test(pricing): cover gpt-5.6 cache-cost plumbing and bedrock_mantle responses billing

[mateo-berri]

TLDR

Problem this solves:

• Follow-up to fix(pricing): correct gpt-5.6 prices for openai, bedrock, and flex long context #35270, closing the two QA-runbook gaps that kept its last checkboxes unchecked
• The cache-cost fields wired through get_model_info there (cache_creation_input_token_cost_flex/_priority, and the *_above_272k_tokens_flex cache variants) had no billing test; re-dropping them from the explicit kwargs list passed the whole suite
• No test exercised a /v1/responses call against a bedrock_mantle gpt-5.6 deployment end to end through the cost calculator

How it solves it:

• test_generic_cost_per_token_gpt56_terra_cache_costs_by_tier_and_context bills a request with cache-read and cache-write tokens across five (service tier, context size) combinations and asserts the exact prompt cost. Mutation-verified: reverting the cache wiring in litellm/utils.py fails the flex, priority, and flex-long-context rows
• test_gpt_5_6_responses_call_cost runs a ResponsesAPIResponse for all three bedrock_mantle gpt-5.6 models through litellm.completion_cost and asserts the exact spend at the corrected rates

The standard above-272k cache-write row passes even without the explicit wiring because _get_model_info_helper has a regex passthrough for keys ending in _above_<n>k_tokens; tier-suffixed keys do not match it, which is exactly why the explicit wiring needs its own coverage.

The remaining wired-but-untested fields (cache_*_above_272k_tokens_priority) have no data in the cost map because OpenAI publishes no long-context column for the Fast tier, so there is no billing behavior to lock down.

Relevant issues

Follow-up to #35270

Pre-Submission checklist

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

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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)

Type

🧪 Tests

Changes

Test-only; no runtime code changes.

[greptile-apps]

Greptile Summary

Adds focused regression coverage for GPT-5.6 billing:

• Verifies OpenAI GPT-5.6 Terra cache-read and cache-write pricing across standard, flex, priority, and long-context tiers
• Verifies Responses API cost calculation for three Bedrock Mantle GPT-5.6 models

Confidence Score: 5/5

The PR appears safe to merge; the added tests exercise the intended pricing paths without changing production behavior

The cache test reaches tier-specific and long-context cache pricing selection, while the Bedrock Mantle test uses the same provider-prefixed model and Responses API usage conversion used by production cost calculation

Important Files Changed

Filename	Overview
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py	Adds exact cache-cost assertions across service-tier and context-size combinations using the calculator's expected OpenAI token accounting shape
tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py	Adds Responses API completion-cost coverage for all registered Bedrock Mantle GPT-5.6 variants

_{Reviews (1): Last reviewed commit: "test(pricing): cover gpt-5.6 cache-cost ..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!