861
+ flex: { inputPerMTok: 0.05, outputPerMTok: 0.25, tiers: openAiLongBand(0.1, 0.375) },
862
+ fast: { inputPerMTok: 0.2, outputPerMTok: 1, tiers: openAiLongBand(0.4, 1.5) },
863
+ },
711
864
  capability: 'small',
712
865
  tier: 'haiku',
713
866
  notes: OPENAI_CURRENT_NOTE,
714
867
  },
715
868
  {
869
+ /*
870
+ * Its model page, read 2026-11-05: "The `gpt-5.6` alias routes requests
871
+ * to GPT-5.6 Sol", so a log that names the alias is this model; and
872
+ * "$4 per million input tokens and $20 per million output tokens, a 20%
873
+ * reduction in input pricing and a 33% reduction in output pricing ...
874
+ * promotional pricing is available until November 30, 2026". The price
875
+ * after the promotion is not published as a figure, and deriving one
876
+ * from two rounded percentages would be a price nobody quoted, so the row
877
+ * carries what the page charges today. When the page changes,
878
+ * `scripts/price-drift.mjs` reports it on the next weekly run; until a
879
+ * person reads the new figure, a call after 2026-11-30 may be priced low,
880
+ * and the note says so to anyone reading the row.
881
+ */
716
882
  id: 'gpt-5.6-sol',
883
+ aliases: ['gpt-5.6'],
717
884
  provider: 'openai',
718
885
  displayName: 'GPT-5.6 Sol',
719
886
  inputPerMTok: 4,
@@ -731,9 +898,14 @@ export const MODELS = [
731
898
  outputPerMTok: 30,
732
899
  },
733
900
  ],
901
+ serviceTiers: {
902
+ flex: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
903
+ fast: { inputPerMTok: 8, outputPerMTok: 40, tiers: openAiLongBand(16, 60) },
904
+ },
734
905
  capability: 'large',
735
906
  tier: 'opus',
736
- notes: OPENAI_CURRENT_NOTE,
907
+ notes: OPENAI_CURRENT_NOTE
908
+ + ' $4/$20 is promotional until 2026-11-30, and OpenAI has not published the price after it as a figure.',
737
909
  },
738
910
  {
739
911
  id: 'gpt-5.6-terra',
@@ -754,6 +926,10 @@ export const MODELS = [
754
926
  outputPerMTok: 18,
755
927
  },
756
928
  ],
929
+ serviceTiers: {
930
+ flex: { inputPerMTok: 1, outputPerMTok: 6, tiers: openAiLongBand(2, 9) },
931
+ fast: { inputPerMTok: 4, outputPerMTok: 24, tiers: openAiLongBand(8, 36) },
932
+ },
757
933
  capability: 'mid',
758
934
  tier: 'sonnet',
759
935
  notes: OPENAI_CURRENT_NOTE,
@@ -777,55 +953,109 @@ export const MODELS = [
777
953
  outputPerMTok: 1.8,
778
954
  },
779
955
  ],
956
+ serviceTiers: {
957
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.6, tiers: openAiLongBand(0.2, 0.9) },
958
+ fast: { inputPerMTok: 0.4, outputPerMTok: 2.4, tiers: openAiLongBand(0.8, 3.6) },
959
+ },
780
960
  capability: 'small',
781
961
  tier: 'haiku',
782
962
  notes: OPENAI_CURRENT_NOTE,
783
963
  },
964
+ {
965
+ /*
966
+ * The pricing page's "Cyber models" table and the model's own page
967
+ * (developers.openai.com/api/docs/models/gpt-5.6-cyber.md), read
968
+ * 2026-11-05: $12.50/$75, cached input $1.25, the 30-minute write $15.625
969
+ * and the 12-hour one $31.25, the same 0.1x, 1.25x and 2.5x as the rest of
970
+ * GPT-5.6. A 400,000-token context window with at most 272,000 input
971
+ * tokens, so the long-prompt band the model page mentions can never be
972
+ * reached and none is modelled; the pricing page lists none either. The
973
+ * Batch endpoint is "Not supported", so there is no batch price.
974
+ *
975
+ * Found by `scripts/price-drift.mjs` on its first run, as a model on the
976
+ * pricing page that read as unpriced. It needs Daybreak programme
977
+ * approval, so it is priced for the logs of people who have it and never
978
+ * recommended to anyone else.
979
+ */
980
+ id: 'gpt-5.6-cyber',
981
+ provider: 'openai',
982
+ displayName: 'GPT-5.6 Cyber',
983
+ recommendable: false,
984
+ inputPerMTok: 12.5,
985
+ outputPerMTok: 75,
986
+ contextWindow: 400_000,
987
+ cacheMinTokens: 1024,
988
+ caching: 'automatic',
989
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1.25, cacheWrite1h: 2.5, batch: null },
990
+ cacheTtl: OPENAI_CACHE_TTL,
991
+ capability: 'large',
992
+ tier: 'opus',
993
+ notes: 'Needs approval through the Daybreak programme. At most 272,000 input tokens, Responses API only, no batch.'
994
+ + ' Cache writes cost 1.25x input for the default 30-minute entry and 2.5x for a 12-hour one.',
995
+ },
784
996
  // ------------------------------------------------------------------------
997
+ // GPT-5, mini and nano. Each model page, read 2026-11-05, opens with
998
+ // "Status: Deprecated", and the catalogue still offered all three as
999
+ // cheaper alternatives, nano as the cheapest OpenAI model there is. Priced
1000
+ // for the logs that hold them; never recommended.
785
1001
  {
786
1002
  id: 'gpt-5',
787
1003
  aliases: ['gpt-5-2025-08-07'],
788
1004
  provider: 'openai',
789
1005
  displayName: 'GPT-5',
1006
+ recommendable: false,
790
1007
  inputPerMTok: 1.25,
791
1008
  outputPerMTok: 10,
792
1009
  contextWindow: 400_000,
793
1010
  cacheMinTokens: 1024,
794
1011
  caching: 'automatic',
795
1012
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1013
+ serviceTiers: {
1014
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1015
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1016
+ },
796
1017
  capability: 'frontier',
797
1018
  tier: 'frontier',
798
- notes: OPENAI_GPT5_NOTE,
1019
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
799
1020
  },
800
1021
  {
801
1022
  id: 'gpt-5-mini',
802
1023
  aliases: ['gpt-5-mini-2025-08-07'],
803
1024
  provider: 'openai',
804
1025
  displayName: 'GPT-5 mini',
1026
+ recommendable: false,
805
1027
  inputPerMTok: 0.25,
806
1028
  outputPerMTok: 2,
807
1029
  contextWindow: 400_000,
808
1030
  cacheMinTokens: 1024,
809
1031
  caching: 'automatic',
810
1032
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1033
+ serviceTiers: {
1034
+ flex: { inputPerMTok: 0.125, outputPerMTok: 1 },
1035
+ fast: { inputPerMTok: 0.45, outputPerMTok: 3.6 },
1036
+ },
811
1037
  capability: 'mid',
812
1038
  tier: 'sonnet',
813
- notes: OPENAI_GPT5_NOTE,
1039
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
814
1040
  },
815
1041
  {
816
1042
  id: 'gpt-5-nano',
817
1043
  aliases: ['gpt-5-nano-2025-08-07'],
818
1044
  provider: 'openai',
819
1045
  displayName: 'GPT-5 nano',
1046
+ recommendable: false,
820
1047
  inputPerMTok: 0.05,
821
1048
  outputPerMTok: 0.4,
822
1049
  contextWindow: 400_000,
823
1050
  cacheMinTokens: 1024,
824
1051
  caching: 'automatic',
825
1052
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1053
+ serviceTiers: {
1054
+ flex: { inputPerMTok: 0.025, outputPerMTok: 0.2 },
1055
+ },
826
1056
  capability: 'small',
827
1057
  tier: 'haiku',
828
- notes: OPENAI_GPT5_NOTE,
1058
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
829
1059
  },
830
1060
  // ------------------------------------------------------------------------
831
1061
  // OpenAI's earlier models, still billed and still in logs.
@@ -862,6 +1092,10 @@ export const MODELS = [
862
1092
  outputPerMTok: 45,
863
1093
  },
864
1094
  ],
1095
+ serviceTiers: {
1096
+ flex: { inputPerMTok: 2.5, outputPerMTok: 15, tiers: openAiLongBand(5, 22.5) },
1097
+ fast: { inputPerMTok: 12.5, outputPerMTok: 75 },
1098
+ },
865
1099
  capability: 'frontier',
866
1100
  tier: 'frontier',
867
1101
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -885,6 +1119,9 @@ export const MODELS = [
885
1119
  outputPerMTok: 270,
886
1120
  },
887
1121
  ],
1122
+ serviceTiers: {
1123
+ flex: { inputPerMTok: 15, outputPerMTok: 90 },
1124
+ },
888
1125
  capability: 'frontier',
889
1126
  tier: 'frontier',
890
1127
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -908,6 +1145,10 @@ export const MODELS = [
908
1145
  outputPerMTok: 22.5,
909
1146
  },
910
1147
  ],
1148
+ serviceTiers: {
1149
+ flex: { inputPerMTok: 1.25, outputPerMTok: 7.5, tiers: openAiLongBand(2.5, 11.25) },
1150
+ fast: { inputPerMTok: 5, outputPerMTok: 30 },
1151
+ },
911
1152
  capability: 'frontier',
912
1153
  tier: 'frontier',
913
1154
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -931,6 +1172,9 @@ export const MODELS = [
931
1172
  outputPerMTok: 270,
932
1173
  },
933
1174
  ],
1175
+ serviceTiers: {
1176
+ flex: { inputPerMTok: 15, outputPerMTok: 90, tiers: openAiLongBand(30, 135) },
1177
+ },
934
1178
  capability: 'frontier',
935
1179
  tier: 'frontier',
936
1180
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -946,6 +1190,10 @@ export const MODELS = [
946
1190
  cacheMinTokens: 1024,
947
1191
  caching: 'automatic',
948
1192
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1193
+ serviceTiers: {
1194
+ flex: { inputPerMTok: 0.375, outputPerMTok: 2.25 },
1195
+ fast: { inputPerMTok: 1.5, outputPerMTok: 9 },
1196
+ },
949
1197
  capability: 'mid',
950
1198
  tier: 'sonnet',
951
1199
  notes: OPENAI_EARLIER_NOTE,
@@ -962,6 +1210,9 @@ export const MODELS = [
962
1210
  cacheMinTokens: 1024,
963
1211
  caching: 'automatic',
964
1212
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1213
+ serviceTiers: {
1214
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.625 },
1215
+ },
965
1216
  capability: 'small',
966
1217
  tier: 'haiku',
967
1218
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -977,6 +1228,10 @@ export const MODELS = [
977
1228
  cacheMinTokens: 1024,
978
1229
  caching: 'automatic',
979
1230
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1231
+ serviceTiers: {
1232
+ flex: { inputPerMTok: 0.875, outputPerMTok: 7 },
1233
+ fast: { inputPerMTok: 3.5, outputPerMTok: 28 },
1234
+ },
980
1235
  capability: 'frontier',
981
1236
  tier: 'frontier',
982
1237
  notes: OPENAI_EARLIER_NOTE,
@@ -1008,6 +1263,10 @@ export const MODELS = [
1008
1263
  cacheMinTokens: 1024,
1009
1264
  caching: 'automatic',
1010
1265
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1266
+ serviceTiers: {
1267
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1268
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1269
+ },
1011
1270
  capability: 'frontier',
1012
1271
  tier: 'frontier',
1013
1272
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1040,6 +1299,9 @@ export const MODELS = [
1040
1299
  cacheMinTokens: 1024,
1041
1300
  caching: 'automatic',
1042
1301
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1302
+ serviceTiers: {
1303
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1304
+ },
1043
1305
  capability: 'large',
1044
1306
  tier: 'opus',
1045
1307
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1056,6 +1318,9 @@ export const MODELS = [
1056
1318
  cacheMinTokens: 1024,
1057
1319
  caching: 'automatic',
1058
1320
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1321
+ serviceTiers: {
1322
+ fast: { inputPerMTok: 0.7, outputPerMTok: 2.8 },
1323
+ },
1059
1324
  capability: 'mid',
1060
1325
  tier: 'sonnet',
1061
1326
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1072,6 +1337,9 @@ export const MODELS = [
1072
1337
  cacheMinTokens: 1024,
1073
1338
  caching: 'automatic',
1074
1339
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1340
+ serviceTiers: {
1341
+ fast: { inputPerMTok: 4.25, outputPerMTok: 17 },
1342
+ },
1075
1343
  capability: 'large',
1076
1344
  tier: 'opus',
1077
1345
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1088,6 +1356,9 @@ export const MODELS = [
1088
1356
  cacheMinTokens: 1024,
1089
1357
  caching: 'automatic',
1090
1358
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1359
+ serviceTiers: {
1360
+ fast: { inputPerMTok: 0.25, outputPerMTok: 1 },
1361
+ },
1091
1362
  capability: 'small',
1092
1363
  tier: 'haiku',
1093
1364
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1104,6 +1375,10 @@ export const MODELS = [
1104
1375
  cacheMinTokens: 1024,
1105
1376
  caching: 'automatic',
1106
1377
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1378
+ serviceTiers: {
1379
+ flex: { inputPerMTok: 1, outputPerMTok: 4 },
1380
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1381
+ },
1107
1382
  capability: 'frontier',
1108
1383
  tier: 'frontier',
1109
1384
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1125,6 +1400,143 @@ export const MODELS = [
1125
1400
  notes: OPENAI_EARLIER_NO_CACHE_NOTE + OPENAI_DEPRECATED_SUFFIX,
1126
1401
  },
1127
1402
  // ------------------------------------------------------------------------
1403
+ // The Codex models and chat-latest. Codex CLI sessions name these, and none
1404
+ // was priced, so every such call read as unpriced. Each from its own model
1405
+ // page (developers.openai.com/api/docs/models/<id>.md), read 2026-11-05,
1406
+ // and gpt-5.3-codex and chat-latest also from the pricing page's
1407
+ // "Specialized models" table. Every page says Batch is "Not supported", so
1408
+ // none has a batch price. gpt-5.3-codex is deprecated and the rest of the
1409
+ // Codex line is shut down; all are priced for the logs that hold them and
1410
+ // never recommended. chat-latest is served and not recommended either: it
1411
+ // points at whatever model ChatGPT uses, so a recommendation would not name
1412
+ // a fixed model. Cache reads are 0.1x everywhere but codex-mini-latest,
1413
+ // whose page gives $0.375 on $1.50, 0.25x. gpt-5-search-api is on the
1414
+ // pricing page too and its model page returns 404, so it has no context
1415
+ // window and is left unpriced (see scripts/price-drift.mjs).
1416
+ // ------------------------------------------------------------------------
1417
+ {
1418
+ id: 'gpt-5.3-codex',
1419
+ provider: 'openai',
1420
+ displayName: 'GPT-5.3-Codex',
1421
+ recommendable: false,
1422
+ inputPerMTok: 1.75,
1423
+ outputPerMTok: 14,
1424
+ contextWindow: 400_000,
1425
+ cacheMinTokens: 1024,
1426
+ caching: 'automatic',
1427
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1428
+ // The pricing page's specialized "Fast" table: $3.50/$28, cached $0.35.
1429
+ serviceTiers: { fast: { inputPerMTok: 3.5, outputPerMTok: 28 } },
1430
+ capability: 'frontier',
1431
+ tier: 'frontier',
1432
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_DEPRECATED_SUFFIX,
1433
+ },
1434
+ {
1435
+ id: 'gpt-5.2-codex',
1436
+ provider: 'openai',
1437
+ displayName: 'GPT-5.2-Codex',
1438
+ recommendable: false,
1439
+ inputPerMTok: 1.75,
1440
+ outputPerMTok: 14,
1441
+ contextWindow: 400_000,
1442
+ cacheMinTokens: 1024,
1443
+ caching: 'automatic',
1444
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1445
+ capability: 'frontier',
1446
+ tier: 'frontier',
1447
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1448
+ },
1449
+ {
1450
+ id: 'gpt-5.1-codex',
1451
+ provider: 'openai',
1452
+ displayName: 'GPT-5.1-Codex',
1453
+ recommendable: false,
1454
+ inputPerMTok: 1.25,
1455
+ outputPerMTok: 10,
1456
+ contextWindow: 400_000,
1457
+ cacheMinTokens: 1024,
1458
+ caching: 'automatic',
1459
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1460
+ capability: 'frontier',
1461
+ tier: 'frontier',
1462
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1463
+ },
1464
+ {
1465
+ id: 'gpt-5.1-codex-max',
1466
+ provider: 'openai',
1467
+ displayName: 'GPT-5.1-Codex-Max',
1468
+ recommendable: false,
1469
+ inputPerMTok: 1.25,
1470
+ outputPerMTok: 10,
1471
+ contextWindow: 400_000,
1472
+ cacheMinTokens: 1024,
1473
+ caching: 'automatic',
1474
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1475
+ capability: 'frontier',
1476
+ tier: 'frontier',
1477
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1478
+ },
1479
+ {
1480
+ id: 'gpt-5.1-codex-mini',
1481
+ provider: 'openai',
1482
+ displayName: 'GPT-5.1-Codex Mini',
1483
+ recommendable: false,
1484
+ inputPerMTok: 0.25,
1485
+ outputPerMTok: 2,
1486
+ contextWindow: 400_000,
1487
+ cacheMinTokens: 1024,
1488
+ caching: 'automatic',
1489
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1490
+ capability: 'mid',
1491
+ tier: 'sonnet',
1492
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1493
+ },
1494
+ {
1495
+ id: 'gpt-5-codex',
1496
+ provider: 'openai',
1497
+ displayName: 'GPT-5-Codex',
1498
+ recommendable: false,
1499
+ inputPerMTok: 1.25,
1500
+ outputPerMTok: 10,
1501
+ contextWindow: 400_000,
1502
+ cacheMinTokens: 1024,
1503
+ caching: 'automatic',
1504
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1505
+ capability: 'frontier',
1506
+ tier: 'frontier',
1507
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1508
+ },
1509
+ {
1510
+ id: 'codex-mini-latest',
1511
+ provider: 'openai',
1512
+ displayName: 'codex-mini-latest',
1513
+ recommendable: false,
1514
+ inputPerMTok: 1.5,
1515
+ outputPerMTok: 6,
1516
+ contextWindow: 200_000,
1517
+ cacheMinTokens: 1024,
1518
+ caching: 'automatic',
1519
+ multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1520
+ capability: 'mid',
1521
+ tier: 'sonnet',
1522
+ notes: OPENAI_CODEX_NOTE + ' A cache read costs 0.25x input.' + OPENAI_SHUT_DOWN_SUFFIX,
1523
+ },
1524
+ {
1525
+ id: 'chat-latest',
1526
+ provider: 'openai',
1527
+ displayName: 'Chat Latest',
1528
+ recommendable: false,
1529
+ inputPerMTok: 5,
1530
+ outputPerMTok: 30,
1531
+ contextWindow: 400_000,
1532
+ cacheMinTokens: 1024,
1533
+ caching: 'automatic',
1534
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1535
+ capability: 'large',
1536
+ tier: 'opus',
1537
+ notes: 'Points at the model ChatGPT currently uses, so it is priced and never recommended. Caching is automatic; a write costs what input costs. At most 272,000 input tokens, and no batch.',
1538
+ },
1539
+ // ------------------------------------------------------------------------
1128
1540
  // Google
1129
1541
  // ------------------------------------------------------------------------
1130
1542
  {
@@ -1905,6 +2317,70 @@ export function getModel(id) {
1905
2317
  export function listModels() {
1906
2318
  return [...MODELS];
1907
2319
  }
2320
+ /**
2321
+ * A response's `usage.inference_geo` as the geography that prices it, or
2322
+ * `undefined` for the global rate. `"global"` is the default and the row's
2323
+ * own price; `"not_available"` is what a model that cannot pin a geography
2324
+ * reports, and is priced the same.
2325
+ */
2326
+ export function inferenceGeoOf(value) {
2327
+ if (typeof value !== 'string')
2328
+ return undefined;
2329
+ return value.trim().toLowerCase() === 'us' ? 'us' : undefined;
2330
+ }
2331
+ /** Rounded to a billionth of a dollar per million tokens, so 1.1x of $5 is $5.5 and not $5.500000000000001. */
2332
+ const scaled = (rate, factor) => Math.round(rate * factor * 1e9) / 1e9;
2333
+ /**
2334
+ * A response's `service_tier` as the tier whose table prices it, or
2335
+ * `undefined` for the standard one.
2336
+ *
2337
+ * OpenAI renamed Priority processing to Fast mode on 2026-07-30, and its guide
2338
+ * says "for GPT-5.6 and earlier models, the response returns `priority`
2339
+ * whether the request specifies `priority` or `fast`", so both spellings are
2340
+ * the Fast table. `"default"` is the standard tier, and is also what a Fast
2341
+ * request downgraded under the ramp-rate limit returns, "and charge standard
2342
+ * rates". `"auto"` and `"scale"`, or a value this table has never seen, are
2343
+ * priced at the standard rate rather than guessed into another tier.
2344
+ */
2345
+ export function serviceTierOf(value) {
2346
+ if (typeof value !== 'string')
2347
+ return undefined;
2348
+ const tier = value.trim().toLowerCase();
2349
+ if (tier === 'flex')
2350
+ return 'flex';
2351
+ if (tier === 'fast' || tier === 'priority')
2352
+ return 'fast';
2353
+ if (tier === 'ultrafast')
2354
+ return 'ultrafast';
2355
+ return undefined;
2356
+ }
2357
+ /**
2358
+ * The row as one service tier prices it: that tier's rates and long band, the
2359
+ * row's own cache and batch ratios. Built once per row and tier.
2360
+ */
2361
+ const serviceTierMemo = new WeakMap();
2362
+ function serviceTierRow(model, tier) {
2363
+ const sheet = model.serviceTiers?.[tier];
2364
+ if (sheet === undefined)
2365
+ return undefined;
2366
+ let rows = serviceTierMemo.get(model);
2367
+ if (rows === undefined) {
2368
+ rows = new Map();
2369
+ serviceTierMemo.set(model, rows);
2370
+ }
2371
+ let row = rows.get(tier);
2372
+ if (row === undefined) {
2373
+ const { history: _history, promo: _promo, tiers: _tiers, serviceTiers: _serviceTiers, ...rest } = model;
2374
+ row = Object.freeze({
2375
+ ...rest,
2376
+ inputPerMTok: sheet.inputPerMTok,
2377
+ outputPerMTok: sheet.outputPerMTok,
2378
+ ...(sheet.tiers !== undefined ? { tiers: sheet.tiers } : {}),
2379
+ });
2380
+ rows.set(tier, row);
2381
+ }
2382
+ return row;
2383
+ }
1908
2384
  const pricingMemo = new WeakMap();
1909
2385
  /**
1910
2386
  * The row as it stood at `on`: itself when the current price applies, or a
@@ -1919,7 +2395,8 @@ function pricedAt(model, on) {
1919
2395
  let periods = periodMemo.get(model);
1920
2396
  if (periods === undefined) {
1921
2397
  periods = history.map((period) => {
1922
- const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, ...rest } = model;
2398
+ // A period's own price only: the service tiers below are today's tables.
2399
+ const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, serviceTiers: _serviceTiers, ...rest } = model;
1923
2400
  const row = Object.freeze({
1924
2401
  ...rest,
1925
2402
  inputPerMTok: period.inputPerMTok,
@@ -2005,12 +2482,54 @@ function holds(when, on, size) {
2005
2482
  * `pricing.test.js` fails a tiered model whose report drops it.
2006
2483
  */
2007
2484
  export function effectivePricing(model, on = new Date(), context = {}) {
2485
+ /*
2486
+ * A geography with a published multiplier scales whatever the rest of the
2487
+ * call decides, fast mode and the long band included, because the page says
2488
+ * it stacks with all of them. Applied once, outside everything else.
2489
+ */
2490
+ if (context.inferenceGeo !== undefined) {
2491
+ const { inferenceGeo, ...rest } = context;
2492
+ const rates = effectivePricing(model, on, rest);
2493
+ const geo = inferenceGeoOf(inferenceGeo);
2494
+ const factor = geo === undefined ? undefined : model.inferenceGeoMultipliers?.[geo];
2495
+ if (geo === undefined || factor === undefined)
2496
+ return rates;
2497
+ return {
2498
+ ...rates,
2499
+ inputPerMTok: scaled(rates.inputPerMTok, factor),
2500
+ outputPerMTok: scaled(rates.outputPerMTok, factor),
2501
+ ...(rates.tier?.floor !== undefined
2502
+ ? {
2503
+ tier: {
2504
+ ...rates.tier,
2505
+ floor: {
2506
+ inputPerMTok: scaled(rates.tier.floor.inputPerMTok, factor),
2507
+ outputPerMTok: scaled(rates.tier.floor.outputPerMTok, factor),
2508
+ },
2509
+ },
2510
+ }
2511
+ : {}),
2512
+ inferenceGeo: geo,
2513
+ };
2514
+ }
2008
2515
  /* A call before a published price change is priced as it was billed then. */
2009
2516
  if (model.history !== undefined) {
2010
2517
  const period = pricedAt(model, on);
2011
2518
  if (period !== model)
2012
2519
  return effectivePricing(period, on, context);
2013
2520
  }
2521
+ /*
2522
+ * A response that names another service tier is priced by that tier's own
2523
+ * table, long band included, and by nothing composed out of two.
2524
+ */
2525
+ if (context.serviceTier !== undefined && model.serviceTiers !== undefined) {
2526
+ const named = serviceTierOf(context.serviceTier);
2527
+ const row = named === undefined ? undefined : serviceTierRow(model, named);
2528
+ if (row !== undefined) {
2529
+ const { serviceTier: _serviceTier, ...size } = context;
2530
+ return { ...effectivePricing(row, on, size), serviceTier: named };
2531
+ }
2532
+ }
2014
2533
  const memo = pricingMemo.get(model) ?? cache(model);
2015
2534
  /**
2016
2535
  * Tiers first, and a promotion never runs on top of one.