fix(pricing): add azure gpt-5.6 cache write rates and correct data zone priority

[mateo-berri]

TLDR

Problem this solves:

• Azure returns gpt-5.6 cache writes as cache_write_tokens, and bills them at 1.25x input
• Our 12 azure gpt-5.6 entries carry no cache write rate, so those tokens are billed at $0
• azure/us and azure/eu priority rates sit at 1.25x Global; Azure's price page says 1.1x

How it solves it:

• Adds cache_creation_input_token_cost (1.25x input) to every tier variant of all 12 entries
• Corrects us/eu priority input, output, and cache read rates to 1.1x Global
• Adds the missing us/eu above-272k priority rates, also at 1.1x Global
• Regression tests pin both invariants against the Azure price page

User Flow

Before: a developer running an agent workload on an azure gpt-5.6-luna deployment sees gateway spend far below their Azure bill, because every prompt cache write is billed at $0

1. They send POST https://litellm-domain/v1/chat/completions with {"model": "azure-gpt-5.6-luna", ...} and a long system prompt
2. The response usage shows prompt_tokens_details.cache_write_tokens: 1313, which Azure bills at $0.25 per million tokens
3. The x-litellm-response-cost-input header reads $0.0000006: only the 3 non-cached tokens were priced, the 1313 cache-write tokens cost $0
4. An admin who pinned the Data Zone entry with priority processing sees the opposite error: priority tokens priced at 1.25x Global where Azure bills 1.1x

After: the same requests price cache writes and Data Zone priority exactly as the Azure price page does

1. They send the same POST https://litellm-domain/v1/chat/completions with {"model": "azure-gpt-5.6-luna", ...} and the same long system prompt
2. The response usage again shows prompt_tokens_details.cache_write_tokens: 1313
3. The x-litellm-response-cost-input header now reads about $0.000329: 3 tokens at $0.20/M plus 1313 cache-write tokens at $0.25/M
4. Data Zone priority requests are priced at 1.1x the Global rate, matching Azure's page

Relevant issues

Fixes #37631
Fixes #37268

Linear ticket

Resolves LIT-6178

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Screenshots / Proof of Fix

Live proxy A/B against a real Azure Foundry gpt-5.6-luna deployment, both legs booted with --num_workers 2 and LITELLM_LOCAL_MODEL_COST_MAP=True: before at the merge base 95285c3, after at this PR's tip b349b9b. Five cases per leg, identical flow: a ~1300-token system prompt led by a fresh random cache-buster forces a cache WRITE on each of /v1/chat/completions, /v1/messages, and /v1/responses, repeating the chat prompt gets the cache READ, and a short prompt is the no-cache control. Azure's published Luna rates: input 2e-07, cache write 2.5e-07 (1.25x input), cache read 2e-08, output 1.2e-06

Exemplar request (same shape on both legs, only the port and buster differ):

curl -sD h.txt -o b.json -H 'Authorization: Bearer sk-lit6178' -H 'Content-Type: application/json' \
  -d @chat_write.json http://localhost:<port>/v1/chat/completions
grep -i x-litellm-response-cost h.txt; jq .usage b.json

Before (95285c3): cache-write tokens billed $0

chat_write came back with "prompt_tokens": 1703, "prompt_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 1700, ...} and

x-litellm-response-cost: 6.6e-06
x-litellm-response-cost-input: 6e-07
x-litellm-response-cost-output: 6e-06

cost-input 6e-07 is exactly the 3 uncached text tokens at 2e-07; the 1700 cache-write tokens contributed $0. The published-rate input side is 3 x 2e-07 + 1700 x 2.5e-07 = 4.256e-04, a ~650x input-side underbill. Same story on every write surface:

case	endpoint	write toks	read toks	cost-input observed	input side at published rates
chat_write	/v1/chat/completions	1700	0	6e-07	4.256e-04
chat_read	/v1/chat/completions	0	1700	6e-07 + 3.4e-05 cache-read	matches (read already priced)
msgs_write	/v1/messages	1696	0	6e-07	4.246e-04
resp_write	/v1/responses	1698	0	6e-07	4.251e-04
chat_short	/v1/chat/completions	0	0	1.8e-06	matches

After (b349b9b): cache writes billed at 2.5e-07, exact

chat_write came back with 1627 cache-write tokens and

x-litellm-response-cost: 0.00041335
x-litellm-response-cost-input: 6.000000000000146e-07
x-litellm-response-cost-cache-creation: 0.00040675
x-litellm-response-cost-output: 6e-06

1627 x 2.5e-07 = 4.0675e-04, exactly the cache-creation amount. Every case matches the published-rate arithmetic:

case	endpoint	write toks	read toks	observed input-side cost	expected
chat_write	/v1/chat/completions	1627	0	6e-07 + 4.0675e-04	exact
chat_read (hit)	/v1/chat/completions	0	1627	6e-07 + 3.254e-05 cache-read	exact (1627 x 2e-08)
msgs_write	/v1/messages	1626	0	6e-07 + 4.065e-04	exact
resp_write	/v1/responses	1627	0	6e-07 + 4.0675e-04	exact
chat_short	/v1/chat/completions	0	0	1.8e-06	exact

The first chat_read attempt (2s after the write) was a server-side cache miss, reported as a second write, and billed as one at 2.5e-07, correctly; the 10s retry hit the cache

Observations from the legs (all pre-existing, this PR leaves them alone):

• Input cost splits across three response-cost headers
• First cache-read attempt after 2s missed; 10s retry hit
• /v1/messages usage keeps cache writes outside input_tokens
• Chat usage reports both cache_write_tokens and cache_creation_tokens
• Azure Luna returns a latency_checkpoint block in usage

Commits since this run: 8d750a2 only regenerates model_prices_and_context_window.schema.json (a generated file consumed by CI validation alone), e97a84a only adds the us/eu above-272k priority rates plus a stricter key-set test, and c5e3b21 only adds that key name to the test-side schema. The QA cases are all sub-272k standard-tier requests, so none of these commits change any behavior they exercise and the proof above stands; the above-272k priority path is pinned by the regression tests against the Azure price page.

Type

🐛 Bug Fix

Caveats (if any)

Low

• Global long-context priority rates predate this PR and have no public page value; cache writes there follow the same 1.25x rule
• Dated snapshot entries (e.g. gpt-5.6-luna-2026-07-09) and azure_ai gpt-5.6 entries are still missing; tracked separately

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

• b349b9b passes /live-pr-risk

• 8d750a2 passes /live-pr-risk

• e97a84a passes /live-pr-risk

• c5e3b21 passes /live-pr-risk

---

Note

Medium Risk
Changes only static pricing metadata but directly affects reported spend and budgets for Azure GPT-5.6 cache writes and Data Zone priority; incorrect rates would under- or over-bill customers.

Overview
Fixes Azure GPT-5.6 cost map entries so prompt cache-write tokens and US/EU Data Zone tiers match Azure’s published pricing.

All 12 azure/.../gpt-5.6* variants (global, sol, terra, luna × global/us/eu) now define cache_creation_input_token_cost (and matching priority / above-272k suffix keys) at 1.25× the corresponding input rate, so cache_creation_input_tokens / cache-write usage is no longer priced at $0 or at plain input when the calculator reads model_prices_and_context_window.json.

For azure/us/ and azure/eu/ entries, priority (and newly filled above-272k priority) input, output, cache-read, and cache-creation rates are corrected from an erroneous 1.25× Global to 1.1× Global, aligned with the Data Zone uplift on standard tiers.

The JSON schema gains cache_creation_input_token_cost_above_272k_tokens_priority, and regression tests assert cache-write billing for azure/gpt-5.6-luna plus map invariants across all twelve keys.

^{Reviewed by Cursor Bugbot for commit c5e3b21. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

This PR corrects Azure GPT-5.6 cache-write and Data Zone priority pricing metadata.

• Adds cache-creation rates across all twelve Azure GPT-5.6 variants.
• Adds the combined above-272k priority rates that the cost resolver selects for long priority requests.
• Keeps the primary and backup pricing maps synchronized and extends the generated schema.
• Adds regression coverage for cache-write billing and regional/global rate parity.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the previously reported combined cache-rate omission is fixed in both pricing maps and the runtime resolver selects the newly supplied field.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds the complete cache-write pricing ladder and corrects US/EU priority rates; the previously missing combined key is present across all affected entries.
litellm/model_prices_and_context_window_backup.json	Mirrors the corrected Azure GPT-5.6 rates and combined long-context priority keys from the primary pricing map.
model_prices_and_context_window.schema.json	Adds schema support for the combined above-272k priority cache-creation field.
tests/test_litellm/test_cost_calculator.py	Adds cache-write cost and pricing-map invariant tests covering all twelve Azure GPT-5.6 variants.
tests/test_litellm/test_utils.py	Extends the test-side pricing schema to recognize the new combined cache-creation rate key.

_{Reviews (4): Last reviewed commit: "test(utils): add priority cache write ti..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit c5e3b21. Configure here.}

[mateo-berri]

@greptileai

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_azure_gpt_5_6_cache_write_pricing (c5e3b21) with litellm_internal_staging (cdb60af)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (80843ae) during the generation of this report, so cdb60af was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩