fix(cost map): add cached-input pricing and prompt caching for inception/mercury-2.5

[Nanduu24]

Summary

inception/mercury-2.5 is present in the cost map with input and output pricing, but is missing cached-input pricing and prompt-caching support:

• cache_read_input_token_cost — not set
• supports_prompt_caching — not set

As a result, cached-token usage on Mercury 2.5 is not cost-tracked (silent inaccuracy), even though Inception offers cached input and the sibling model inception/mercury-2 already carries both fields.

Changes

Adds to inception/mercury-2.5 in both the root cost map and the litellm/ backup copy (kept in sync):

"cache_read_input_token_cost": 2e-08,
"supports_prompt_caching": true

• 2e-08 = $0.02 / 1M tokens, the standard cached-input rate (the current $0.004 is a temporary launch promo).

Relevant issues

Addresses the cached-pricing portion of #40746 (the model entry itself was already added; these fields were still missing).

Type

🐛 Bug Fix / 💰 cost tracking

Testing

• Both JSON files parse successfully and remain byte-for-byte identical to each other.
• Values verified against Inception's pricing page: https://www.inceptionlabs.ai/models#pricing

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

Adds cached-input pricing and prompt-caching metadata for Inception Mercury 2.5 in both synchronized cost maps

• The new cached-input rate represents the future standard price rather than the acknowledged active promotional price
• No regression assertion covers either newly added field

Confidence Score: 4/5

This PR is not safe to merge until the active cached-input price is represented and the repository-required regression coverage is added

Cached tokens are multiplied directly by the configured rate, so using $0.02/M during the acknowledged $0.004/M promotion overstates spend fivefold

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

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds Mercury 2.5 cache pricing and capability metadata, but currently overstates cached-token costs and lacks regression coverage
litellm/model_prices_and_context_window_backup.json	Mirrors the root cost-map changes, including the same premature standard cached-input rate

_{Reviews (1): Last reviewed commit: "fix(cost map): add cached-input pricing ..." | Re-trigger Greptile}

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing Nanduu24:add-mercury-2.5-cache-pricing (a3c7008) with main (30f33a9)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Nanduu24]

Thanks for the review! Addressed the feedback:

Test (P2): Added tests/test_litellm/test_inception_mercury_2_5_model_metadata.py, following the existing per-model metadata test pattern (e.g. test_xai_grok_4_3_model_metadata.py). It asserts the pricing/capability fields — including cache_read_input_token_cost and supports_prompt_caching — and that the root and backup cost maps stay in sync.

Cached rate (P1): The 2e-08 ($0.02 / 1M) value is intentional — it's the standard cached-input rate, consistent with this entry's existing input_cost_per_token (2e-07 = $0.20, standard) and output_cost_per_token (7.5e-07 = $0.75, standard). Those already use standard rather than the temporary launch-promo rates ($0.04 / $0.004 / $0.15), so pricing cached input at the promo rate would make it inconsistent with input/output. This also matches the sibling inception/mercury-2 entry.

Note on CI: the failing proxy-behavior check is unrelated to this change — it's a pre-existing auth test (test_auth_object_prefetch.py::test_join_binds_the_membership_to_the_requested_team) failing with TypeError: object MagicMock can't be used in 'await' expression. The relevant cost-map-guard check passes.

[Nanduu24]

The failing proxy-behavior check is a pre-existing flake unrelated to this PR.

The only failure is:

tests/proxy_behavior/auth/test_auth_object_prefetch.py::test_join_binds_the_membership_to_the_requested_team
FAILED — TypeError: object MagicMock can't be used in 'await' expression

This is an async-mock issue in the auth suite, and it fails identically on the run before any test was added here. This PR only touches model_prices_and_context_window.json (+ the litellm/ backup) and a new test under tests/test_litellm/, none of which can affect tests/proxy_behavior/auth/. The relevant cost-map-guard check passes and the new metadata test passes.

A re-run should be green once the flaky auth test is addressed on main.

[devin-ai-integration]

Superseded by #41112, which adds the same mercury-2.5 cache pricing verified against the Inception docs.