@trazum/core 2.7.1 → 2.7.2

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/src/pricing.ts CHANGED
@@ -1,4 +1,4 @@
1
- import type { ModelPricing, PricingTier, TierCondition } from './types.js';
1
+ import type { 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'
@@ -464,6 +482,53 @@ export const MODELS: ModelPricing[] = [
464
482
  tier: 'opus',
465
483
  notes: '200K context window and a 64K output cap.',
466
484
  },
485
+ {
486
+ /*
487
+ * Retired on the Claude API on 2026-08-05 (the deprecations page, read
488
+ * 2026-11-05) and still served on Claude Models on Google Cloud (Google
489
+ * managed) and on Bedrock. Priced because a year of logs carries it and,
490
+ * unpriced, every one of those calls read as a gap. The pricing page, read
491
+ * the same day: $15/$75, writes $18.75 and $30, cache hits $1.50, batch
492
+ * $7.50/$37.50, which are the default multipliers. Its model page is gone;
493
+ * the context-windows page gives every Claude model outside its 1M list a
494
+ * 200k window, and the prompt-caching page a 1,024-token cache minimum.
495
+ * The dated id is the one the deprecations page names and what a log
496
+ * carries; no shorter alias is on any page read, so none is declared.
497
+ */
498
+ id: 'claude-opus-4-1-20250805',
499
+ provider: 'anthropic',
500
+ displayName: 'Claude Opus 4.1',
501
+ recommendable: false,
502
+ inputPerMTok: 15,
503
+ outputPerMTok: 75,
504
+ contextWindow: 200_000,
505
+ cacheMinTokens: 1024,
506
+ capability: 'large',
507
+ tier: 'opus',
508
+ notes:
509
+ 'Retired on the Claude API on 2026-08-05; still served on Google Cloud (Google managed) and Bedrock.'
510
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
511
+ },
512
+ {
513
+ /*
514
+ * Retired on the Claude API on 2026-06-15; still served on Claude Models
515
+ * on Google Cloud (Google managed). The same pages and the same figures as
516
+ * Opus 4.1 above: $15/$75, 200k context, a 1,024-token cache minimum.
517
+ */
518
+ id: 'claude-opus-4-20250514',
519
+ provider: 'anthropic',
520
+ displayName: 'Claude Opus 4',
521
+ recommendable: false,
522
+ inputPerMTok: 15,
523
+ outputPerMTok: 75,
524
+ contextWindow: 200_000,
525
+ cacheMinTokens: 1024,
526
+ capability: 'large',
527
+ tier: 'opus',
528
+ notes:
529
+ 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed).'
530
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
531
+ },
467
532
  {
468
533
  /*
469
534
  * The pricing page and the Sonnet 5.5 model page, read 2026-10-10. The
@@ -530,6 +595,30 @@ export const MODELS: ModelPricing[] = [
530
595
  tier: 'sonnet',
531
596
  notes: 'Deprecated; retires 2026-11-30. Recommended replacement: Claude Sonnet 5.5.',
532
597
  },
598
+ {
599
+ /*
600
+ * Retired on the Claude API on 2026-06-15 (the deprecations page, read
601
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
602
+ * managed) and on Bedrock. The pricing page, read the same day: $3/$15,
603
+ * writes $3.75 and $6, cache hits $0.30, batch $1.50/$7.50. 200k context
604
+ * from the context-windows page, a 1,024-token cache minimum from the
605
+ * prompt-caching page. The pricing page publishes one rate for it and no
606
+ * long-context band, so every call is priced at that rate.
607
+ */
608
+ id: 'claude-sonnet-4-20250514',
609
+ provider: 'anthropic',
610
+ displayName: 'Claude Sonnet 4',
611
+ recommendable: false,
612
+ inputPerMTok: 3,
613
+ outputPerMTok: 15,
614
+ contextWindow: 200_000,
615
+ cacheMinTokens: 1024,
616
+ capability: 'mid',
617
+ tier: 'sonnet',
618
+ notes:
619
+ 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed) and Bedrock.'
620
+ + ' Anthropic names Claude Sonnet 5.5 as the replacement.',
621
+ },
533
622
  {
534
623
  /*
535
624
  * The pricing page, the models overview and the prompt-caching page, read
@@ -585,6 +674,29 @@ export const MODELS: ModelPricing[] = [
585
674
  tier: 'haiku',
586
675
  notes: '200K context window and a 64K output cap, smaller than the rest.',
587
676
  },
677
+ {
678
+ /*
679
+ * Retired on the Claude API on 2026-02-19 (the deprecations page, read
680
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
681
+ * managed). The pricing page, read the same day: $0.80/$4, writes $1 and
682
+ * $1.60, cache hits $0.08, batch $0.40/$2, the default multipliers. 200k
683
+ * context from the context-windows page, and the prompt-caching page's
684
+ * 2,048-token cache minimum.
685
+ */
686
+ id: 'claude-3-5-haiku-20241022',
687
+ provider: 'anthropic',
688
+ displayName: 'Claude Haiku 3.5',
689
+ recommendable: false,
690
+ inputPerMTok: 0.8,
691
+ outputPerMTok: 4,
692
+ contextWindow: 200_000,
693
+ cacheMinTokens: 2048,
694
+ capability: 'small',
695
+ tier: 'haiku',
696
+ notes:
697
+ 'Retired on the Claude API on 2026-02-19; still served on Google Cloud (Google managed).'
698
+ + ' Anthropic names Claude Haiku 4.5 as the replacement.',
699
+ },
588
700
 
589
701
  // ------------------------------------------------------------------------
590
702
  // OpenAI
@@ -645,6 +757,11 @@ export const MODELS: ModelPricing[] = [
645
757
  outputPerMTok: 75,
646
758
  },
647
759
  ],
760
+ serviceTiers: {
761
+ flex: { inputPerMTok: 5, outputPerMTok: 25, tiers: openAiLongBand(10, 37.5) },
762
+ fast: { inputPerMTok: 20, outputPerMTok: 100, tiers: openAiLongBand(40, 150) },
763
+ ultrafast: { inputPerMTok: 60, outputPerMTok: 300, tiers: openAiLongBand(120, 450) },
764
+ },
648
765
  capability: 'frontier',
649
766
  tier: 'frontier',
650
767
  notes: OPENAI_CURRENT_NOTE,
@@ -668,6 +785,11 @@ export const MODELS: ModelPricing[] = [
668
785
  outputPerMTok: 15,
669
786
  },
670
787
  ],
788
+ serviceTiers: {
789
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
790
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
791
+ ultrafast: { inputPerMTok: 12, outputPerMTok: 60, tiers: openAiLongBand(24, 90) },
792
+ },
671
793
  capability: 'large',
672
794
  tier: 'opus',
673
795
  notes: OPENAI_CURRENT_NOTE,
@@ -697,6 +819,10 @@ export const MODELS: ModelPricing[] = [
697
819
  outputPerMTok: 7.5,
698
820
  },
699
821
  ],
822
+ serviceTiers: {
823
+ flex: { inputPerMTok: 0.5, outputPerMTok: 2.5, tiers: openAiLongBand(1, 3.75) },
824
+ fast: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
825
+ },
700
826
  capability: 'mid',
701
827
  tier: 'sonnet',
702
828
  notes: OPENAI_CURRENT_NOTE,
@@ -720,6 +846,10 @@ export const MODELS: ModelPricing[] = [
720
846
  outputPerMTok: 15,
721
847
  },
722
848
  ],
849
+ serviceTiers: {
850
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
851
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
852
+ },
723
853
  capability: 'large',
724
854
  tier: 'opus',
725
855
  notes: OPENAI_CURRENT_NOTE,
@@ -743,12 +873,30 @@ export const MODELS: ModelPricing[] = [
743
873
  outputPerMTok: 0.75,
744
874
  },
745
875
  ],
876
+ serviceTiers: {
877
+ flex: { inputPerMTok: 0.05, outputPerMTok: 0.25, tiers: openAiLongBand(0.1, 0.375) },
878
+ fast: { inputPerMTok: 0.2, outputPerMTok: 1, tiers: openAiLongBand(0.4, 1.5) },
879
+ },
746
880
  capability: 'small',
747
881
  tier: 'haiku',
748
882
  notes: OPENAI_CURRENT_NOTE,
749
883
  },
750
884
  {
885
+ /*
886
+ * Its model page, read 2026-11-05: "The `gpt-5.6` alias routes requests
887
+ * to GPT-5.6 Sol", so a log that names the alias is this model; and
888
+ * "$4 per million input tokens and $20 per million output tokens, a 20%
889
+ * reduction in input pricing and a 33% reduction in output pricing ...
890
+ * promotional pricing is available until November 30, 2026". The price
891
+ * after the promotion is not published as a figure, and deriving one
892
+ * from two rounded percentages would be a price nobody quoted, so the row
893
+ * carries what the page charges today. When the page changes,
894
+ * `scripts/price-drift.mjs` reports it on the next weekly run; until a
895
+ * person reads the new figure, a call after 2026-11-30 may be priced low,
896
+ * and the note says so to anyone reading the row.
897
+ */
751
898
  id: 'gpt-5.6-sol',
899
+ aliases: ['gpt-5.6'],
752
900
  provider: 'openai',
753
901
  displayName: 'GPT-5.6 Sol',
754
902
  inputPerMTok: 4,
@@ -766,9 +914,15 @@ export const MODELS: ModelPricing[] = [
766
914
  outputPerMTok: 30,
767
915
  },
768
916
  ],
917
+ serviceTiers: {
918
+ flex: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
919
+ fast: { inputPerMTok: 8, outputPerMTok: 40, tiers: openAiLongBand(16, 60) },
920
+ },
769
921
  capability: 'large',
770
922
  tier: 'opus',
771
- notes: OPENAI_CURRENT_NOTE,
923
+ notes:
924
+ OPENAI_CURRENT_NOTE
925
+ + ' $4/$20 is promotional until 2026-11-30, and OpenAI has not published the price after it as a figure.',
772
926
  },
773
927
  {
774
928
  id: 'gpt-5.6-terra',
@@ -789,6 +943,10 @@ export const MODELS: ModelPricing[] = [
789
943
  outputPerMTok: 18,
790
944
  },
791
945
  ],
946
+ serviceTiers: {
947
+ flex: { inputPerMTok: 1, outputPerMTok: 6, tiers: openAiLongBand(2, 9) },
948
+ fast: { inputPerMTok: 4, outputPerMTok: 24, tiers: openAiLongBand(8, 36) },
949
+ },
792
950
  capability: 'mid',
793
951
  tier: 'sonnet',
794
952
  notes: OPENAI_CURRENT_NOTE,
@@ -812,56 +970,111 @@ export const MODELS: ModelPricing[] = [
812
970
  outputPerMTok: 1.8,
813
971
  },
814
972
  ],
973
+ serviceTiers: {
974
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.6, tiers: openAiLongBand(0.2, 0.9) },
975
+ fast: { inputPerMTok: 0.4, outputPerMTok: 2.4, tiers: openAiLongBand(0.8, 3.6) },
976
+ },
815
977
  capability: 'small',
816
978
  tier: 'haiku',
817
979
  notes: OPENAI_CURRENT_NOTE,
818
980
  },
981
+ {
982
+ /*
983
+ * The pricing page's "Cyber models" table and the model's own page
984
+ * (developers.openai.com/api/docs/models/gpt-5.6-cyber.md), read
985
+ * 2026-11-05: $12.50/$75, cached input $1.25, the 30-minute write $15.625
986
+ * and the 12-hour one $31.25, the same 0.1x, 1.25x and 2.5x as the rest of
987
+ * GPT-5.6. A 400,000-token context window with at most 272,000 input
988
+ * tokens, so the long-prompt band the model page mentions can never be
989
+ * reached and none is modelled; the pricing page lists none either. The
990
+ * Batch endpoint is "Not supported", so there is no batch price.
991
+ *
992
+ * Found by `scripts/price-drift.mjs` on its first run, as a model on the
993
+ * pricing page that read as unpriced. It needs Daybreak programme
994
+ * approval, so it is priced for the logs of people who have it and never
995
+ * recommended to anyone else.
996
+ */
997
+ id: 'gpt-5.6-cyber',
998
+ provider: 'openai',
999
+ displayName: 'GPT-5.6 Cyber',
1000
+ recommendable: false,
1001
+ inputPerMTok: 12.5,
1002
+ outputPerMTok: 75,
1003
+ contextWindow: 400_000,
1004
+ cacheMinTokens: 1024,
1005
+ caching: 'automatic',
1006
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1.25, cacheWrite1h: 2.5, batch: null },
1007
+ cacheTtl: OPENAI_CACHE_TTL,
1008
+ capability: 'large',
1009
+ tier: 'opus',
1010
+ notes:
1011
+ 'Needs approval through the Daybreak programme. At most 272,000 input tokens, Responses API only, no batch.'
1012
+ + ' Cache writes cost 1.25x input for the default 30-minute entry and 2.5x for a 12-hour one.',
1013
+ },
819
1014
 
820
1015
  // ------------------------------------------------------------------------
1016
+ // GPT-5, mini and nano. Each model page, read 2026-11-05, opens with
1017
+ // "Status: Deprecated", and the catalogue still offered all three as
1018
+ // cheaper alternatives, nano as the cheapest OpenAI model there is. Priced
1019
+ // for the logs that hold them; never recommended.
821
1020
  {
822
1021
  id: 'gpt-5',
823
1022
  aliases: ['gpt-5-2025-08-07'],
824
1023
  provider: 'openai',
825
1024
  displayName: 'GPT-5',
1025
+ recommendable: false,
826
1026
  inputPerMTok: 1.25,
827
1027
  outputPerMTok: 10,
828
1028
  contextWindow: 400_000,
829
1029
  cacheMinTokens: 1024,
830
1030
  caching: 'automatic',
831
1031
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1032
+ serviceTiers: {
1033
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1034
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1035
+ },
832
1036
  capability: 'frontier',
833
1037
  tier: 'frontier',
834
- notes: OPENAI_GPT5_NOTE,
1038
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
835
1039
  },
836
1040
  {
837
1041
  id: 'gpt-5-mini',
838
1042
  aliases: ['gpt-5-mini-2025-08-07'],
839
1043
  provider: 'openai',
840
1044
  displayName: 'GPT-5 mini',
1045
+ recommendable: false,
841
1046
  inputPerMTok: 0.25,
842
1047
  outputPerMTok: 2,
843
1048
  contextWindow: 400_000,
844
1049
  cacheMinTokens: 1024,
845
1050
  caching: 'automatic',
846
1051
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1052
+ serviceTiers: {
1053
+ flex: { inputPerMTok: 0.125, outputPerMTok: 1 },
1054
+ fast: { inputPerMTok: 0.45, outputPerMTok: 3.6 },
1055
+ },
847
1056
  capability: 'mid',
848
1057
  tier: 'sonnet',
849
- notes: OPENAI_GPT5_NOTE,
1058
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
850
1059
  },
851
1060
  {
852
1061
  id: 'gpt-5-nano',
853
1062
  aliases: ['gpt-5-nano-2025-08-07'],
854
1063
  provider: 'openai',
855
1064
  displayName: 'GPT-5 nano',
1065
+ recommendable: false,
856
1066
  inputPerMTok: 0.05,
857
1067
  outputPerMTok: 0.4,
858
1068
  contextWindow: 400_000,
859
1069
  cacheMinTokens: 1024,
860
1070
  caching: 'automatic',
861
1071
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1072
+ serviceTiers: {
1073
+ flex: { inputPerMTok: 0.025, outputPerMTok: 0.2 },
1074
+ },
862
1075
  capability: 'small',
863
1076
  tier: 'haiku',
864
- notes: OPENAI_GPT5_NOTE,
1077
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
865
1078
  },
866
1079
 
867
1080
  // ------------------------------------------------------------------------
@@ -899,6 +1112,10 @@ export const MODELS: ModelPricing[] = [
899
1112
  outputPerMTok: 45,
900
1113
  },
901
1114
  ],
1115
+ serviceTiers: {
1116
+ flex: { inputPerMTok: 2.5, outputPerMTok: 15, tiers: openAiLongBand(5, 22.5) },
1117
+ fast: { inputPerMTok: 12.5, outputPerMTok: 75 },
1118
+ },
902
1119
  capability: 'frontier',
903
1120
  tier: 'frontier',
904
1121
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -922,6 +1139,9 @@ export const MODELS: ModelPricing[] = [
922
1139
  outputPerMTok: 270,
923
1140
  },
924
1141
  ],
1142
+ serviceTiers: {
1143
+ flex: { inputPerMTok: 15, outputPerMTok: 90 },
1144
+ },
925
1145
  capability: 'frontier',
926
1146
  tier: 'frontier',
927
1147
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -945,6 +1165,10 @@ export const MODELS: ModelPricing[] = [
945
1165
  outputPerMTok: 22.5,
946
1166
  },
947
1167
  ],
1168
+ serviceTiers: {
1169
+ flex: { inputPerMTok: 1.25, outputPerMTok: 7.5, tiers: openAiLongBand(2.5, 11.25) },
1170
+ fast: { inputPerMTok: 5, outputPerMTok: 30 },
1171
+ },
948
1172
  capability: 'frontier',
949
1173
  tier: 'frontier',
950
1174
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -968,6 +1192,9 @@ export const MODELS: ModelPricing[] = [
968
1192
  outputPerMTok: 270,
969
1193
  },
970
1194
  ],
1195
+ serviceTiers: {
1196
+ flex: { inputPerMTok: 15, outputPerMTok: 90, tiers: openAiLongBand(30, 135) },
1197
+ },
971
1198
  capability: 'frontier',
972
1199
  tier: 'frontier',
973
1200
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -983,6 +1210,10 @@ export const MODELS: ModelPricing[] = [
983
1210
  cacheMinTokens: 1024,
984
1211
  caching: 'automatic',
985
1212
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1213
+ serviceTiers: {
1214
+ flex: { inputPerMTok: 0.375, outputPerMTok: 2.25 },
1215
+ fast: { inputPerMTok: 1.5, outputPerMTok: 9 },
1216
+ },
986
1217
  capability: 'mid',
987
1218
  tier: 'sonnet',
988
1219
  notes: OPENAI_EARLIER_NOTE,
@@ -999,6 +1230,9 @@ export const MODELS: ModelPricing[] = [
999
1230
  cacheMinTokens: 1024,
1000
1231
  caching: 'automatic',
1001
1232
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1233
+ serviceTiers: {
1234
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.625 },
1235
+ },
1002
1236
  capability: 'small',
1003
1237
  tier: 'haiku',
1004
1238
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1014,6 +1248,10 @@ export const MODELS: ModelPricing[] = [
1014
1248
  cacheMinTokens: 1024,
1015
1249
  caching: 'automatic',
1016
1250
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1251
+ serviceTiers: {
1252
+ flex: { inputPerMTok: 0.875, outputPerMTok: 7 },
1253
+ fast: { inputPerMTok: 3.5, outputPerMTok: 28 },
1254
+ },
1017
1255
  capability: 'frontier',
1018
1256
  tier: 'frontier',
1019
1257
  notes: OPENAI_EARLIER_NOTE,
@@ -1045,6 +1283,10 @@ export const MODELS: ModelPricing[] = [
1045
1283
  cacheMinTokens: 1024,
1046
1284
  caching: 'automatic',
1047
1285
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1286
+ serviceTiers: {
1287
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1288
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1289
+ },
1048
1290
  capability: 'frontier',
1049
1291
  tier: 'frontier',
1050
1292
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1077,6 +1319,9 @@ export const MODELS: ModelPricing[] = [
1077
1319
  cacheMinTokens: 1024,
1078
1320
  caching: 'automatic',
1079
1321
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1322
+ serviceTiers: {
1323
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1324
+ },
1080
1325
  capability: 'large',
1081
1326
  tier: 'opus',
1082
1327
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1093,6 +1338,9 @@ export const MODELS: ModelPricing[] = [
1093
1338
  cacheMinTokens: 1024,
1094
1339
  caching: 'automatic',
1095
1340
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1341
+ serviceTiers: {
1342
+ fast: { inputPerMTok: 0.7, outputPerMTok: 2.8 },
1343
+ },
1096
1344
  capability: 'mid',
1097
1345
  tier: 'sonnet',
1098
1346
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1109,6 +1357,9 @@ export const MODELS: ModelPricing[] = [
1109
1357
  cacheMinTokens: 1024,
1110
1358
  caching: 'automatic',
1111
1359
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1360
+ serviceTiers: {
1361
+ fast: { inputPerMTok: 4.25, outputPerMTok: 17 },
1362
+ },
1112
1363
  capability: 'large',
1113
1364
  tier: 'opus',
1114
1365
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1125,6 +1376,9 @@ export const MODELS: ModelPricing[] = [
1125
1376
  cacheMinTokens: 1024,
1126
1377
  caching: 'automatic',
1127
1378
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1379
+ serviceTiers: {
1380
+ fast: { inputPerMTok: 0.25, outputPerMTok: 1 },
1381
+ },
1128
1382
  capability: 'small',
1129
1383
  tier: 'haiku',
1130
1384
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1141,6 +1395,10 @@ export const MODELS: ModelPricing[] = [
1141
1395
  cacheMinTokens: 1024,
1142
1396
  caching: 'automatic',
1143
1397
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1398
+ serviceTiers: {
1399
+ flex: { inputPerMTok: 1, outputPerMTok: 4 },
1400
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1401
+ },
1144
1402
  capability: 'frontier',
1145
1403
  tier: 'frontier',
1146
1404
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1161,6 +1419,143 @@ export const MODELS: ModelPricing[] = [
1161
1419
  tier: 'frontier',
1162
1420
  notes: OPENAI_EARLIER_NO_CACHE_NOTE + OPENAI_DEPRECATED_SUFFIX,
1163
1421
  },
1422
+ // ------------------------------------------------------------------------
1423
+ // The Codex models and chat-latest. Codex CLI sessions name these, and none
1424
+ // was priced, so every such call read as unpriced. Each from its own model
1425
+ // page (developers.openai.com/api/docs/models/<id>.md), read 2026-11-05,
1426
+ // and gpt-5.3-codex and chat-latest also from the pricing page's
1427
+ // "Specialized models" table. Every page says Batch is "Not supported", so
1428
+ // none has a batch price. gpt-5.3-codex is deprecated and the rest of the
1429
+ // Codex line is shut down; all are priced for the logs that hold them and
1430
+ // never recommended. chat-latest is served and not recommended either: it
1431
+ // points at whatever model ChatGPT uses, so a recommendation would not name
1432
+ // a fixed model. Cache reads are 0.1x everywhere but codex-mini-latest,
1433
+ // whose page gives $0.375 on $1.50, 0.25x. gpt-5-search-api is on the
1434
+ // pricing page too and its model page returns 404, so it has no context
1435
+ // window and is left unpriced (see scripts/price-drift.mjs).
1436
+ // ------------------------------------------------------------------------
1437
+ {
1438
+ id: 'gpt-5.3-codex',
1439
+ provider: 'openai',
1440
+ displayName: 'GPT-5.3-Codex',
1441
+ recommendable: false,
1442
+ inputPerMTok: 1.75,
1443
+ outputPerMTok: 14,
1444
+ contextWindow: 400_000,
1445
+ cacheMinTokens: 1024,
1446
+ caching: 'automatic',
1447
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1448
+ // The pricing page's specialized "Fast" table: $3.50/$28, cached $0.35.
1449
+ serviceTiers: { fast: { inputPerMTok: 3.5, outputPerMTok: 28 } },
1450
+ capability: 'frontier',
1451
+ tier: 'frontier',
1452
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_DEPRECATED_SUFFIX,
1453
+ },
1454
+ {
1455
+ id: 'gpt-5.2-codex',
1456
+ provider: 'openai',
1457
+ displayName: 'GPT-5.2-Codex',
1458
+ recommendable: false,
1459
+ inputPerMTok: 1.75,
1460
+ outputPerMTok: 14,
1461
+ contextWindow: 400_000,
1462
+ cacheMinTokens: 1024,
1463
+ caching: 'automatic',
1464
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1465
+ capability: 'frontier',
1466
+ tier: 'frontier',
1467
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1468
+ },
1469
+ {
1470
+ id: 'gpt-5.1-codex',
1471
+ provider: 'openai',
1472
+ displayName: 'GPT-5.1-Codex',
1473
+ recommendable: false,
1474
+ inputPerMTok: 1.25,
1475
+ outputPerMTok: 10,
1476
+ contextWindow: 400_000,
1477
+ cacheMinTokens: 1024,
1478
+ caching: 'automatic',
1479
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1480
+ capability: 'frontier',
1481
+ tier: 'frontier',
1482
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1483
+ },
1484
+ {
1485
+ id: 'gpt-5.1-codex-max',
1486
+ provider: 'openai',
1487
+ displayName: 'GPT-5.1-Codex-Max',
1488
+ recommendable: false,
1489
+ inputPerMTok: 1.25,
1490
+ outputPerMTok: 10,
1491
+ contextWindow: 400_000,
1492
+ cacheMinTokens: 1024,
1493
+ caching: 'automatic',
1494
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1495
+ capability: 'frontier',
1496
+ tier: 'frontier',
1497
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1498
+ },
1499
+ {
1500
+ id: 'gpt-5.1-codex-mini',
1501
+ provider: 'openai',
1502
+ displayName: 'GPT-5.1-Codex Mini',
1503
+ recommendable: false,
1504
+ inputPerMTok: 0.25,
1505
+ outputPerMTok: 2,
1506
+ contextWindow: 400_000,
1507
+ cacheMinTokens: 1024,
1508
+ caching: 'automatic',
1509
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1510
+ capability: 'mid',
1511
+ tier: 'sonnet',
1512
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1513
+ },
1514
+ {
1515
+ id: 'gpt-5-codex',
1516
+ provider: 'openai',
1517
+ displayName: 'GPT-5-Codex',
1518
+ recommendable: false,
1519
+ inputPerMTok: 1.25,
1520
+ outputPerMTok: 10,
1521
+ contextWindow: 400_000,
1522
+ cacheMinTokens: 1024,
1523
+ caching: 'automatic',
1524
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1525
+ capability: 'frontier',
1526
+ tier: 'frontier',
1527
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1528
+ },
1529
+ {
1530
+ id: 'codex-mini-latest',
1531
+ provider: 'openai',
1532
+ displayName: 'codex-mini-latest',
1533
+ recommendable: false,
1534
+ inputPerMTok: 1.5,
1535
+ outputPerMTok: 6,
1536
+ contextWindow: 200_000,
1537
+ cacheMinTokens: 1024,
1538
+ caching: 'automatic',
1539
+ multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1540
+ capability: 'mid',
1541
+ tier: 'sonnet',
1542
+ notes: OPENAI_CODEX_NOTE + ' A cache read costs 0.25x input.' + OPENAI_SHUT_DOWN_SUFFIX,
1543
+ },
1544
+ {
1545
+ id: 'chat-latest',
1546
+ provider: 'openai',
1547
+ displayName: 'Chat Latest',
1548
+ recommendable: false,
1549
+ inputPerMTok: 5,
1550
+ outputPerMTok: 30,
1551
+ contextWindow: 400_000,
1552
+ cacheMinTokens: 1024,
1553
+ caching: 'automatic',
1554
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1555
+ capability: 'large',
1556
+ tier: 'opus',
1557
+ 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.',
1558
+ },
1164
1559
 
1165
1560
  // ------------------------------------------------------------------------
1166
1561
  // Google
@@ -2042,6 +2437,8 @@ interface Rates {
2042
2437
  */
2043
2438
  floor?: { inputPerMTok: number; outputPerMTok: number };
2044
2439
  } | null;
2440
+ /** The service tier whose table priced the call, when it was not the standard one. */
2441
+ serviceTier?: ServiceTier;
2045
2442
  }
2046
2443
 
2047
2444
  /**
@@ -2064,6 +2461,61 @@ export interface PromptSize {
2064
2461
  inputTokensAtMost?: number;
2065
2462
  /** The response's `usage.speed`, when the log kept it: `"fast"` or `"standard"`. */
2066
2463
  speed?: string;
2464
+ /**
2465
+ * The response's `service_tier`, when the log kept it: OpenAI's `"flex"`,
2466
+ * `"priority"` or `"fast"`, `"ultrafast"`, or `"default"`. Normalised by
2467
+ * `serviceTierOf`; anything else prices at the standard rate.
2468
+ */
2469
+ serviceTier?: string;
2470
+ }
2471
+
2472
+ /**
2473
+ * A response's `service_tier` as the tier whose table prices it, or
2474
+ * `undefined` for the standard one.
2475
+ *
2476
+ * OpenAI renamed Priority processing to Fast mode on 2026-07-30, and its guide
2477
+ * says "for GPT-5.6 and earlier models, the response returns `priority`
2478
+ * whether the request specifies `priority` or `fast`", so both spellings are
2479
+ * the Fast table. `"default"` is the standard tier, and is also what a Fast
2480
+ * request downgraded under the ramp-rate limit returns, "and charge standard
2481
+ * rates". `"auto"` and `"scale"`, or a value this table has never seen, are
2482
+ * priced at the standard rate rather than guessed into another tier.
2483
+ */
2484
+ export function serviceTierOf(value: unknown): ServiceTier | undefined {
2485
+ if (typeof value !== 'string') return undefined;
2486
+ const tier = value.trim().toLowerCase();
2487
+ if (tier === 'flex') return 'flex';
2488
+ if (tier === 'fast' || tier === 'priority') return 'fast';
2489
+ if (tier === 'ultrafast') return 'ultrafast';
2490
+ return undefined;
2491
+ }
2492
+
2493
+ /**
2494
+ * The row as one service tier prices it: that tier's rates and long band, the
2495
+ * row's own cache and batch ratios. Built once per row and tier.
2496
+ */
2497
+ const serviceTierMemo = new WeakMap<ModelPricing, Map<ServiceTier, ModelPricing>>();
2498
+
2499
+ function serviceTierRow(model: ModelPricing, tier: ServiceTier): ModelPricing | undefined {
2500
+ const sheet = model.serviceTiers?.[tier];
2501
+ if (sheet === undefined) return undefined;
2502
+ let rows = serviceTierMemo.get(model);
2503
+ if (rows === undefined) {
2504
+ rows = new Map();
2505
+ serviceTierMemo.set(model, rows);
2506
+ }
2507
+ let row = rows.get(tier);
2508
+ if (row === undefined) {
2509
+ const { history: _history, promo: _promo, tiers: _tiers, serviceTiers: _serviceTiers, ...rest } = model;
2510
+ row = Object.freeze({
2511
+ ...rest,
2512
+ inputPerMTok: sheet.inputPerMTok,
2513
+ outputPerMTok: sheet.outputPerMTok,
2514
+ ...(sheet.tiers !== undefined ? { tiers: sheet.tiers } : {}),
2515
+ });
2516
+ rows.set(tier, row);
2517
+ }
2518
+ return row;
2067
2519
  }
2068
2520
 
2069
2521
  const pricingMemo = new WeakMap<
@@ -2084,7 +2536,8 @@ function pricedAt(model: ModelPricing, on: Date): ModelPricing {
2084
2536
  let periods = periodMemo.get(model);
2085
2537
  if (periods === undefined) {
2086
2538
  periods = history.map((period) => {
2087
- const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, ...rest } = model;
2539
+ // A period's own price only: the service tiers below are today's tables.
2540
+ const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, serviceTiers: _serviceTiers, ...rest } = model;
2088
2541
  const row: ModelPricing = Object.freeze({
2089
2542
  ...rest,
2090
2543
  inputPerMTok: period.inputPerMTok,
@@ -2181,6 +2634,18 @@ export function effectivePricing(
2181
2634
  const period = pricedAt(model, on);
2182
2635
  if (period !== model) return effectivePricing(period, on, context);
2183
2636
  }
2637
+ /*
2638
+ * A response that names another service tier is priced by that tier's own
2639
+ * table, long band included, and by nothing composed out of two.
2640
+ */
2641
+ if (context.serviceTier !== undefined && model.serviceTiers !== undefined) {
2642
+ const named = serviceTierOf(context.serviceTier);
2643
+ const row = named === undefined ? undefined : serviceTierRow(model, named);
2644
+ if (row !== undefined) {
2645
+ const { serviceTier: _serviceTier, ...size } = context;
2646
+ return { ...effectivePricing(row, on, size), serviceTier: named };
2647
+ }
2648
+ }
2184
2649
  const memo = pricingMemo.get(model) ?? cache(model);
2185
2650
 
2186
2651
  /**