fix(cost): bill Azure Foundry GPT-5.6 prompt cache writes

[omertal323]

TLDR

Problem this solves:

• Azure Foundry GPT-5.6 cache write tokens were billed at $0
• Data-zone sol/gpt-5.6 priority writes were 25% above Azure's published cell

How it solves it:

• Adds cache write prices to all 12 azure GPT-5.6 entries at 1.25x input
• Sets data-zone sol/gpt-5.6 priority to Azure's $11 prompt / $13.75 write
• Pins every added rate with cost-calculator regression tests

User Flow

Before: a developer whose Azure GPT-5.6 completion includes prompt-cache write tokens is billed as if those writes were free, so spend is only uncached input plus output

1. They POST http://localhost:4000/spend/calculate with a completion_response for azure/gpt-5.6, 100000 prompt tokens, 80000 cache_write_tokens, and 10 completion tokens
2. The response is {"cost":0.1003} (20000 uncached input x $5/1M + 10 output x $30/1M). The 80000 write tokens add $0
3. The same payload for azure/us/gpt-5.6 with "service_tier":"priority" returns {"cost":0.275825}. Writes add $0
4. The same payload for azure/gpt-5.6-luna returns {"cost":0.004012}. Writes add $0

After: the same payloads bill cache writes at 1.25x input. Data-zone sol/gpt-5.6 priority uses Azure's $11 / $13.75 cells. Luna writes stay on the 7/30 cut ($0.25/1M)

1. They POST the same azure/gpt-5.6 payload. Response is {"cost":0.6003} (80000 writes x $6.25/1M = $0.50 added)
2. They POST the same azure/us/gpt-5.6 priority payload. Response is {"cost":1.32066} (priority input $11/1M, writes $13.75/1M)
3. They POST the same azure/gpt-5.6-luna payload. Response is {"cost":0.024012} (input $0.20/1M, writes $0.25/1M)

Relevant issues

Linear ticket

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)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review)

Screenshots / Proof of Fix

Azure added prompt cache write pricing for the GPT-5.6 series on 2026-07-29. Write is 1.25x the matching input cell. Sources disagree for terra/luna, so this PR does not treat the calculator API as the last word for those two

Sol (unchanged by the 7/30 cut): the calculator API still matches the Foundry blog. Global short write $6.25, data-zone short write $6.875, data-zone priority write $13.75 (language-models-gpt-5-6-sol-short-priority-area-cache-writes). Azure has no gpt-5-6 base offer; azure/gpt-5.6 mirrors sol

Terra and luna: the calculator still shows pre-cut cells (terra input $2.50 / luna $1.00). The Foundry launch blog says the table "reflects the latest discounts announced by OpenAI on 7/30" and lists Standard Global short as terra $2.00 / $2.50 write and luna $0.20 / $0.25 write. OpenAI's live list is the same. A Microsoft Q&A moderator (2026-08-11) and Azure Support (2026-08-10) say the cut applied on Azure from 2026-08-01 and that the public pricing page lags. This PR follows the blog for terra/luna, including 1.25x writes on those cut inputs, and #35481 for the input/output cells

We don't have an Azure GPT-5.6 deployment to burn real dollars on, so the proof uses /spend/calculate with the usage shape Azure returns. After proxy: LITELLM_LOCAL_MODEL_COST_MAP=True python -m litellm.proxy.proxy_cli --config litellm/proxy/dev_config.yaml --port 4001 at bcc2d54. Before was the same payload against a proxy still on the merge-base map (no cache-write fields)

Shared payload (model and optional service_tier vary per case):

curl -s http://localhost:4001/spend/calculate -H 'Authorization: Bearer sk-1234' -H 'Content-Type: application/json' -d '{
  "completion_response": {
    "id": "chatcmpl-proof",
    "object": "chat.completion",
    "created": 1753000000,
    "model": "azure/gpt-5.6",
    "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}],
    "usage": {
      "prompt_tokens": 100000,
      "completion_tokens": 10,
      "total_tokens": 100010,
      "prompt_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 80000}
    }
  }
}'

Before (d542c82)

Global azure/gpt-5.6 cache write

1. POST /spend/calculate with model: azure/gpt-5.6 and the shared usage
2. Observed: {"cost":0.1003}. 80000 write tokens add $0 (20000 uncached input x $5/1M + 10 output x $30/1M only)

Data-zone priority azure/us/gpt-5.6

1. Same payload with model: azure/us/gpt-5.6 and "service_tier":"priority"
2. Observed: {"cost":0.275825}. Writes add $0. Priority input is the stale $13.75/1M (Azure publishes $11)

Luna azure/gpt-5.6-luna

1. Same payload with model: azure/gpt-5.6-luna
2. Observed: {"cost":0.004012}. Writes add $0. Input is already the 7/30 cut ($0.20/1M)

After (bcc2d54)

Global azure/gpt-5.6 cache write

1. Same POST against the branch proxy on port 4001
2. Observed: {"cost":0.6003}. 80000 write tokens add $0.50 at Azure's published $6.25/1M

Data-zone priority azure/us/gpt-5.6

1. Same priority payload against port 4001
2. Observed: {"cost":1.32066}. Priority input $11/1M, writes $13.75/1M (Azure's gpt-5-6-sol-short-priority-area-* cells). An earlier revision of this PR billed those writes at $17.1875 by applying 1.25x to the stale $13.75 input

Luna azure/gpt-5.6-luna

1. Same luna payload against port 4001
2. Observed: {"cost":0.024012}. Input $0.20/1M, writes $0.25/1M (Foundry blog / OpenAI cut). Data-zone luna on the same payload is {"cost":0.0264132} at $0.22 / $0.275

Type

🐛 Bug Fix

Changes

Adds cache_creation_input_token_cost, cache_creation_input_token_cost_above_272k_tokens, and cache_creation_input_token_cost_priority (plus _above_272k_tokens_priority on the four global entries) to all 12 azure/gpt-5.6* entries in model_prices_and_context_window.json and the backup copy

Global and data-zone short/long writes for gpt-5.6 and -sol match Azure's published calculator cells ($6.25 / $12.50 global, $6.875 / $13.75 data zone). Data-zone priority for those two models was wrong on an earlier revision: 1.25x of our stale $13.75 priority input produced $17.1875, while Azure publishes $11 prompt and $13.75 write (language-models-gpt-5-6-sol-short-priority-area-*). This revision also corrects the data-zone priority input, cache-read, and output cells to that same Azure row ($11 / $1.10 / $66)

Luna and terra writes are 1.25x of the cut inputs from #35481, which match the Foundry blog's post-7/30 table (luna $0.25 / $0.275, terra $2.50 / $2.75). The calculator API still lists pre-cut luna $1.25 and terra $3.125 writes. Do not copy those back. Azure publishes no luna priority cells, so priority and long+priority luna rates stay inferred from the cut inputs

cache_creation_input_token_cost_above_272k_tokens_priority on the four global entries is inferred (1.25x of our long+priority input). Azure publishes no long+priority cell for any 5.6 variant. For us/eu entries, which also have no published long-context priority prices, the test pins the fallback to the standard long-context write rate. azure/gpt-5.6 mirrors sol. Azure has no gpt-5-6 base offer

An earlier revision of this PR also declared the cache_creation_input_token_cost_flex/_priority/_above_272k_tokens[_priority] fields in ModelInfoBase, without which get_model_info strips them and cache writes fall back to the standard rate. #35270 landed the same field declarations first, so after rebasing this PR carries only the pricing data, the regenerated prices schema, and the tests

Tests: test_generic_cost_per_token_azure_gpt_5_6_cache_write_tokens pins the write rates for all 12 azure entries across standard, priority, and long-context tiers through generic_cost_per_token. test_azure_data_zone_gpt56_priority_matches_azure_list pins the data-zone sol/gpt-5.6 priority row to Azure's $11 / $13.75 cells so the 1.25x-of-stale-input bug cannot return. test_generic_cost_per_token_azure_gpt56 pins terra/luna to the 7/30 cut. test_generic_cost_per_token_openai_gpt_5_6_cache_write_service_tiers pins the flex/priority plumbing via the OpenAI entry. The new _above_272k_tokens_priority key is on the intended-schema whitelist in tests/test_litellm/test_utils.py

Caveats (if any)

• The calculator API still shows pre-cut terra ($2.50 input) and luna ($1.00 input). The Foundry blog, OpenAI list, and Azure Support say the 7/30 cut applies. Follow the blog for those two
• Long+priority cache writes on the four global entries are inferred. Azure publishes no long+priority cell for any 5.6 variant
• Azure has no gpt-5-6 base offer. azure/gpt-5.6 is sol rates
• Luna priority / long+priority are inferred. Azure publishes no luna priority cells

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

---

Note

Medium Risk
Changes spend calculation for Azure GPT-5.6 (including previously free cache writes and lower data-zone priority rates). Wrong cells would systematically over- or under-bill customers.

Overview
Azure GPT-5.6 cache-write tokens were billed at $0. This adds cache_creation_input_token_cost (and long-context / priority variants) to all 12 azure/gpt-5.6* entries at 1.25x the matching input cell.

Also corrects US/EU data-zone priority rates for gpt-5.6/-sol: prompt $11/1M, cache-read $1.10/1M, output $66/1M, writes $13.75/1M (previously 2.5x input instead of Azure’s 2x). Global long+priority writes are inferred; data-zone long+priority falls back to the standard long write rate.

Schema whitelist now includes cache_creation_input_token_cost_above_272k_tokens_priority. Tests pin write billing across standard/priority/long context and OpenAI flex/priority cache-write plumbing.

^{Reviewed by Cursor Bugbot for commit bcc2d54. Bugbot is set up for automated code reviews on this repo. Configure here.}

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR corrects Azure Foundry GPT-5.6 prompt-cache write billing and regional priority pricing

• Adds standard, priority, and long-context cache-write rates for all 12 Azure GPT-5.6 pricing entries
• Aligns US and EU GPT-5.6 and Sol priority input, cache-read, and output rates with the published data-zone pricing
• Extends the pricing schema and regression tests for cache-write tier selection
• Replaces the previously skipped regional long-priority case with an explicit fallback assertion

Confidence Score: 5/5

The PR appears safe to merge

No blocking failure remains

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds Azure GPT-5.6 cache-write pricing and corrects data-zone GPT-5.6 and Sol priority rates
litellm/model_prices_and_context_window_backup.json	Keeps the bundled backup pricing map synchronized with the canonical pricing data
model_prices_and_context_window.schema.json	Declares the combined long-context priority cache-write pricing field
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py	Covers all Azure GPT-5.6 cache-write tiers, including explicit regional long-priority fallback behavior
tests/test_litellm/test_utils.py	Updates pricing-map validation for the combined long-context priority cache-write field

_{Reviews (7): Last reviewed commit: "fix(cost): restore Azure GPT-5.6 luna to..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[omertal323]

@greptileai

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing omertal323:litellm_azure_gpt56_cache_write_prices (bcc2d54) with litellm_internal_staging (d542c82)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (79cac36) during the generation of this report, so d542c82 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[omertal323]

@greptileai

[omertal323]

@greptileai

[omertal323]

@mubashir1osmani / @mateo-berri this complements your gpt-5.6 pricing work in #35481: azure cache writes for the 5.6 series currently bill at $0. This PR adds them for all 12 azure entries at the published rate (1.25x of input in every cell Azure lists), rebased on top of your terra/luna cut so the write prices track the new inputs. Greptile 5/5, tests pin every rate

[omertal323]

@greptileai

[omertal323]

@mateo-berri any chance for a review here?

[mubashir1osmani]

Pushed a follow-up: data-zone priority writes are now $13.75/1M, luna tracks Azure list. @greptileai

[mubashir1osmani]

Reverted luna to the 7/30 Foundry cut ($0.25 writes). Data-zone sol priority stays at $13.75. @greptileai

[mubashir1osmani]

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 bcc2d54. Configure here.}