fix(utils): resolve bedrock regional inference profiles to regional pricing in get_model_info (LIT-4056)

[mateo-berri]

Relevant issues

Linear ticket

Resolves LIT-4056

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)

Screenshots / Proof of Fix

Live proxy on localhost with no mocks and no database, started from a clean worktree venv with LITELLM_LOCAL_MODEL_COST_MAP=True and LITELLM_MASTER_KEY=sk-qa-lit4056, using this qa_config.yaml, which mirrors the deployment shape from the ticket:

model_list:
  - model_name: claude-opus-4-8
    litellm_params:
      model: bedrock/au.anthropic.claude-opus-4-8
      aws_region_name: ap-southeast-2

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

Both runs use the exact same startup and curl commands on the same port; the only difference is the checked-out commit

.venv/bin/python litellm/proxy/proxy_cli.py --config qa_config.yaml --port 59668

curl -s "http://127.0.0.1:59668/v1/model/info" -H "Authorization: Bearer sk-qa-lit4056" | jq '.data[] | {model_name, litellm_model: .litellm_params.model, key: .model_info.key, input_cost_per_token: .model_info.input_cost_per_token, output_cost_per_token: .model_info.output_cost_per_token}'

Before (base commit ff6dc33)

The AU regional inference profile resolves to the region-stripped base cost-map entry and shows base pricing

{
  "model_name": "claude-opus-4-8",
  "litellm_model": "bedrock/au.anthropic.claude-opus-4-8",
  "key": "anthropic.claude-opus-4-8",
  "input_cost_per_token": 0.000005,
  "output_cost_per_token": 0.000025
}

After (PR head commit dadf926)

The same deployment resolves to the exact regional cost-map entry with the AU premium pricing

{
  "model_name": "claude-opus-4-8",
  "litellm_model": "bedrock/au.anthropic.claude-opus-4-8",
  "key": "au.anthropic.claude-opus-4-8",
  "input_cost_per_token": 0.0000055,
  "output_cost_per_token": 0.0000275
}

Type

🐛 Bug Fix

Changes

get_model_info resolved Bedrock regional inference profiles configured with the standard bedrock/ routing prefix (including bedrock/converse/ and bedrock/invoke/) to the region-stripped base cost-map entry instead of the exact regional entry. For example bedrock/au.anthropic.claude-opus-4-8 returned key anthropic.claude-opus-4-8 at $5.00/$25.00 per 1M tokens instead of au.anthropic.claude-opus-4-8 at $5.50/$27.50, so /v1/model/info and the Admin UI model overview showed base pricing for premium-priced regional profiles. The unprefixed form (model: au.anthropic.claude-opus-4-8 with custom_llm_provider: bedrock) already resolved correctly, so the two config spellings of the same deployment disagreed

Root cause: _get_model_info_helper tries lookup candidates in a fixed order and placed split_model (the exact name after the provider prefix) last, after the region-stripped candidates produced by _strip_model_name. For Bedrock, stripping removes the cross-region prefix (us., eu., au., and so on), so the base entry always matched first whenever the model carried the bedrock/ prefix. For the converse/ and invoke/ route forms, split_model additionally kept the route prefix, so the exact regional key was never even a candidate

The fix tries the exact split_model candidate before the stripped fallbacks in both _get_model_info_helper and _get_model_info_from_generalization, and normalizes Bedrock route prefixes off split_model in _get_potential_model_names. The existing provider-match guard keeps split_model hits from leaking across providers, and the region-stripped fallback still applies for regional profiles that have no dedicated cost-map entry, for example bedrock/jp.anthropic.claude-opus-4-8 still resolves to the base entry. Runtime spend tracking through cost_per_token was already correct because completion flows pass the unprefixed model plus provider; this change aligns get_model_info with it

Regression tests in tests/test_litellm/test_utils.py cover the plain, converse/, and invoke/ prefixed forms resolving to the regional entry, the unprefixed control form, and the base-entry fallback for a regional profile without a dedicated entry

One intended side effect: a doubled provider prefix such as bedrock/bedrock/us.anthropic.claude-sonnet-4-6 now resolves too, because the second bedrock/ is treated as the routing prefix it already is at request time (strip_bedrock_routing_prefix removes it before the call is sent), so model info now matches what the request actually bills as. test_register_model_inherits_builtin_cache_pricing_for_unmapped_key relied on that shape being unresolvable to exercise the register_model fallback path, so its fixture moved to a triple prefix, which still cannot resolve; the new resolution behavior is locked in by test_get_model_info_bedrock_double_provider_prefix_resolves

[greptile-apps]

Greptile Summary

This PR fixes get_model_info incorrectly resolving Bedrock regional inference profiles (e.g. bedrock/au.anthropic.claude-opus-4-8) to the region-stripped base cost-map entry instead of the exact regional entry, causing /v1/model/info and the Admin UI to show base pricing when premium-priced AU/EU/US regional profiles were configured with the bedrock/ routing prefix.

• The lookup order in both _get_model_info_helper and _get_model_info_from_generalization is changed so that split_model (the exact name after stripping the provider prefix) is tried before the stripped fallbacks.
• A new Bedrock-specific block in _get_potential_model_names calls strip_bedrock_routing_prefix to normalize routing sub-prefixes (converse/, invoke/) off split_model, ensuring all three call forms resolve to the same entry.
• The fixture in test_register_model_inherits_builtin_cache_pricing_for_unmapped_key is updated from a double-prefix to a triple-prefix shape because double-prefix keys now resolve correctly, and the test needs a genuinely unresolvable key to exercise the fallback path; the new resolution behavior is locked in by an added test.

Confidence Score: 5/5

Safe to merge — the change only affects the model-info lookup order and does not touch any request-execution or billing paths at runtime.

The fix is narrowly scoped to get_model_info resolution order. The fallback chain is preserved (unmapped regional profiles still fall through to the base entry), the provider-match guard prevents cross-provider leakage, and the modified test fixture is demonstrably necessary because the double-prefix shape now resolves by design. The new tests cover the three routing-prefix forms, the fallback case, and the double-prefix case.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/utils.py	Lookup order in _get_model_info_helper and _get_model_info_from_generalization reordered to try split_model before stripped fallbacks; Bedrock routing-prefix normalization added to _get_potential_model_names. Logic is correct and follows the existing dispatch pattern.
tests/test_litellm/test_utils.py	Three new regression tests added covering plain, converse/, and invoke/ prefixed forms, the fallback for unmapped regional profiles, and the double-prefix resolution case. All use local_model_cost_map fixture — no real network calls.
tests/test_litellm/test_register_model_custom_pricing.py	Fixture key bumped from bedrock/bedrock/ to bedrock/bedrock/bedrock/ so the test still exercises the unresolvable-key fallback path; docstring updated to match. The change is intentional and the new behavior is locked in by a separate added test.

_{Reviews (2): Last reviewed commit: "test(register_model): use a triple provi..." | 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

✅ 30 untouched benchmarks

---

_{Comparing litellm_lit4056_bedrock_regional_model_info (a4cee67) with litellm_internal_staging (4b0ac8b)}

[mateo-berri]

@greptileai