feat(model_prices): add missing Databricks Foundation Model API models and refresh Gemini 2.5 rates

[leecoder]

Summary

Adds 19 Databricks Foundation Model API models missing from model_prices_and_context_window.json (and the packaged backup), based on the official Databricks supported models docs and pricing pages.

Added models

• OpenAI: databricks-gpt-5-6-sol, databricks-gpt-5-6-terra, databricks-gpt-5-6-luna, databricks-gpt-5-5, databricks-gpt-5-5-pro
• Google: databricks-gemini-3-8-flash, databricks-gemini-3-7-flash, databricks-gemini-3-6-flash, databricks-gemini-3-5-flash, databricks-gemini-3-5-flash-lite, databricks-gemini-3-1-flash-image, databricks-gemini-3-pro-image
• Alibaba: databricks-qwen35-122b-a10b, databricks-qwen3-next-80b-a3b-instruct, databricks-qwen3-embedding-0-6b
• Zhipu: databricks-glm-5-3, databricks-glm-5-3-flash (DBU rates now published on the FMS pricing page)
• Others: databricks-grok-4-6 (xAI), databricks-inkling (Thinking Machines), databricks-claude-fable-5-1 (Anthropic)

Updates

• databricks-gemini-2-5-flash / databricks-gemini-2-5-pro: adds cache write/read rates per the official cache-read DBU columns (rates keep the promotional pricing the repo already tracks)
• Adds cache write/read cost fields + supports_prompt_caching to newly published models per the official cache DBU columns
• gemini-3-7-flash / gemini-3-8-flash: endpoint metadata only — Databricks has not published DBU rates for these yet
• Tests: PUBLISHED_DBU_PER_MILLION extended with the 12 newly published models; unpublished-cache-rate tripwire 14 → 18

Pricing methodology

All rates are derived from the official Databricks Foundation Model Serving DBU tables using the LiteLLM convention USD = DBU × $0.07 per 1M tokens (*_dbu_cost_per_token fields are used in actual cost calculations).

Source pages:

• https://www.databricks.com/product/pricing/foundation-model-serving
• https://www.databricks.com/product/pricing/proprietary-foundation-model-serving
• https://docs.databricks.com/aws/en/machine-learning/foundation-models/supported-models

[greptile-apps]

Greptile Summary

The PR expands both model-price registries with Databricks Foundation Model metadata, refreshes Gemini 2.5 pricing, and adds prompt-cache pricing and capability fields. The two registry copies are synchronized, but the Databricks cost path does not consume the newly supplied cache rates.

• Adds Databricks model identifiers, context limits, capabilities, and pricing.
• Updates existing Gemini 2.5 rates.
• Adds cache-read and cache-creation pricing across existing and new Databricks entries.

Confidence Score: 4/5

The PR should not merge until Databricks cached-token usage is charged using the cache-read and cache-creation rates added by this change.

Databricks requests bypass the generic cache-aware calculator, causing cached prompt tokens to remain priced at the full input rate in recorded spend and budget accounting.

Files Needing Attention: model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds and updates Databricks pricing metadata, but its cache-rate fields are ineffective in the provider-specific cost path.
litellm/model_prices_and_context_window_backup.json	Mirrors the root pricing registry, including the same ineffective Databricks cache-rate metadata.

_{Reviews (1): Last reviewed commit: "feat(model_prices): add missing Databric..." | Re-trigger Greptile}

[greptile-apps]

Cache rates bypass cost accounting

When a Databricks completion reports cache-read or cache-creation prompt tokens, the provider-specific calculator charges every prompt token at input_cost_per_token instead of the cache rates added here, causing recorded spend and budget accounting to overcharge cached traffic.

[leecoder]

Thanks for the review — this is actually already handled.

The Databricks provider routes through generic_cost_per_token (see litellm/llms/databricks/cost_calculator.py), which calls get_billable_input_tokens() — it subtracts cached tokens from prompt tokens and charges them at cache_read_input_token_cost / cache_creation_input_token_cost. The existing tests in tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py (test_cached_tokens_bill_at_cache_rates) assert exactly this path, and they pass on the latest commit.

The cache rates added by this PR were previously missing for the new models, which meant cached tokens fell back to the full input rate — the fix in 6961442e4b completes those fields (and corrects a few copied-by-mistake values), so cached Databricks traffic is now billed at the published cache DBU rates.

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing leecoder:feat/databricks-model-prices (6961442) with litellm_internal_staging (c8635ec)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[devin-ai-integration]

Superseded by #39388, which carries these Databricks entries with values re-read from the Databricks pricing and supported-models pages. Thanks for the catalog work