@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/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).';
@@ -257,6 +270,9 @@ 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.';
@@ -432,6 +448,51 @@ export const MODELS = [
432
448
  tier: 'opus',
433
449
  notes: '200K context window and a 64K output cap.',
434
450
  },
451
+ {
452
+ /*
453
+ * Retired on the Claude API on 2026-08-05 (the deprecations page, read
454
+ * 2026-11-05) and still served on Claude Models on Google Cloud (Google
455
+ * managed) and on Bedrock. Priced because a year of logs carries it and,
456
+ * unpriced, every one of those calls read as a gap. The pricing page, read
457
+ * the same day: $15/$75, writes $18.75 and $30, cache hits $1.50, batch
458
+ * $7.50/$37.50, which are the default multipliers. Its model page is gone;
459
+ * the context-windows page gives every Claude model outside its 1M list a
460
+ * 200k window, and the prompt-caching page a 1,024-token cache minimum.
461
+ * The dated id is the one the deprecations page names and what a log
462
+ * carries; no shorter alias is on any page read, so none is declared.
463
+ */
464
+ id: 'claude-opus-4-1-20250805',
465
+ provider: 'anthropic',
466
+ displayName: 'Claude Opus 4.1',
467
+ recommendable: false,
468
+ inputPerMTok: 15,
469
+ outputPerMTok: 75,
470
+ contextWindow: 200_000,
471
+ cacheMinTokens: 1024,
472
+ capability: 'large',
473
+ tier: 'opus',
474
+ notes: 'Retired on the Claude API on 2026-08-05; still served on Google Cloud (Google managed) and Bedrock.'
475
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
476
+ },
477
+ {
478
+ /*
479
+ * Retired on the Claude API on 2026-06-15; still served on Claude Models
480
+ * on Google Cloud (Google managed). The same pages and the same figures as
481
+ * Opus 4.1 above: $15/$75, 200k context, a 1,024-token cache minimum.
482
+ */
483
+ id: 'claude-opus-4-20250514',
484
+ provider: 'anthropic',
485
+ displayName: 'Claude Opus 4',
486
+ recommendable: false,
487
+ inputPerMTok: 15,
488
+ outputPerMTok: 75,
489
+ contextWindow: 200_000,
490
+ cacheMinTokens: 1024,
491
+ capability: 'large',
492
+ tier: 'opus',
493
+ notes: 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed).'
494
+ + ' Anthropic names Claude Opus 4.8 as the replacement.',
495
+ },
435
496
  {
436
497
  /*
437
498
  * The pricing page and the Sonnet 5.5 model page, read 2026-10-10. The
@@ -497,6 +558,29 @@ export const MODELS = [
497
558
  tier: 'sonnet',
498
559
  notes: 'Deprecated; retires 2026-11-30. Recommended replacement: Claude Sonnet 5.5.',
499
560
  },
561
+ {
562
+ /*
563
+ * Retired on the Claude API on 2026-06-15 (the deprecations page, read
564
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
565
+ * managed) and on Bedrock. The pricing page, read the same day: $3/$15,
566
+ * writes $3.75 and $6, cache hits $0.30, batch $1.50/$7.50. 200k context
567
+ * from the context-windows page, a 1,024-token cache minimum from the
568
+ * prompt-caching page. The pricing page publishes one rate for it and no
569
+ * long-context band, so every call is priced at that rate.
570
+ */
571
+ id: 'claude-sonnet-4-20250514',
572
+ provider: 'anthropic',
573
+ displayName: 'Claude Sonnet 4',
574
+ recommendable: false,
575
+ inputPerMTok: 3,
576
+ outputPerMTok: 15,
577
+ contextWindow: 200_000,
578
+ cacheMinTokens: 1024,
579
+ capability: 'mid',
580
+ tier: 'sonnet',
581
+ notes: 'Retired on the Claude API on 2026-06-15; still served on Google Cloud (Google managed) and Bedrock.'
582
+ + ' Anthropic names Claude Sonnet 5.5 as the replacement.',
583
+ },
500
584
  {
501
585
  /*
502
586
  * The pricing page, the models overview and the prompt-caching page, read
@@ -551,6 +635,28 @@ export const MODELS = [
551
635
  tier: 'haiku',
552
636
  notes: '200K context window and a 64K output cap, smaller than the rest.',
553
637
  },
638
+ {
639
+ /*
640
+ * Retired on the Claude API on 2026-02-19 (the deprecations page, read
641
+ * 2026-11-05); still served on Claude Models on Google Cloud (Google
642
+ * managed). The pricing page, read the same day: $0.80/$4, writes $1 and
643
+ * $1.60, cache hits $0.08, batch $0.40/$2, the default multipliers. 200k
644
+ * context from the context-windows page, and the prompt-caching page's
645
+ * 2,048-token cache minimum.
646
+ */
647
+ id: 'claude-3-5-haiku-20241022',
648
+ provider: 'anthropic',
649
+ displayName: 'Claude Haiku 3.5',
650
+ recommendable: false,
651
+ inputPerMTok: 0.8,
652
+ outputPerMTok: 4,
653
+ contextWindow: 200_000,
654
+ cacheMinTokens: 2048,
655
+ capability: 'small',
656
+ tier: 'haiku',
657
+ notes: 'Retired on the Claude API on 2026-02-19; still served on Google Cloud (Google managed).'
658
+ + ' Anthropic names Claude Haiku 4.5 as the replacement.',
659
+ },
554
660
  // ------------------------------------------------------------------------
555
661
  // OpenAI
556
662
  //
@@ -610,6 +716,11 @@ export const MODELS = [
610
716
  outputPerMTok: 75,
611
717
  },
612
718
  ],
719
+ serviceTiers: {
720
+ flex: { inputPerMTok: 5, outputPerMTok: 25, tiers: openAiLongBand(10, 37.5) },
721
+ fast: { inputPerMTok: 20, outputPerMTok: 100, tiers: openAiLongBand(40, 150) },
722
+ ultrafast: { inputPerMTok: 60, outputPerMTok: 300, tiers: openAiLongBand(120, 450) },
723
+ },
613
724
  capability: 'frontier',
614
725
  tier: 'frontier',
615
726
  notes: OPENAI_CURRENT_NOTE,
@@ -633,6 +744,11 @@ export const MODELS = [
633
744
  outputPerMTok: 15,
634
745
  },
635
746
  ],
747
+ serviceTiers: {
748
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
749
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
750
+ ultrafast: { inputPerMTok: 12, outputPerMTok: 60, tiers: openAiLongBand(24, 90) },
751
+ },
636
752
  capability: 'large',
637
753
  tier: 'opus',
638
754
  notes: OPENAI_CURRENT_NOTE,
@@ -662,6 +778,10 @@ export const MODELS = [
662
778
  outputPerMTok: 7.5,
663
779
  },
664
780
  ],
781
+ serviceTiers: {
782
+ flex: { inputPerMTok: 0.5, outputPerMTok: 2.5, tiers: openAiLongBand(1, 3.75) },
783
+ fast: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
784
+ },
665
785
  capability: 'mid',
666
786
  tier: 'sonnet',
667
787
  notes: OPENAI_CURRENT_NOTE,
@@ -685,6 +805,10 @@ export const MODELS = [
685
805
  outputPerMTok: 15,
686
806
  },
687
807
  ],
808
+ serviceTiers: {
809
+ flex: { inputPerMTok: 1, outputPerMTok: 5, tiers: openAiLongBand(2, 7.5) },
810
+ fast: { inputPerMTok: 4, outputPerMTok: 20, tiers: openAiLongBand(8, 30) },
811
+ },
688
812
  capability: 'large',
689
813
  tier: 'opus',
690
814
  notes: OPENAI_CURRENT_NOTE,
@@ -708,12 +832,30 @@ export const MODELS = [
708
832
  outputPerMTok: 0.75,
709
833
  },
710
834
  ],
835
+ serviceTiers: {
836
+ flex: { inputPerMTok: 0.05, outputPerMTok: 0.25, tiers: openAiLongBand(0.1, 0.375) },
837
+ fast: { inputPerMTok: 0.2, outputPerMTok: 1, tiers: openAiLongBand(0.4, 1.5) },
838
+ },
711
839
  capability: 'small',
712
840
  tier: 'haiku',
713
841
  notes: OPENAI_CURRENT_NOTE,
714
842
  },
715
843
  {
844
+ /*
845
+ * Its model page, read 2026-11-05: "The `gpt-5.6` alias routes requests
846
+ * to GPT-5.6 Sol", so a log that names the alias is this model; and
847
+ * "$4 per million input tokens and $20 per million output tokens, a 20%
848
+ * reduction in input pricing and a 33% reduction in output pricing ...
849
+ * promotional pricing is available until November 30, 2026". The price
850
+ * after the promotion is not published as a figure, and deriving one
851
+ * from two rounded percentages would be a price nobody quoted, so the row
852
+ * carries what the page charges today. When the page changes,
853
+ * `scripts/price-drift.mjs` reports it on the next weekly run; until a
854
+ * person reads the new figure, a call after 2026-11-30 may be priced low,
855
+ * and the note says so to anyone reading the row.
856
+ */
716
857
  id: 'gpt-5.6-sol',
858
+ aliases: ['gpt-5.6'],
717
859
  provider: 'openai',
718
860
  displayName: 'GPT-5.6 Sol',
719
861
  inputPerMTok: 4,
@@ -731,9 +873,14 @@ export const MODELS = [
731
873
  outputPerMTok: 30,
732
874
  },
733
875
  ],
876
+ serviceTiers: {
877
+ flex: { inputPerMTok: 2, outputPerMTok: 10, tiers: openAiLongBand(4, 15) },
878
+ fast: { inputPerMTok: 8, outputPerMTok: 40, tiers: openAiLongBand(16, 60) },
879
+ },
734
880
  capability: 'large',
735
881
  tier: 'opus',
736
- notes: OPENAI_CURRENT_NOTE,
882
+ notes: OPENAI_CURRENT_NOTE
883
+ + ' $4/$20 is promotional until 2026-11-30, and OpenAI has not published the price after it as a figure.',
737
884
  },
738
885
  {
739
886
  id: 'gpt-5.6-terra',
@@ -754,6 +901,10 @@ export const MODELS = [
754
901
  outputPerMTok: 18,
755
902
  },
756
903
  ],
904
+ serviceTiers: {
905
+ flex: { inputPerMTok: 1, outputPerMTok: 6, tiers: openAiLongBand(2, 9) },
906
+ fast: { inputPerMTok: 4, outputPerMTok: 24, tiers: openAiLongBand(8, 36) },
907
+ },
757
908
  capability: 'mid',
758
909
  tier: 'sonnet',
759
910
  notes: OPENAI_CURRENT_NOTE,
@@ -777,55 +928,109 @@ export const MODELS = [
777
928
  outputPerMTok: 1.8,
778
929
  },
779
930
  ],
931
+ serviceTiers: {
932
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.6, tiers: openAiLongBand(0.2, 0.9) },
933
+ fast: { inputPerMTok: 0.4, outputPerMTok: 2.4, tiers: openAiLongBand(0.8, 3.6) },
934
+ },
780
935
  capability: 'small',
781
936
  tier: 'haiku',
782
937
  notes: OPENAI_CURRENT_NOTE,
783
938
  },
939
+ {
940
+ /*
941
+ * The pricing page's "Cyber models" table and the model's own page
942
+ * (developers.openai.com/api/docs/models/gpt-5.6-cyber.md), read
943
+ * 2026-11-05: $12.50/$75, cached input $1.25, the 30-minute write $15.625
944
+ * and the 12-hour one $31.25, the same 0.1x, 1.25x and 2.5x as the rest of
945
+ * GPT-5.6. A 400,000-token context window with at most 272,000 input
946
+ * tokens, so the long-prompt band the model page mentions can never be
947
+ * reached and none is modelled; the pricing page lists none either. The
948
+ * Batch endpoint is "Not supported", so there is no batch price.
949
+ *
950
+ * Found by `scripts/price-drift.mjs` on its first run, as a model on the
951
+ * pricing page that read as unpriced. It needs Daybreak programme
952
+ * approval, so it is priced for the logs of people who have it and never
953
+ * recommended to anyone else.
954
+ */
955
+ id: 'gpt-5.6-cyber',
956
+ provider: 'openai',
957
+ displayName: 'GPT-5.6 Cyber',
958
+ recommendable: false,
959
+ inputPerMTok: 12.5,
960
+ outputPerMTok: 75,
961
+ contextWindow: 400_000,
962
+ cacheMinTokens: 1024,
963
+ caching: 'automatic',
964
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1.25, cacheWrite1h: 2.5, batch: null },
965
+ cacheTtl: OPENAI_CACHE_TTL,
966
+ capability: 'large',
967
+ tier: 'opus',
968
+ notes: 'Needs approval through the Daybreak programme. At most 272,000 input tokens, Responses API only, no batch.'
969
+ + ' Cache writes cost 1.25x input for the default 30-minute entry and 2.5x for a 12-hour one.',
970
+ },
784
971
  // ------------------------------------------------------------------------
972
+ // GPT-5, mini and nano. Each model page, read 2026-11-05, opens with
973
+ // "Status: Deprecated", and the catalogue still offered all three as
974
+ // cheaper alternatives, nano as the cheapest OpenAI model there is. Priced
975
+ // for the logs that hold them; never recommended.
785
976
  {
786
977
  id: 'gpt-5',
787
978
  aliases: ['gpt-5-2025-08-07'],
788
979
  provider: 'openai',
789
980
  displayName: 'GPT-5',
981
+ recommendable: false,
790
982
  inputPerMTok: 1.25,
791
983
  outputPerMTok: 10,
792
984
  contextWindow: 400_000,
793
985
  cacheMinTokens: 1024,
794
986
  caching: 'automatic',
795
987
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
988
+ serviceTiers: {
989
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
990
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
991
+ },
796
992
  capability: 'frontier',
797
993
  tier: 'frontier',
798
- notes: OPENAI_GPT5_NOTE,
994
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
799
995
  },
800
996
  {
801
997
  id: 'gpt-5-mini',
802
998
  aliases: ['gpt-5-mini-2025-08-07'],
803
999
  provider: 'openai',
804
1000
  displayName: 'GPT-5 mini',
1001
+ recommendable: false,
805
1002
  inputPerMTok: 0.25,
806
1003
  outputPerMTok: 2,
807
1004
  contextWindow: 400_000,
808
1005
  cacheMinTokens: 1024,
809
1006
  caching: 'automatic',
810
1007
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1008
+ serviceTiers: {
1009
+ flex: { inputPerMTok: 0.125, outputPerMTok: 1 },
1010
+ fast: { inputPerMTok: 0.45, outputPerMTok: 3.6 },
1011
+ },
811
1012
  capability: 'mid',
812
1013
  tier: 'sonnet',
813
- notes: OPENAI_GPT5_NOTE,
1014
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
814
1015
  },
815
1016
  {
816
1017
  id: 'gpt-5-nano',
817
1018
  aliases: ['gpt-5-nano-2025-08-07'],
818
1019
  provider: 'openai',
819
1020
  displayName: 'GPT-5 nano',
1021
+ recommendable: false,
820
1022
  inputPerMTok: 0.05,
821
1023
  outputPerMTok: 0.4,
822
1024
  contextWindow: 400_000,
823
1025
  cacheMinTokens: 1024,
824
1026
  caching: 'automatic',
825
1027
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1028
+ serviceTiers: {
1029
+ flex: { inputPerMTok: 0.025, outputPerMTok: 0.2 },
1030
+ },
826
1031
  capability: 'small',
827
1032
  tier: 'haiku',
828
- notes: OPENAI_GPT5_NOTE,
1033
+ notes: OPENAI_GPT5_NOTE + OPENAI_DEPRECATED_SUFFIX,
829
1034
  },
830
1035
  // ------------------------------------------------------------------------
831
1036
  // OpenAI's earlier models, still billed and still in logs.
@@ -862,6 +1067,10 @@ export const MODELS = [
862
1067
  outputPerMTok: 45,
863
1068
  },
864
1069
  ],
1070
+ serviceTiers: {
1071
+ flex: { inputPerMTok: 2.5, outputPerMTok: 15, tiers: openAiLongBand(5, 22.5) },
1072
+ fast: { inputPerMTok: 12.5, outputPerMTok: 75 },
1073
+ },
865
1074
  capability: 'frontier',
866
1075
  tier: 'frontier',
867
1076
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -885,6 +1094,9 @@ export const MODELS = [
885
1094
  outputPerMTok: 270,
886
1095
  },
887
1096
  ],
1097
+ serviceTiers: {
1098
+ flex: { inputPerMTok: 15, outputPerMTok: 90 },
1099
+ },
888
1100
  capability: 'frontier',
889
1101
  tier: 'frontier',
890
1102
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -908,6 +1120,10 @@ export const MODELS = [
908
1120
  outputPerMTok: 22.5,
909
1121
  },
910
1122
  ],
1123
+ serviceTiers: {
1124
+ flex: { inputPerMTok: 1.25, outputPerMTok: 7.5, tiers: openAiLongBand(2.5, 11.25) },
1125
+ fast: { inputPerMTok: 5, outputPerMTok: 30 },
1126
+ },
911
1127
  capability: 'frontier',
912
1128
  tier: 'frontier',
913
1129
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -931,6 +1147,9 @@ export const MODELS = [
931
1147
  outputPerMTok: 270,
932
1148
  },
933
1149
  ],
1150
+ serviceTiers: {
1151
+ flex: { inputPerMTok: 15, outputPerMTok: 90, tiers: openAiLongBand(30, 135) },
1152
+ },
934
1153
  capability: 'frontier',
935
1154
  tier: 'frontier',
936
1155
  notes: OPENAI_EARLIER_LONG_NOTE,
@@ -946,6 +1165,10 @@ export const MODELS = [
946
1165
  cacheMinTokens: 1024,
947
1166
  caching: 'automatic',
948
1167
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1168
+ serviceTiers: {
1169
+ flex: { inputPerMTok: 0.375, outputPerMTok: 2.25 },
1170
+ fast: { inputPerMTok: 1.5, outputPerMTok: 9 },
1171
+ },
949
1172
  capability: 'mid',
950
1173
  tier: 'sonnet',
951
1174
  notes: OPENAI_EARLIER_NOTE,
@@ -962,6 +1185,9 @@ export const MODELS = [
962
1185
  cacheMinTokens: 1024,
963
1186
  caching: 'automatic',
964
1187
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1188
+ serviceTiers: {
1189
+ flex: { inputPerMTok: 0.1, outputPerMTok: 0.625 },
1190
+ },
965
1191
  capability: 'small',
966
1192
  tier: 'haiku',
967
1193
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -977,6 +1203,10 @@ export const MODELS = [
977
1203
  cacheMinTokens: 1024,
978
1204
  caching: 'automatic',
979
1205
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1206
+ serviceTiers: {
1207
+ flex: { inputPerMTok: 0.875, outputPerMTok: 7 },
1208
+ fast: { inputPerMTok: 3.5, outputPerMTok: 28 },
1209
+ },
980
1210
  capability: 'frontier',
981
1211
  tier: 'frontier',
982
1212
  notes: OPENAI_EARLIER_NOTE,
@@ -1008,6 +1238,10 @@ export const MODELS = [
1008
1238
  cacheMinTokens: 1024,
1009
1239
  caching: 'automatic',
1010
1240
  multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1241
+ serviceTiers: {
1242
+ flex: { inputPerMTok: 0.625, outputPerMTok: 5 },
1243
+ fast: { inputPerMTok: 2.5, outputPerMTok: 20 },
1244
+ },
1011
1245
  capability: 'frontier',
1012
1246
  tier: 'frontier',
1013
1247
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1040,6 +1274,9 @@ export const MODELS = [
1040
1274
  cacheMinTokens: 1024,
1041
1275
  caching: 'automatic',
1042
1276
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1277
+ serviceTiers: {
1278
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1279
+ },
1043
1280
  capability: 'large',
1044
1281
  tier: 'opus',
1045
1282
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1056,6 +1293,9 @@ export const MODELS = [
1056
1293
  cacheMinTokens: 1024,
1057
1294
  caching: 'automatic',
1058
1295
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1296
+ serviceTiers: {
1297
+ fast: { inputPerMTok: 0.7, outputPerMTok: 2.8 },
1298
+ },
1059
1299
  capability: 'mid',
1060
1300
  tier: 'sonnet',
1061
1301
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1072,6 +1312,9 @@ export const MODELS = [
1072
1312
  cacheMinTokens: 1024,
1073
1313
  caching: 'automatic',
1074
1314
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1315
+ serviceTiers: {
1316
+ fast: { inputPerMTok: 4.25, outputPerMTok: 17 },
1317
+ },
1075
1318
  capability: 'large',
1076
1319
  tier: 'opus',
1077
1320
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1088,6 +1331,9 @@ export const MODELS = [
1088
1331
  cacheMinTokens: 1024,
1089
1332
  caching: 'automatic',
1090
1333
  multipliers: { cacheRead: 0.5, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1334
+ serviceTiers: {
1335
+ fast: { inputPerMTok: 0.25, outputPerMTok: 1 },
1336
+ },
1091
1337
  capability: 'small',
1092
1338
  tier: 'haiku',
1093
1339
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1104,6 +1350,10 @@ export const MODELS = [
1104
1350
  cacheMinTokens: 1024,
1105
1351
  caching: 'automatic',
1106
1352
  multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: 0.5 },
1353
+ serviceTiers: {
1354
+ flex: { inputPerMTok: 1, outputPerMTok: 4 },
1355
+ fast: { inputPerMTok: 3.5, outputPerMTok: 14 },
1356
+ },
1107
1357
  capability: 'frontier',
1108
1358
  tier: 'frontier',
1109
1359
  notes: OPENAI_EARLIER_NOTE + OPENAI_DEPRECATED_SUFFIX,
@@ -1125,6 +1375,143 @@ export const MODELS = [
1125
1375
  notes: OPENAI_EARLIER_NO_CACHE_NOTE + OPENAI_DEPRECATED_SUFFIX,
1126
1376
  },
1127
1377
  // ------------------------------------------------------------------------
1378
+ // The Codex models and chat-latest. Codex CLI sessions name these, and none
1379
+ // was priced, so every such call read as unpriced. Each from its own model
1380
+ // page (developers.openai.com/api/docs/models/<id>.md), read 2026-11-05,
1381
+ // and gpt-5.3-codex and chat-latest also from the pricing page's
1382
+ // "Specialized models" table. Every page says Batch is "Not supported", so
1383
+ // none has a batch price. gpt-5.3-codex is deprecated and the rest of the
1384
+ // Codex line is shut down; all are priced for the logs that hold them and
1385
+ // never recommended. chat-latest is served and not recommended either: it
1386
+ // points at whatever model ChatGPT uses, so a recommendation would not name
1387
+ // a fixed model. Cache reads are 0.1x everywhere but codex-mini-latest,
1388
+ // whose page gives $0.375 on $1.50, 0.25x. gpt-5-search-api is on the
1389
+ // pricing page too and its model page returns 404, so it has no context
1390
+ // window and is left unpriced (see scripts/price-drift.mjs).
1391
+ // ------------------------------------------------------------------------
1392
+ {
1393
+ id: 'gpt-5.3-codex',
1394
+ provider: 'openai',
1395
+ displayName: 'GPT-5.3-Codex',
1396
+ recommendable: false,
1397
+ inputPerMTok: 1.75,
1398
+ outputPerMTok: 14,
1399
+ contextWindow: 400_000,
1400
+ cacheMinTokens: 1024,
1401
+ caching: 'automatic',
1402
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1403
+ // The pricing page's specialized "Fast" table: $3.50/$28, cached $0.35.
1404
+ serviceTiers: { fast: { inputPerMTok: 3.5, outputPerMTok: 28 } },
1405
+ capability: 'frontier',
1406
+ tier: 'frontier',
1407
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_DEPRECATED_SUFFIX,
1408
+ },
1409
+ {
1410
+ id: 'gpt-5.2-codex',
1411
+ provider: 'openai',
1412
+ displayName: 'GPT-5.2-Codex',
1413
+ recommendable: false,
1414
+ inputPerMTok: 1.75,
1415
+ outputPerMTok: 14,
1416
+ contextWindow: 400_000,
1417
+ cacheMinTokens: 1024,
1418
+ caching: 'automatic',
1419
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1420
+ capability: 'frontier',
1421
+ tier: 'frontier',
1422
+ notes: OPENAI_CODEX_NOTE + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1423
+ },
1424
+ {
1425
+ id: 'gpt-5.1-codex',
1426
+ provider: 'openai',
1427
+ displayName: 'GPT-5.1-Codex',
1428
+ recommendable: false,
1429
+ inputPerMTok: 1.25,
1430
+ outputPerMTok: 10,
1431
+ contextWindow: 400_000,
1432
+ cacheMinTokens: 1024,
1433
+ caching: 'automatic',
1434
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1435
+ capability: 'frontier',
1436
+ tier: 'frontier',
1437
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1438
+ },
1439
+ {
1440
+ id: 'gpt-5.1-codex-max',
1441
+ provider: 'openai',
1442
+ displayName: 'GPT-5.1-Codex-Max',
1443
+ recommendable: false,
1444
+ inputPerMTok: 1.25,
1445
+ outputPerMTok: 10,
1446
+ contextWindow: 400_000,
1447
+ cacheMinTokens: 1024,
1448
+ caching: 'automatic',
1449
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1450
+ capability: 'frontier',
1451
+ tier: 'frontier',
1452
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1453
+ },
1454
+ {
1455
+ id: 'gpt-5.1-codex-mini',
1456
+ provider: 'openai',
1457
+ displayName: 'GPT-5.1-Codex Mini',
1458
+ recommendable: false,
1459
+ inputPerMTok: 0.25,
1460
+ outputPerMTok: 2,
1461
+ contextWindow: 400_000,
1462
+ cacheMinTokens: 1024,
1463
+ caching: 'automatic',
1464
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1465
+ capability: 'mid',
1466
+ tier: 'sonnet',
1467
+ notes: OPENAI_CODEX_NOTE + OPENAI_SHUT_DOWN_SUFFIX,
1468
+ },
1469
+ {
1470
+ id: 'gpt-5-codex',
1471
+ provider: 'openai',
1472
+ displayName: 'GPT-5-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 + ' At most 272,000 input tokens.' + OPENAI_SHUT_DOWN_SUFFIX,
1483
+ },
1484
+ {
1485
+ id: 'codex-mini-latest',
1486
+ provider: 'openai',
1487
+ displayName: 'codex-mini-latest',
1488
+ recommendable: false,
1489
+ inputPerMTok: 1.5,
1490
+ outputPerMTok: 6,
1491
+ contextWindow: 200_000,
1492
+ cacheMinTokens: 1024,
1493
+ caching: 'automatic',
1494
+ multipliers: { cacheRead: 0.25, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1495
+ capability: 'mid',
1496
+ tier: 'sonnet',
1497
+ notes: OPENAI_CODEX_NOTE + ' A cache read costs 0.25x input.' + OPENAI_SHUT_DOWN_SUFFIX,
1498
+ },
1499
+ {
1500
+ id: 'chat-latest',
1501
+ provider: 'openai',
1502
+ displayName: 'Chat Latest',
1503
+ recommendable: false,
1504
+ inputPerMTok: 5,
1505
+ outputPerMTok: 30,
1506
+ contextWindow: 400_000,
1507
+ cacheMinTokens: 1024,
1508
+ caching: 'automatic',
1509
+ multipliers: { cacheRead: 0.1, cacheWrite5m: 1, cacheWrite1h: 1, batch: null },
1510
+ capability: 'large',
1511
+ tier: 'opus',
1512
+ 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.',
1513
+ },
1514
+ // ------------------------------------------------------------------------
1128
1515
  // Google
1129
1516
  // ------------------------------------------------------------------------
1130
1517
  {
@@ -1905,6 +2292,57 @@ export function getModel(id) {
1905
2292
  export function listModels() {
1906
2293
  return [...MODELS];
1907
2294
  }
2295
+ /**
2296
+ * A response's `service_tier` as the tier whose table prices it, or
2297
+ * `undefined` for the standard one.
2298
+ *
2299
+ * OpenAI renamed Priority processing to Fast mode on 2026-07-30, and its guide
2300
+ * says "for GPT-5.6 and earlier models, the response returns `priority`
2301
+ * whether the request specifies `priority` or `fast`", so both spellings are
2302
+ * the Fast table. `"default"` is the standard tier, and is also what a Fast
2303
+ * request downgraded under the ramp-rate limit returns, "and charge standard
2304
+ * rates". `"auto"` and `"scale"`, or a value this table has never seen, are
2305
+ * priced at the standard rate rather than guessed into another tier.
2306
+ */
2307
+ export function serviceTierOf(value) {
2308
+ if (typeof value !== 'string')
2309
+ return undefined;
2310
+ const tier = value.trim().toLowerCase();
2311
+ if (tier === 'flex')
2312
+ return 'flex';
2313
+ if (tier === 'fast' || tier === 'priority')
2314
+ return 'fast';
2315
+ if (tier === 'ultrafast')
2316
+ return 'ultrafast';
2317
+ return undefined;
2318
+ }
2319
+ /**
2320
+ * The row as one service tier prices it: that tier's rates and long band, the
2321
+ * row's own cache and batch ratios. Built once per row and tier.
2322
+ */
2323
+ const serviceTierMemo = new WeakMap();
2324
+ function serviceTierRow(model, tier) {
2325
+ const sheet = model.serviceTiers?.[tier];
2326
+ if (sheet === undefined)
2327
+ return undefined;
2328
+ let rows = serviceTierMemo.get(model);
2329
+ if (rows === undefined) {
2330
+ rows = new Map();
2331
+ serviceTierMemo.set(model, rows);
2332
+ }
2333
+ let row = rows.get(tier);
2334
+ if (row === undefined) {
2335
+ const { history: _history, promo: _promo, tiers: _tiers, serviceTiers: _serviceTiers, ...rest } = model;
2336
+ row = Object.freeze({
2337
+ ...rest,
2338
+ inputPerMTok: sheet.inputPerMTok,
2339
+ outputPerMTok: sheet.outputPerMTok,
2340
+ ...(sheet.tiers !== undefined ? { tiers: sheet.tiers } : {}),
2341
+ });
2342
+ rows.set(tier, row);
2343
+ }
2344
+ return row;
2345
+ }
1908
2346
  const pricingMemo = new WeakMap();
1909
2347
  /**
1910
2348
  * The row as it stood at `on`: itself when the current price applies, or a
@@ -1919,7 +2357,8 @@ function pricedAt(model, on) {
1919
2357
  let periods = periodMemo.get(model);
1920
2358
  if (periods === undefined) {
1921
2359
  periods = history.map((period) => {
1922
- const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, ...rest } = model;
2360
+ // A period's own price only: the service tiers below are today's tables.
2361
+ const { history: _history, promo: _promo, tiers: _tiers, multipliers: _multipliers, serviceTiers: _serviceTiers, ...rest } = model;
1923
2362
  const row = Object.freeze({
1924
2363
  ...rest,
1925
2364
  inputPerMTok: period.inputPerMTok,
@@ -2011,6 +2450,18 @@ export function effectivePricing(model, on = new Date(), context = {}) {
2011
2450
  if (period !== model)
2012
2451
  return effectivePricing(period, on, context);
2013
2452
  }
2453
+ /*
2454
+ * A response that names another service tier is priced by that tier's own
2455
+ * table, long band included, and by nothing composed out of two.
2456
+ */
2457
+ if (context.serviceTier !== undefined && model.serviceTiers !== undefined) {
2458
+ const named = serviceTierOf(context.serviceTier);
2459
+ const row = named === undefined ? undefined : serviceTierRow(model, named);
2460
+ if (row !== undefined) {
2461
+ const { serviceTier: _serviceTier, ...size } = context;
2462
+ return { ...effectivePricing(row, on, size), serviceTier: named };
2463
+ }
2464
+ }
2014
2465
  const memo = pricingMemo.get(model) ?? cache(model);
2015
2466
  /**
2016
2467
  * Tiers first, and a promotion never runs on top of one.