@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/src/pricing.ts CHANGED
@@ -1,4 +1,4 @@
1
- import type { ModelPricing, PricingTier, TierCondition } from './types.js';
1
+ import type { InferenceGeo, ModelPricing, PricingTier, ServiceTier, TierCondition } from './types.js';
2
2
 
3
3
  /**
4
4
  * Model and pricing catalogue (USD per million tokens).
@@ -41,10 +41,12 @@ export const PROVIDER_REVIEWED: Readonly<Record<string, string>> = Object.freeze
41
41
  * which bills Opus 5.5 at $8/$40 and Opus 5 and 4.8 at $10/$50 and which
42
42
  * every response records in `usage.speed`, so fast calls had been priced at
43
43
  * half their cost; and Claude Sonnet 4.5, deprecated but still served until
44
- * 2026-11-30, which read as unpriced. Both are here now. The retired Opus
45
- * 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are on the pricing page too and are
46
- * not priced here yet: a row needs a context window, which their own pages
47
- * would give, and those were not read.
44
+ * 2026-11-30, which read as unpriced. Both are here now. So are the
45
+ * retired Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5, still on the pricing
46
+ * page and in every log from before their retirement: their own model pages
47
+ * are gone, and the context-windows page gives each the 200k window every
48
+ * Claude model outside its 1M list has. Claude Sonnet 3.7 and Haiku 3 are
49
+ * retired and absent from the pricing page, so they stay unpriced.
48
50
  */
49
51
  anthropic: '2026-11-05',
50
52
  /*
@@ -269,6 +271,18 @@ const NO_PUBLISHED_CACHE_MINIMUM =
269
271
  */
270
272
  const OPENAI_CACHE_TTL = Object.freeze({ writeMs: 30 * 60_000, longerMs: 12 * 60 * 60_000 });
271
273
 
274
+ /**
275
+ * OpenAI's service-tier tables, read from developers.openai.com/api/docs/pricing.md
276
+ * on 2026-11-05: the Flex, Fast and Ultrafast tables beside the Standard one,
277
+ * each model's row as printed, and its "Long context" columns where the table
278
+ * publishes them ("Short context: <=272K input tokens. Long context: >272K").
279
+ * A row whose long columns read "-" gets no long band here: GPT-5.5 and 5.4 in
280
+ * Fast, GPT-5.5 pro in Flex. See `ModelPricing.serviceTiers`.
281
+ */
282
+ const openAiLongBand = (inputPerMTok: number, outputPerMTok: number): PricingTier[] => [
283
+ { id: 'over-272k', when: { kind: 'input-tokens-above', tokens: 272_000 }, inputPerMTok, outputPerMTok },
284
+ ];
285
+
272
286
  const OPENAI_CURRENT_NOTE =
273
287
  'Priced per request by prompt length: over 272,000 input tokens, 2x input and cache and 1.5x output for'
274
288
  + ' the whole request. Cache writes cost 1.25x input for the default 30-minute entry and 2.5x for a'
@@ -282,6 +296,10 @@ const OPENAI_EARLIER_LONG_NOTE =
282
296
  + OPENAI_EARLIER_NOTE;
283
297
  const OPENAI_EARLIER_NO_CACHE_NOTE = 'OpenAI publishes no cached-input price for this model, so caching is not priced.';
284
298
  const OPENAI_DEPRECATED_SUFFIX = ' Deprecated by OpenAI: priced for the logs that hold it, never recommended.';
299
+ const OPENAI_SHUT_DOWN_SUFFIX = ' Shut down by OpenAI: priced for the logs that hold it, never recommended.';
300
+ const OPENAI_CODEX_NOTE =
301
+ 'Caching is automatic; a write costs what input costs. Responses API only, and no batch. OpenAI says the minimum for'
302
+ + ' models before GPT-5.6 varies by request settings, so 1,024 here is the nearest published floor.';
285
303
 
286
304
  const OPENAI_GPT5_NOTE =
287
305
  'Caching is automatic, with no cache_control to set. OpenAI says the minimum for models before GPT-5.6 varies by'
@@ -292,6 +310,19 @@ const MISTRAL_CACHE_NOTE =
292
310
  'Cached input costs 10% of the input price (mistral.ai/cloud/inference, read 2026-10-11). How caching starts and'
293
311
  + ' its minimum are not published anywhere Trazum can read, so no caching advice is given for it.';
294
312
 
313
+ /**
314
+ * US-only inference on Claude 4.6 and later.
315
+ *
316
+ * platform.claude.com/docs/en/about-claude/pricing, "Data residency pricing",
317
+ * read 2026-11-05: "For Claude 4.6 and later models, specifying US-only
318
+ * inference through the `inference_geo` parameter incurs a 1.1x multiplier on
319
+ * all token pricing categories, including input tokens, output tokens, cache
320
+ * writes, and cache reads." It stacks with fast mode and the Batch API
321
+ * discount, the same page says. Earlier models reject the parameter, so
322
+ * they carry no multiplier. `price-drift.mjs` reads the sentence weekly.
323
+ */
324
+ const US_ONLY_INFERENCE = Object.freeze({ us: 1.1 });
325
+
295
326
  export const MODELS: ModelPricing[] = [
296
327
  {
297
328
  /*
@@ -303,6 +334,7 @@ export const MODELS: ModelPricing[] = [
303
334
  */
304
335
  id: 'claude-fable-5-1',
305
336
  provider: 'anthropic',
337
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
306
338
  displayName: 'Claude Fable 5.1',
307
339
  inputPerMTok: 10,
308
340
  outputPerMTok: 50,
@@ -321,6 +353,7 @@ export const MODELS: ModelPricing[] = [
321
353
  */
322
354
  id: 'claude-mythos-5-1',
323
355
  provider: 'anthropic',
356
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
324
357
  displayName: 'Claude Mythos 5.1',
325
358
  recommendable: false,
326
359
  inputPerMTok: 10,
@@ -337,6 +370,7 @@ export const MODELS: ModelPricing[] = [
337
370
  {
338
371
  id: 'claude-fable-5',
339
372
  provider: 'anthropic',
373
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
340
374
  displayName: 'Claude Fable 5',
341
375
  inputPerMTok: 10,
342
376
  outputPerMTok: 50,
@@ -349,6 +383,7 @@ export const MODELS: ModelPricing[] = [
349
383
  {
350
384
  id: 'claude-mythos-5',
351
385
  provider: 'anthropic',
386
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
352
387
  displayName: 'Claude Mythos 5',
353
388
  recommendable: false,
354
389
  inputPerMTok: 10,
@@ -370,6 +405,7 @@ export const MODELS: ModelPricing[] = [
370
405
  */
371
406
  id: 'claude-opus-5-5',
372
407
  provider: 'anthropic',
408
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
373
409
  displayName: 'Claude Opus 5.5',
374
410
  inputPerMTok: 4,
375
411
  outputPerMTok: 20,
@@ -392,6 +428,7 @@ export const MODELS: ModelPricing[] = [
392
428
  {
393
429
  id: 'claude-opus-5',
394
430
  provider: 'anthropic',
431
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
395
432
  displayName: 'Claude Opus 5',
396
433
  inputPerMTok: 5,
397
434
  outputPerMTok: 25,
@@ -406,6 +443,7 @@ export const MODELS: ModelPricing[] = [
406
443
  {
407
444
  id: 'claude-opus-4-8',
408
445
  provider: 'anthropic',
446
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
409
447
  displayName: 'Claude Opus 4.8',
410
448
  inputPerMTok: 5,
411
449
  outputPerMTok: 25,
@@ -424,6 +462,7 @@ export const MODELS: ModelPricing[] = [
424
462
  {
425
463
  id: 'claude-opus-4-7',
426
464
  provider: 'anthropic',
465
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
427
466
  displayName: 'Claude Opus 4.7',
428
467
  inputPerMTok: 5,
429
468
  outputPerMTok: 25,
@@ -435,6 +474,7 @@ export const MODELS: ModelPricing[] = [
435
474
  {
436
475
  id: 'claude-opus-4-6',
437
476
  provider: 'anthropic',
477
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
438
478
  displayName: 'Claude Opus 4.6',
439
479
  inputPerMTok: 5,
440
480
  outputPerMTok: 25,
@@ -464,6 +504,53 @@ export const MODELS: ModelPricing[] = [
464
504
  tier: 'opus',
465
505
  notes: '200K context window and a 64K output cap.',
466
506
  },
507
+ {
508
+ /*
509
+ * Retired on the Claude API on 2026-08-05 (the deprecations page, read
510
+ * 2026-11-05) and still served on Claude Models on Google Cloud (Google
511
+ * managed) and on Bedrock. Priced because a year of logs carries it and,
512
+ * unpriced, every one of those calls read as a gap. The pricing page, read
513
+ * the same day: $15/$75, writes $18.75 and $30, cache hits $1.50, batch
514
+ * $7.50/$37.50, which are the default multipliers. Its model page is gone;
515
+ * the context-windows page gives every Claude model outside its 1M list a
516
+ * 200k window, and the prompt-caching page a 1,024-token cache minimum.
517
+ * The dated id is the one the deprecations page names and what a log
518
+ * carries; no shorter alias is on any page read, so none is declared.
519
+ */
520
+ id: 'claude-opus-4-1-20250805',
521
+ provider: 'anthropic',
522
+ displayName: 'Claude Opus 4.1',
523
+ recommendable: false,
524
+ inputPerMTok: 15,
525
+ outputPerMTok: 75,
526
+ contextWindow: 200_000,
527
+ cacheMinTokens: 1024,
528
+ capability: 'large',
529
+ tier: 'opus',
530
+ notes:
531
+ 'Retired on the Claude API on 2026-08-05; still served on Google Cloud (Google managed) and Bedrock.'
532
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
533
+ },
534
+ {
535
+ /*
536
+ * Retired on the Claude API on 2026-06-15; still served on Claude Models
537
+ * on Google Cloud (Google managed). The same pages and the same figures as
538
+ * Opus 4.1 above: $15/$75, 200k context, a 1,024-token cache minimum.
539
+ */
540
+ id: 'claude-opus-4-20250514',
541
+ provider: 'anthropic',
542
+ displayName: 'Claude Opus 4',
543
+ recommendable: false,
544
+ inputPerMTok: 15,
545
+ outputPerMTok: 75,
546
+ contextWindow: 200_000,
547
+ cacheMinTokens: 1024,
548
+ capability: 'large',
549
+ tier: 'opus',
550
+ notes:
551
+ 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed).'
552
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
553
+ },
467
554
  {
468
555
  /*
469
556
  * The pricing page and the Sonnet 5.5 model page, read 2026-10-10. The
@@ -472,6 +559,7 @@ export const MODELS: ModelPricing[] = [
472
559
  */
473
560
  id: 'claude-sonnet-5-5',
474
561
  provider: 'anthropic',
562
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
475
563
  displayName: 'Claude Sonnet 5.5',
476
564
  inputPerMTok: 2,
477
565
  outputPerMTok: 10,
@@ -485,6 +573,7 @@ export const MODELS: ModelPricing[] = [
485
573
  {
486
574
  id: 'claude-sonnet-5',
487
575
  provider: 'anthropic',
576
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
488
577
  displayName: 'Claude Sonnet 5',
489
578
  inputPerMTok: 2,
490
579
  outputPerMTok: 10,
@@ -499,6 +588,7 @@ export const MODELS: ModelPricing[] = [
499
588
  {
500
589
  id: 'claude-sonnet-4-6',
501
590
  provider: 'anthropic',
591
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
502
592
  displayName: 'Claude Sonnet 4.6',
503
593
  inputPerMTok: 3,
504
594
  outputPerMTok: 15,
@@ -530,6 +620,30 @@ export const MODELS: ModelPricing[] = [
530
620
  tier: 'sonnet',
531
621
  notes: 'Deprecated; retires 2026-11-30. Recommended replacement: Claude Sonnet 5.5.',
532
622
  },
623
+ {
624
+ /*
625
+ * Retired on the Claude API on 2026-06-15 (the deprecations page, read
626
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
627
+ * managed) and on Bedrock. The pricing page, read the same day: $3/$15,
628
+ * writes $3.75 and $6, cache hits $0.30, batch $1.50/$7.50. 200k context
629
+ * from the context-windows page, a 1,024-token cache minimum from the
630
+ * prompt-caching page. The pricing page publishes one rate for it and no
631
+ * long-context band, so every call is priced at that rate.
632
+ */
633
+ id: 'claude-sonnet-4-20250514',
634
+ provider: 'anthropic',
635
+ displayName: 'Claude Sonnet 4',
636
+ recommendable: false,
637
+ inputPerMTok: 3,
638
+ outputPerMTok: 15,
639
+ contextWindow: 200_000,
640
+ cacheMinTokens: 1024,
641
+ capability: 'mid',
642
+ tier: 'sonnet',
643
+ notes:
644
+ 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed) and Bedrock.'
645
+ + ' Anthropic names Claude Sonnet 5.5 as the replacement.',
646
+ },
533
647
  {
534
648
  /*
535
649
  * The pricing page, the models overview and the prompt-caching page, read
@@ -549,6 +663,7 @@ export const MODELS: ModelPricing[] = [
549
663
  */
550
664
  id: 'claude-haiku-5-5',
551
665
  provider: 'anthropic',
666
+ inferenceGeoMultipliers: US_ONLY_INFERENCE,
552
667
  displayName: 'Claude Haiku 5.5',
553
668
  inputPerMTok: 0.1,
554
669
  outputPerMTok: 0.5,
@@ -585,6 +700,29 @@ export const MODELS: ModelPricing[] = [
585
700
  tier: 'haiku',
586
701
  notes: '200K context window and a 64K output cap, smaller than the rest.',
587
702
  },
703
+ {
704
+ /*
705
+ * Retired on the Claude API on 2026-02-19 (the deprecations page, read
706
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
707
+ * managed). The pricing page, read the same day: $0.80/$4, writes $1 and
708
+ * $1.60, cache hits $0.08, batch $0.40/$2, the default multipliers. 200k
709
+ * context from the context-windows page, and the prompt-caching page's
710
+ * 2,048-token cache minimum.
711
+ */
712
+ id: 'claude-3-5-haiku-20241022',
713
+ provider: 'anthropic',
714
+ displayName: 'Claude Haiku 3.5',
715
+ recommendable: false,
716
+ inputPerMTok: 0.8,
717
+ outputPerMTok: 4,
718
+ contextWindow: 200_000,
719
+ cacheMinTokens: 2048,
720
+ capability: 'small',
721
+ tier: 'haiku',
722
+ notes:
723
+ 'Retired on the Claude API on 2026-02-19; still served on Google Cloud (Google managed).'
724
+ + ' Anthropic names Claude Haiku 4.5 as the replacement.',
725
+ },
588
726
 
589
727
  // ------------------------------------------------------------------------
590
728
  // OpenAI
@@ -645,6 +783,11 @@ export const MODELS: ModelPricing[] = [
645
783
  outputPerMTok: 75,
646
784
  },
647
785
  ],
786
+ serviceTiers: {
787
+ flex: { inputPerMTok: 5, outputPerMTok: 25, tiers: openAiLongBand(10, 37.5) },
788
+ fast: { inputPerMTok: 20, outputPerMTok: 100, tiers: openAiLongBand(40, 150) },
789
+ ultrafast: { inputPerMTok: 60, outputPerMTok: 300, tiers: openAiLongBand(120, 450) },
790
+ },
648
791
  capability: 'frontier',
649
792
  tier: 'frontier',
650
793
  notes: OPENAI_CURRENT_NOTE,
@@ -668,6 +811,11 @@ export const MODELS: ModelPricing[] = [
668
811
  outputPerMTok: 15,
669
812
  },
670
813
  ],
814
+ serviceTiers: {
815
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
816
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
817
+ ultrafast: { inputPerMTok: 12, outputPerMTok: 60, tiers: openAiLongBand(24, 90) },
818
+ },
671
819
  capability: 'large',
672
820
  tier: 'opus',
673
821
  notes: OPENAI_CURRENT_NOTE,
@@ -697,6 +845,10 @@ export const MODELS: ModelPricing[] = [
697
845
  outputPerMTok: 7.5,
698
846
  },
699
847
  ],
848
+ serviceTiers: {
849
+ flex: { inputPerMTok: 0.5, outputPerMTok: 2.5, tiers: openAiLongBand(1, 3.75) },
850
+ fast: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
851
+ },
700
852
  capability: 'mid',
701
853
  tier: 'sonnet',
702
854
  notes: OPENAI_CURRENT_NOTE,
@@ -720,6 +872,10 @@ export const MODELS: ModelPricing[] = [
720
872
  outputPerMTok: 15,
721
873
  },
722
874
  ],
875
+ serviceTiers: {
876
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
877
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
878
+ },
723
879
  capability: 'large',
724
880
  tier: 'opus',
725
881
  notes: OPENAI_CURRENT_NOTE,
@@ -743,12 +899,30 @@ export const MODELS: ModelPricing[] = [
743
899
  outputPerMTok: 0.75,
744
900
  },
745
901
  ],
902
+ serviceTiers: {
903
+ flex: { inputPerMTok: 0.05, outputPerMTok: 0.25, tiers: openAiLongBand(0.1, 0.375) },
904
+ fast: { inputPerMTok: 0.2, outputPerMTok: 1, tiers: openAiLongBand(0.4, 1.5) },
905
+ },
746
906
  capability: 'small',
747
907
  tier: 'haiku',
748
908
  notes: OPENAI_CURRENT_NOTE,
749
909
  },
750
910
  {
911
+ /*
912
+ * Its model page, read 2026-11-05: "The `gpt-5.6` alias routes requests
913
+ * to GPT-5.6 Sol", so a log that names the alias is this model; and
914
+ * "$4 per million input tokens and $20 per million output tokens, a 20%
915
+ * reduction in input pricing and a 33% reduction in output pricing ...
916
+ * promotional pricing is available until November 30, 2026". The price
917
+ * after the promotion is not published as a figure, and deriving one
918
+ * from two rounded percentages would be a price nobody quoted, so the row
919
+ * carries what the page charges today. When the page changes,
920
+ * `scripts/price-drift.mjs` reports it on the next weekly run; until a
921
+ * person reads the new figure, a call after 2026-11-30 may be priced low,
922
+ * and the note says so to anyone reading the row.
923
+ */
751
924
  id: 'gpt-5.6-sol',
925
+ aliases: ['gpt-5.6'],
752
926
  provider: 'openai',
753
927
  displayName: 'GPT-5.6 Sol',
754
928
  inputPerMTok: 4,
@@ -766,9 +940,15 @@ export const MODELS: ModelPricing[] = [
766
940
  outputPerMTok: 30,
767
941
  },
768
942
  ],
943
+ serviceTiers: {
944
+ flex: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
945
+ fast: { inputPerMTok: 8, outputPerMTok: 40, tiers: openAiLongBand(16, 60) },
946
+ },
769
947
  capability: 'large',
770
948
  tier: 'opus',
771
- notes: OPENAI_CURRENT_NOTE,
949
+ notes:
950
+ OPENAI_CURRENT_NOTE
951
+ + ' $4/$20 is promotional until 2026-11-30, and OpenAI has not published the price after it as a figure.',
772
952
  },
773
953
  {
774
954
  id: 'gpt-5.6-terra',
@@ -789,6 +969,10 @@ export const MODELS: ModelPricing[] = [
789
969
  outputPerMTok: 18,
790
970
  },
791
971
  ],
972
+ serviceTiers: {
973
+ flex: { inputPerMTok: 1, outputPerMTok: 6, tiers: openAiLongBand(2, 9) },
974
+ fast: { inputPerMTok: 4, outputPerMTok: 24, tiers: openAiLongBand(8, 36) },
975
+ },
792
976
  capability: 'mid',
793
977
  tier: 'sonnet',
794
978
  notes: OPENAI_CURRENT_NOTE,
@@ -812,56 +996,111 @@ export const MODELS: ModelPricing[] = [
812
996
  outputPerMTok: 1.8,
813
997
  },
814
998
  ],
999
+ serviceTiers: {
1000
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.6, tiers: openAiLongBand(0.2, 0.9) },
1001
+ fast: { inputPerMTok: 0.4, outputPerMTok: 2.4, tiers: openAiLongBand(0.8, 3.6) },
1002
+ },
815
1003
  capability: 'small',
816
1004
  tier: 'haiku',
817
1005
  notes: OPENAI_CURRENT_NOTE,
818
1006
  },
1007
+ {
1008
+ /*
1009
+ * The pricing page's "Cyber models" table and the model's own page
1010
+ * (developers.openai.com/api/docs/models/gpt-5.6-cyber.md), read
1011
+ * 2026-11-05: $12.50/$75, cached input $1.25, the 30-minute write $15.625
1012
+ * and the 12-hour one $31.25, the same 0.1x, 1.25x and 2.5x as the rest of
1013
+ * GPT-5.6. A 400,000-token context window with at most 272,000 input
1014
+ * tokens, so the long-prompt band the model page mentions can never be
1015
+ * reached and none is modelled; the pricing page lists none either. The
1016
+ * Batch endpoint is "Not supported", so there is no batch price.
1017
+ *
1018
+ * Found by `scripts/price-drift.mjs` on its first run, as a model on the
1019
+ * pricing page that read as unpriced. It needs Daybreak programme
1020
+ * approval, so it is priced for the logs of people who have it and never
1021
+ * recommended to anyone else.
1022
+ */
1023
+ id: 'gpt-5.6-cyber',
1024
+ provider: 'openai',
1025
+ displayName: 'GPT-5.6 Cyber',
1026
+ recommendable: false,
1027
+ inputPerMTok: 12.5,
1028
+ outputPerMTok: 75,
1029
+ contextWindow: 400_000,
1030
+ cacheMinTokens: 1024,
1031
+ caching: 'automatic',
1032
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1.25, cacheWrite1h: 2.5, batch: null },
1033
+ cacheTtl: OPENAI_CACHE_TTL,
1034
+ capability: 'large',
1035
+ tier: 'opus',
1036
+ notes:
1037
+ 'Needs approval through the Daybreak programme. At most 272,000 input tokens, Responses API only, no batch.'
1038
+ + ' Cache writes cost 1.25x input for the default 30-minute entry and 2.5x for a 12-hour one.',
1039
+ },
819
1040
 
820
1041
  // ------------------------------------------------------------------------
1042
+ // GPT-5, mini and nano. Each model page, read 2026-11-05, opens with
1043
+ // "Status: Deprecated", and the catalogue still offered all three as
1044
+ // cheaper alternatives, nano as the cheapest OpenAI model there is. Priced
1045
+ // for the logs that hold them; never recommended.
821
1046
  {
822
1047
  id: 'gpt-5',
823
1048
  aliases: ['gpt-5-2025-08-07'],
824
1049
  provider: 'openai',
825
1050
  displayName: 'GPT-5',
1051
+ recommendable: false,
826
1052
  inputPerMTok: 1.25,
827
1053
  outputPerMTok: 10,
828
1054
  contextWindow: 400_000,
829
1055
  cacheMinTokens: 1024,
830
1056
  caching: 'automatic',
831
1057
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1058
+ serviceTiers: {
1059
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1060
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1061
+ },
832
1062
  capability: 'frontier',
833
1063
  tier: 'frontier',
834
- notes: OPENAI_GPT5_NOTE,
1064
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
835
1065
  },
836
1066
  {
837
1067
  id: 'gpt-5-mini',
838
1068
  aliases: ['gpt-5-mini-2025-08-07'],
839
1069
  provider: 'openai',
840
1070
  displayName: 'GPT-5 mini',
1071
+ recommendable: false,
841
1072
  inputPerMTok: 0.25,
842
1073
  outputPerMTok: 2,
843
1074
  contextWindow: 400_000,
844
1075
  cacheMinTokens: 1024,
845
1076
  caching: 'automatic',
846
1077
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1078
+ serviceTiers: {
1079
+ flex: { inputPerMTok: 0.125, outputPerMTok: 1 },
1080
+ fast: { inputPerMTok: 0.45, outputPerMTok: 3.6 },
1081
+ },
847
1082
  capability: 'mid',
848
1083
  tier: 'sonnet',
849
- notes: OPENAI_GPT5_NOTE,
1084
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
850
1085
  },
851
1086
  {
852
1087
  id: 'gpt-5-nano',
853
1088
  aliases: ['gpt-5-nano-2025-08-07'],
854
1089
  provider: 'openai',
855
1090
  displayName: 'GPT-5 nano',
1091
+ recommendable: false,
856
1092
  inputPerMTok: 0.05,
857
1093
  outputPerMTok: 0.4,
858
1094
  contextWindow: 400_000,
859
1095
  cacheMinTokens: 1024,
860
1096
  caching: 'automatic',
861
1097
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1098
+ serviceTiers: {
1099
+ flex: { inputPerMTok: 0.025, outputPerMTok: 0.2 },
1100
+ },
862
1101
  capability: 'small',
863
1102
  tier: 'haiku',
864
- notes: OPENAI_GPT5_NOTE,
1103
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
865
1104
  },
866
1105
 
867
1106
  // ------------------------------------------------------------------------
@@ -899,6 +1138,10 @@ export const MODELS: ModelPricing[] = [
899
1138
  outputPerMTok: 45,
900
1139
  },
901
1140
  ],
1141
+ serviceTiers: {
1142
+ flex: { inputPerMTok: 2.5, outputPerMTok: 15, tiers: openAiLongBand(5, 22.5) },
1143
+ fast: { inputPerMTok: 12.5, outputPerMTok: 75 },
1144
+ },
902
1145
  capability: 'frontier',
903
1146
  tier: 'frontier',
904
1147
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -922,6 +1165,9 @@ export const MODELS: ModelPricing[] = [
922
1165
  outputPerMTok: 270,
923
1166
  },
924
1167
  ],
1168
+ serviceTiers: {
1169
+ flex: { inputPerMTok: 15, outputPerMTok: 90 },
1170
+ },
925
1171
  capability: 'frontier',
926
1172
  tier: 'frontier',
927
1173
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -945,6 +1191,10 @@ export const MODELS: ModelPricing[] = [
945
1191
  outputPerMTok: 22.5,
946
1192
  },
947
1193
  ],
1194
+ serviceTiers: {
1195
+ flex: { inputPerMTok: 1.25, outputPerMTok: 7.5, tiers: openAiLongBand(2.5, 11.25) },
1196
+ fast: { inputPerMTok: 5, outputPerMTok: 30 },
1197
+ },
948
1198
  capability: 'frontier',
949
1199
  tier: 'frontier',
950
1200
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -968,6 +1218,9 @@ export const MODELS: ModelPricing[] = [
968
1218
  outputPerMTok: 270,
969
1219
  },
970
1220
  ],
1221
+ serviceTiers: {
1222
+ flex: { inputPerMTok: 15, outputPerMTok: 90, tiers: openAiLongBand(30, 135) },
1223
+ },
971
1224
  capability: 'frontier',
972
1225
  tier: 'frontier',
973
1226
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -983,6 +1236,10 @@ export const MODELS: ModelPricing[] = [
983
1236
  cacheMinTokens: 1024,
984
1237
  caching: 'automatic',
985
1238
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1239
+ serviceTiers: {
1240
+ flex: { inputPerMTok: 0.375, outputPerMTok: 2.25 },
1241
+ fast: { inputPerMTok: 1.5, outputPerMTok: 9 },
1242
+ },
986
1243
  capability: 'mid',
987
1244
  tier: 'sonnet',
988
1245
  notes: OPENAI_EARLIER_NOTE,
@@ -999,6 +1256,9 @@ export const MODELS: ModelPricing[] = [
999
1256
  cacheMinTokens: 1024,
1000
1257
  caching: 'automatic',
1001
1258
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1259
+ serviceTiers: {
1260
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.625 },
1261
+ },
1002
1262
  capability: 'small',
1003
1263
  tier: 'haiku',
1004
1264
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1014,6 +1274,10 @@ export const MODELS: ModelPricing[] = [
1014
1274
  cacheMinTokens: 1024,
1015
1275
  caching: 'automatic',
1016
1276
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1277
+ serviceTiers: {
1278
+ flex: { inputPerMTok: 0.875, outputPerMTok: 7 },
1279
+ fast: { inputPerMTok: 3.5, outputPerMTok: 28 },
1280
+ },
1017
1281
  capability: 'frontier',
1018
1282
  tier: 'frontier',
1019
1283
  notes: OPENAI_EARLIER_NOTE,
@@ -1045,6 +1309,10 @@ export const MODELS: ModelPricing[] = [
1045
1309
  cacheMinTokens: 1024,
1046
1310
  caching: 'automatic',
1047
1311
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1312
+ serviceTiers: {
1313
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1314
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1315
+ },
1048
1316
  capability: 'frontier',
1049
1317
  tier: 'frontier',
1050
1318
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1077,6 +1345,9 @@ export const MODELS: ModelPricing[] = [
1077
1345
  cacheMinTokens: 1024,
1078
1346
  caching: 'automatic',
1079
1347
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1348
+ serviceTiers: {
1349
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1350
+ },
1080
1351
  capability: 'large',
1081
1352
  tier: 'opus',
1082
1353
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1093,6 +1364,9 @@ export const MODELS: ModelPricing[] = [
1093
1364
  cacheMinTokens: 1024,
1094
1365
  caching: 'automatic',
1095
1366
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1367
+ serviceTiers: {
1368
+ fast: { inputPerMTok: 0.7, outputPerMTok: 2.8 },
1369
+ },
1096
1370
  capability: 'mid',
1097
1371
  tier: 'sonnet',
1098
1372
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1109,6 +1383,9 @@ export const MODELS: ModelPricing[] = [
1109
1383
  cacheMinTokens: 1024,
1110
1384
  caching: 'automatic',
1111
1385
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1386
+ serviceTiers: {
1387
+ fast: { inputPerMTok: 4.25, outputPerMTok: 17 },
1388
+ },
1112
1389
  capability: 'large',
1113
1390
  tier: 'opus',
1114
1391
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1125,6 +1402,9 @@ export const MODELS: ModelPricing[] = [
1125
1402
  cacheMinTokens: 1024,
1126
1403
  caching: 'automatic',
1127
1404
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1405
+ serviceTiers: {
1406
+ fast: { inputPerMTok: 0.25, outputPerMTok: 1 },
1407
+ },
1128
1408
  capability: 'small',
1129
1409
  tier: 'haiku',
1130
1410
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1141,6 +1421,10 @@ export const MODELS: ModelPricing[] = [
1141
1421
  cacheMinTokens: 1024,
1142
1422
  caching: 'automatic',
1143
1423
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1424
+ serviceTiers: {
1425
+ flex: { inputPerMTok: 1, outputPerMTok: 4 },
1426
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1427
+ },
1144
1428
  capability: 'frontier',
1145
1429
  tier: 'frontier',
1146
1430
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1161,6 +1445,143 @@ export const MODELS: ModelPricing[] = [
1161
1445
  tier: 'frontier',
1162
1446
  notes: OPENAI_EARLIER_NO_CACHE_NOTE + OPENAI_DEPRECATED_SUFFIX,
1163
1447
  },
1448
+ // ------------------------------------------------------------------------
1449
+ // The Codex models and chat-latest. Codex CLI sessions name these, and none
1450
+ // was priced, so every such call read as unpriced. Each from its own model
1451
+ // page (developers.openai.com/api/docs/models/<id>.md), read 2026-11-05,
1452
+ // and gpt-5.3-codex and chat-latest also from the pricing page's
1453
+ // "Specialized models" table. Every page says Batch is "Not supported", so
1454
+ // none has a batch price. gpt-5.3-codex is deprecated and the rest of the
1455
+ // Codex line is shut down; all are priced for the logs that hold them and
1456
+ // never recommended. chat-latest is served and not recommended either: it
1457
+ // points at whatever model ChatGPT uses, so a recommendation would not name
1458
+ // a fixed model. Cache reads are 0.1x everywhere but codex-mini-latest,
1459
+ // whose page gives $0.375 on $1.50, 0.25x. gpt-5-search-api is on the
1460
+ // pricing page too and its model page returns 404, so it has no context
1461
+ // window and is left unpriced (see scripts/price-drift.mjs).
1462
+ // ------------------------------------------------------------------------
1463
+ {
1464
+ id: 'gpt-5.3-codex',
1465
+ provider: 'openai',
1466
+ displayName: 'GPT-5.3-Codex',
1467
+ recommendable: false,
1468
+ inputPerMTok: 1.75,
1469
+ outputPerMTok: 14,
1470
+ contextWindow: 400_000,
1471
+ cacheMinTokens: 1024,
1472
+ caching: 'automatic',
1473
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1474
+ // The pricing page's specialized "Fast" table: $3.50/$28, cached $0.35.
1475
+ serviceTiers: { fast: { inputPerMTok: 3.5, outputPerMTok: 28 } },
1476
+ capability: 'frontier',
1477
+ tier: 'frontier',
1478
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_DEPRECATED_SUFFIX,
1479
+ },
1480
+ {
1481
+ id: 'gpt-5.2-codex',
1482
+ provider: 'openai',
1483
+ displayName: 'GPT-5.2-Codex',
1484
+ recommendable: false,
1485
+ inputPerMTok: 1.75,
1486
+ outputPerMTok: 14,
1487
+ contextWindow: 400_000,
1488
+ cacheMinTokens: 1024,
1489
+ caching: 'automatic',
1490
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1491
+ capability: 'frontier',
1492
+ tier: 'frontier',
1493
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1494
+ },
1495
+ {
1496
+ id: 'gpt-5.1-codex',
1497
+ provider: 'openai',
1498
+ displayName: 'GPT-5.1-Codex',
1499
+ recommendable: false,
1500
+ inputPerMTok: 1.25,
1501
+ outputPerMTok: 10,
1502
+ contextWindow: 400_000,
1503
+ cacheMinTokens: 1024,
1504
+ caching: 'automatic',
1505
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1506
+ capability: 'frontier',
1507
+ tier: 'frontier',
1508
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1509
+ },
1510
+ {
1511
+ id: 'gpt-5.1-codex-max',
1512
+ provider: 'openai',
1513
+ displayName: 'GPT-5.1-Codex-Max',
1514
+ recommendable: false,
1515
+ inputPerMTok: 1.25,
1516
+ outputPerMTok: 10,
1517
+ contextWindow: 400_000,
1518
+ cacheMinTokens: 1024,
1519
+ caching: 'automatic',
1520
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1521
+ capability: 'frontier',
1522
+ tier: 'frontier',
1523
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1524
+ },
1525
+ {
1526
+ id: 'gpt-5.1-codex-mini',
1527
+ provider: 'openai',
1528
+ displayName: 'GPT-5.1-Codex Mini',
1529
+ recommendable: false,
1530
+ inputPerMTok: 0.25,
1531
+ outputPerMTok: 2,
1532
+ contextWindow: 400_000,
1533
+ cacheMinTokens: 1024,
1534
+ caching: 'automatic',
1535
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1536
+ capability: 'mid',
1537
+ tier: 'sonnet',
1538
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1539
+ },
1540
+ {
1541
+ id: 'gpt-5-codex',
1542
+ provider: 'openai',
1543
+ displayName: 'GPT-5-Codex',
1544
+ recommendable: false,
1545
+ inputPerMTok: 1.25,
1546
+ outputPerMTok: 10,
1547
+ contextWindow: 400_000,
1548
+ cacheMinTokens: 1024,
1549
+ caching: 'automatic',
1550
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1551
+ capability: 'frontier',
1552
+ tier: 'frontier',
1553
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1554
+ },
1555
+ {
1556
+ id: 'codex-mini-latest',
1557
+ provider: 'openai',
1558
+ displayName: 'codex-mini-latest',
1559
+ recommendable: false,
1560
+ inputPerMTok: 1.5,
1561
+ outputPerMTok: 6,
1562
+ contextWindow: 200_000,
1563
+ cacheMinTokens: 1024,
1564
+ caching: 'automatic',
1565
+ multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1566
+ capability: 'mid',
1567
+ tier: 'sonnet',
1568
+ notes: OPENAI_CODEX_NOTE + ' A cache read costs 0.25x input.' + OPENAI_SHUT_DOWN_SUFFIX,
1569
+ },
1570
+ {
1571
+ id: 'chat-latest',
1572
+ provider: 'openai',
1573
+ displayName: 'Chat Latest',
1574
+ recommendable: false,
1575
+ inputPerMTok: 5,
1576
+ outputPerMTok: 30,
1577
+ contextWindow: 400_000,
1578
+ cacheMinTokens: 1024,
1579
+ caching: 'automatic',
1580
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1581
+ capability: 'large',
1582
+ tier: 'opus',
1583
+ 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.',
1584
+ },
1164
1585
 
1165
1586
  // ------------------------------------------------------------------------
1166
1587
  // Google
@@ -2042,6 +2463,10 @@ interface Rates {
2042
2463
  */
2043
2464
  floor?: { inputPerMTok: number; outputPerMTok: number };
2044
2465
  } | null;
2466
+ /** The service tier whose table priced the call, when it was not the standard one. */
2467
+ serviceTier?: ServiceTier;
2468
+ /** The geography whose published multiplier priced the call, when it was not global. */
2469
+ inferenceGeo?: InferenceGeo;
2045
2470
  }
2046
2471
 
2047
2472
  /**
@@ -2064,6 +2489,81 @@ export interface PromptSize {
2064
2489
  inputTokensAtMost?: number;
2065
2490
  /** The response's `usage.speed`, when the log kept it: `"fast"` or `"standard"`. */
2066
2491
  speed?: string;
2492
+ /**
2493
+ * The response's `service_tier`, when the log kept it: OpenAI's `"flex"`,
2494
+ * `"priority"` or `"fast"`, `"ultrafast"`, or `"default"`. Normalised by
2495
+ * `serviceTierOf`; anything else prices at the standard rate.
2496
+ */
2497
+ serviceTier?: string;
2498
+ /**
2499
+ * The response's `usage.inference_geo`, when the log kept it: Anthropic's
2500
+ * `"us"`, `"global"` or `"not_available"`. Normalised by `inferenceGeoOf`;
2501
+ * only a geography the row publishes a multiplier for changes the price.
2502
+ */
2503
+ inferenceGeo?: string;
2504
+ }
2505
+
2506
+ /**
2507
+ * A response's `usage.inference_geo` as the geography that prices it, or
2508
+ * `undefined` for the global rate. `"global"` is the default and the row's
2509
+ * own price; `"not_available"` is what a model that cannot pin a geography
2510
+ * reports, and is priced the same.
2511
+ */
2512
+ export function inferenceGeoOf(value: unknown): InferenceGeo | undefined {
2513
+ if (typeof value !== 'string') return undefined;
2514
+ return value.trim().toLowerCase() === 'us' ? 'us' : undefined;
2515
+ }
2516
+
2517
+ /** Rounded to a billionth of a dollar per million tokens, so 1.1x of $5 is $5.5 and not $5.500000000000001. */
2518
+ const scaled = (rate: number, factor: number): number => Math.round(rate * factor * 1e9) / 1e9;
2519
+
2520
+ /**
2521
+ * A response's `service_tier` as the tier whose table prices it, or
2522
+ * `undefined` for the standard one.
2523
+ *
2524
+ * OpenAI renamed Priority processing to Fast mode on 2026-07-30, and its guide
2525
+ * says "for GPT-5.6 and earlier models, the response returns `priority`
2526
+ * whether the request specifies `priority` or `fast`", so both spellings are
2527
+ * the Fast table. `"default"` is the standard tier, and is also what a Fast
2528
+ * request downgraded under the ramp-rate limit returns, "and charge standard
2529
+ * rates". `"auto"` and `"scale"`, or a value this table has never seen, are
2530
+ * priced at the standard rate rather than guessed into another tier.
2531
+ */
2532
+ export function serviceTierOf(value: unknown): ServiceTier | undefined {
2533
+ if (typeof value !== 'string') return undefined;
2534
+ const tier = value.trim().toLowerCase();
2535
+ if (tier === 'flex') return 'flex';
2536
+ if (tier === 'fast' || tier === 'priority') return 'fast';
2537
+ if (tier === 'ultrafast') return 'ultrafast';
2538
+ return undefined;
2539
+ }
2540
+
2541
+ /**
2542
+ * The row as one service tier prices it: that tier's rates and long band, the
2543
+ * row's own cache and batch ratios. Built once per row and tier.
2544
+ */
2545
+ const serviceTierMemo = new WeakMap<ModelPricing, Map<ServiceTier, ModelPricing>>();
2546
+
2547
+ function serviceTierRow(model: ModelPricing, tier: ServiceTier): ModelPricing | undefined {
2548
+ const sheet = model.serviceTiers?.[tier];
2549
+ if (sheet === undefined) return undefined;
2550
+ let rows = serviceTierMemo.get(model);
2551
+ if (rows === undefined) {
2552
+ rows = new Map();
2553
+ serviceTierMemo.set(model, rows);
2554
+ }
2555
+ let row = rows.get(tier);
2556
+ if (row === undefined) {
2557
+ const { history: _history, promo: _promo, tiers: _tiers, serviceTiers: _serviceTiers, ...rest } = model;
2558
+ row = Object.freeze({
2559
+ ...rest,
2560
+ inputPerMTok: sheet.inputPerMTok,
2561
+ outputPerMTok: sheet.outputPerMTok,
2562
+ ...(sheet.tiers !== undefined ? { tiers: sheet.tiers } : {}),
2563
+ });
2564
+ rows.set(tier, row);
2565
+ }
2566
+ return row;
2067
2567
  }
2068
2568
 
2069
2569
  const pricingMemo = new WeakMap<
@@ -2084,7 +2584,8 @@ function pricedAt(model: ModelPricing, on: Date): ModelPricing {
2084
2584
  let periods = periodMemo.get(model);
2085
2585
  if (periods === undefined) {
2086
2586
  periods = history.map((period) => {
2087
- const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, ...rest } = model;
2587
+ // A period's own price only: the service tiers below are today's tables.
2588
+ const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, serviceTiers: _serviceTiers, ...rest } = model;
2088
2589
  const row: ModelPricing = Object.freeze({
2089
2590
  ...rest,
2090
2591
  inputPerMTok: period.inputPerMTok,
@@ -2176,11 +2677,52 @@ export function effectivePricing(
2176
2677
  on: Date = new Date(),
2177
2678
  context: PromptSize = {},
2178
2679
  ): Rates {
2680
+ /*
2681
+ * A geography with a published multiplier scales whatever the rest of the
2682
+ * call decides, fast mode and the long band included, because the page says
2683
+ * it stacks with all of them. Applied once, outside everything else.
2684
+ */
2685
+ if (context.inferenceGeo !== undefined) {
2686
+ const { inferenceGeo, ...rest } = context;
2687
+ const rates = effectivePricing(model, on, rest);
2688
+ const geo = inferenceGeoOf(inferenceGeo);
2689
+ const factor = geo === undefined ? undefined : model.inferenceGeoMultipliers?.[geo];
2690
+ if (geo === undefined || factor === undefined) return rates;
2691
+ return {
2692
+ ...rates,
2693
+ inputPerMTok: scaled(rates.inputPerMTok, factor),
2694
+ outputPerMTok: scaled(rates.outputPerMTok, factor),
2695
+ ...(rates.tier?.floor !== undefined
2696
+ ? {
2697
+ tier: {
2698
+ ...rates.tier,
2699
+ floor: {
2700
+ inputPerMTok: scaled(rates.tier.floor.inputPerMTok, factor),
2701
+ outputPerMTok: scaled(rates.tier.floor.outputPerMTok, factor),
2702
+ },
2703
+ },
2704
+ }
2705
+ : {}),
2706
+ inferenceGeo: geo,
2707
+ };
2708
+ }
2179
2709
  /* A call before a published price change is priced as it was billed then. */
2180
2710
  if (model.history !== undefined) {
2181
2711
  const period = pricedAt(model, on);
2182
2712
  if (period !== model) return effectivePricing(period, on, context);
2183
2713
  }
2714
+ /*
2715
+ * A response that names another service tier is priced by that tier's own
2716
+ * table, long band included, and by nothing composed out of two.
2717
+ */
2718
+ if (context.serviceTier !== undefined && model.serviceTiers !== undefined) {
2719
+ const named = serviceTierOf(context.serviceTier);
2720
+ const row = named === undefined ? undefined : serviceTierRow(model, named);
2721
+ if (row !== undefined) {
2722
+ const { serviceTier: _serviceTier, ...size } = context;
2723
+ return { ...effectivePricing(row, on, size), serviceTier: named };
2724
+ }
2725
+ }
2184
2726
  const memo = pricingMemo.get(model) ?? cache(model);
2185
2727
 
2186
2728
  /**