fix(pricing): add missing Azure gpt-5.6 cache-write rates

[jlan-nl]

TLDR

Problem this solves:

• Azure gpt-5.6 cache-write tokens are billed at zero
• Spend under-reported by ~98% on cache-heavy prompts

How it solves it:

• Adds the missing cache-write rate to 12 azure gpt-5.6 entries
• Derives each from the OpenAI twin's cache-to-input ratio

User Flow

Before: a platform team charging back Azure OpenAI spend sees almost nothing billed for their RAG service, whose requests carry a large cached system prompt

1. Their RAG service sends POST https://litellm-domain/v1/chat/completions with "model": "azure-gpt-5.6-sol" and a ~2,200-token system prompt
2. The response comes back 200 with "usage": {"prompt_tokens": 2219, "completion_tokens": 5, "prompt_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 2216}}
3. They open https://litellm-domain/ui/?page=logs and the request shows $0.000165 spend, as if the 2,216 cache-write tokens were free
4. Their monthly Azure invoice for that traffic is roughly 85x the total the gateway reports, so per-team chargeback is unusable

After: the same request is billed at the cache-write rate, and the gateway total lines up with the invoice

1. Their RAG service sends the same POST https://litellm-domain/v1/chat/completions with "model": "azure-gpt-5.6-sol" and the same ~2,200-token system prompt
2. The response comes back 200 with the same usage object
3. https://litellm-domain/ui/?page=logs now shows that request at $0.014015 spend
4. The monthly total tracks the Azure invoice, so chargeback per team is accurate

Relevant issues

Fixes #37631

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)

Screenshots / Proof of Fix

Shared setup, a live Azure OpenAI gpt-5.6-sol deployment behind the proxy:

model_list:
  - model_name: azure-gpt-5.6-sol
    litellm_params:
      model: azure/gpt-5.6-sol
      api_base: os.environ/AZURE_OPENAI_ENDPOINT
      api_version: "2025-04-01-preview"

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
  database_url: os.environ/DATABASE_URL

The request body, sized so the prompt is large enough for Azure to write it to cache:

python - <<'EOF'
import json
prefix = 'You are a documentation assistant. ' + (
    'The following is reference material about cloud architecture patterns. ' * 220)
json.dump({"model": "azure-gpt-5.6-sol",
           "messages": [{"role": "system", "content": prefix},
                        {"role": "user", "content": "say ok"}],
           "max_tokens": 10}, open("req.json", "w"))
EOF

Before (31a6756)

1. Start the proxy at the merge base and send the request, confirming the provider reports cache writes:

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" --data @req.json | jq .usage

2. Read back what the gateway recorded for it:

curl -s "http://localhost:4000/spend/logs" \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq '.[-1] | {model, spend, prompt_tokens}'

Expected to show the cache-write tokens costing nothing, around $0.000165 against a correct $0.014015

After (0400f0b)

1. Restart the proxy at this PR's tip and send the identical request:

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" --data @req.json | jq .usage

2. Read back the recorded spend:

curl -s "http://localhost:4000/spend/logs" \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" | jq '.[-1] | {model, spend, prompt_tokens}'

Expected to show the same usage object now billed at the cache-write rate, around $0.014015

Both runs are pending: no Azure gpt-5.6 deployment was available to whoever prepared this branch, so the output above is stated as an expectation rather than a captured run. The commands are ready to paste and this section will be replaced with the real output before review. In the meantime the regression test in tests/test_litellm/test_cost_calculator.py covers the same billing path, and every one of its 37 cases fails at the merge base and passes here

Type

🐛 Bug Fix

✅ Test

Caveats (if any)

• Purely additive: no existing price is modified
• Azure priority input rates look wrong, but that is separate
• Also fixes the azure/eu entries, which the issue omits

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

Adds cache-write pricing for 12 Azure GPT-5.6 model entries so cached prompt creation is included in spend calculations

• Adds global, EU, and US cache-creation rates for standard, long-context, and priority tiers
• Keeps the primary and backup model-price catalogs synchronized
• Adds metadata consistency and end-to-end cost-calculator regression coverage

Confidence Score: 5/5

The PR appears safe to merge, with consistent pricing metadata and focused regression coverage

All affected Azure GPT-5.6 variants receive cache-creation rates derived consistently from their corresponding input rates, both catalogs remain synchronized, and the billing path is exercised

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds internally consistent cache-creation rates to all 12 affected Azure GPT-5.6 entries
litellm/model_prices_and_context_window_backup.json	Mirrors the primary pricing-catalog changes without divergence
tests/test_litellm/test_cost_calculator.py	Verifies every added rate against its OpenAI twin and exercises cache-write billing through the generic calculator

_{Reviews (1): Last reviewed commit: "fix(pricing): add missing Azure gpt-5.6 ..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing jlan-nl:litellm-azure-gpt56-cache-write-pricing (0400f0b) with litellm_internal_staging (31a6756)}

[jlan-nl]

Closing, superseded by #38370