fix(router): register model info under responses/-stripped variant

[krisxia0506]

Relevant issues

When the Router registers a deployment into litellm.model_cost, the downstream responses_api_bridge_check() strips the responses/ prefix from the model name. If a deployment is only registered under azure/responses/..., lookups for azure/... miss, which can affect flags like supports_native_streaming (e.g. falsely triggering fake streaming).

Linear ticket

N/A

Pre-Submission checklist

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

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Screenshots / Proof of Fix

Unit tests covering both registration paths:

• test_responses_prefix_stripped_alias_registered_for_model_list
• test_responses_prefix_stripped_alias_registered_for_add_deployment

Type

🐛 Bug Fix
✅ Test

Changes

Bug Fix

• _create_deployment: after registering the shared backend key, if the deployment's litellm_params.model contains responses/, additionally register an alias with that prefix stripped, so post-bridge lookups resolve correctly.
• add_deployment: on top of the existing model_id-level custom pricing registration, mirror the OLD MODEL REGISTRATION and the stripped alias from _create_deployment, so runtime-added deployments match the behavior of YAML-bootstrapped ones.

Affected files

• litellm/router.py
• tests/test_litellm/test_router_model_cost_isolation.py

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing qiniu:fix/router-register-model-cost-responses-stripped (73eefc8) with main (fa81017)}

[greptile-apps]

Greptile Summary

This PR fixes a lookup miss in litellm.model_cost for deployments whose litellm_params.model starts with "responses/": because responses_api_bridge_check() strips that prefix before performing model-info lookups, only registering azure/responses/gpt-4o left azure/gpt-4o unresolvable, which could falsely suppress flags like supports_native_streaming.

• _create_deployment: after building the shared backend alias in litellm.model_cost, a second entry without the responses/ segment is now also registered.
• add_deployment: mirrors the same dual-registration so dynamically added deployments (e.g. from DB) match the YAML-bootstrapped path; _model_info_dict construction is moved up to be shared between both registration calls.
• Tests: two new mock-only tests cover both registration paths, each verifying both the prefixed and stripped keys exist in litellm.model_cost with the expected field values.

Confidence Score: 4/5

Safe to merge; the fix is narrow and well-tested, with only a minor guard condition that could be tightened.

The core logic is correct and both changed code paths are covered by the new tests. The only concern is the guard "responses/" in _model_name operating on the fully-constructed provider/model string instead of the raw litellm_params.model field — a model whose name incidentally contains "responses/" mid-string would produce a spurious stripped alias in litellm.model_cost. This is an edge case in practice but the fix (using startswith on the raw field and capping replace to count=1) is straightforward.

The two alias-registration blocks in litellm/router.py (around lines 7079–7083 and 7787–7791) both share the same broad guard condition and are worth a second look before merging.

Important Files Changed

Filename	Overview
litellm/router.py	Registers a stripped alias (without responses/ prefix) in litellm.model_cost for both _create_deployment and add_deployment. The guard condition "responses/" in _model_name is broader than intended — it operates on the fully-formed provider/model string rather than on litellm_params.model directly, which could cause false-positive alias creation for non-responses-API models that happen to contain that substring.
tests/test_litellm/test_router_model_cost_isolation.py	Adds two new unit tests verifying that the stripped alias is registered for both model-list and add_deployment paths. Tests use fake API keys and only inspect litellm.model_cost in memory — no real network calls are made, consistent with the mock-only rule for this test directory.

_{Reviews (1): Last reviewed commit: "fix(router): register model info under r..." | Re-trigger Greptile}

[greptile-apps]

The check "responses/" in _model_name operates on the fully-constructed provider/model string, so a model named something like azure/custom-responses/v2 (where "responses/" is not a prefix of the model field) would also trigger the alias stripping and produce azure/custom-v2 — a potentially incorrect or conflicting key. A tighter guard that checks only the litellm_params.model field (before prepending the provider) avoids this false-positive. str.replace without count=1 would also replace multiple occurrences of "responses/" if they exist.

Suggested change

	_backend_alias_cost = {_model_name: _shared_model_info}
	if "responses/" in _model_name:
	_stripped_model_name = _model_name.replace("responses/", "")
	_backend_alias_cost[_stripped_model_name] = _shared_model_info
	litellm.register_model(model_cost=_backend_alias_cost)
	## Check if LLM Deployment is allowed for this deployment
	_backend_alias_cost = {_model_name: _shared_model_info}
	if deployment.litellm_params.model.startswith("responses/"):
	_stripped_model_name = _model_name.replace("responses/", "", 1)
	_backend_alias_cost[_stripped_model_name] = _shared_model_info
	litellm.register_model(model_cost=_backend_alias_cost)
	## Check if LLM Deployment is allowed for this deployment

[greptile-apps]

Same overly-broad guard in add_deployment: "responses/" in _model_name can match a model whose name happens to contain that substring in a non-prefix position. Checking deployment.litellm_params.model.startswith("responses/") is the precise equivalent of what responses_api_bridge_check actually tests, and passing count=1 to replace keeps the replacement bounded.

Suggested change

	_backend_alias_cost = {_model_name: _shared_model_info}
	if "responses/" in _model_name:
	_stripped_model_name = _model_name.replace("responses/", "")
	_backend_alias_cost[_stripped_model_name] = _shared_model_info
	litellm.register_model(model_cost=_backend_alias_cost)
	# add to model names
	_backend_alias_cost = {_model_name: _shared_model_info}
	if deployment.litellm_params.model.startswith("responses/"):
	_stripped_model_name = _model_name.replace("responses/", "", 1)
	_backend_alias_cost[_stripped_model_name] = _shared_model_info
	litellm.register_model(model_cost=_backend_alias_cost)
	# add to model names

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[oss-pr-review-agent-shin]

🤖 litellm-agent: Squash-merged into staging branch shin_agent_oss_staging_05_09_2026. Staging PR: #27549

---

Triage Summary
Gathered PR data only — the triage LLM step did not produce a valid report, so failing-check classification and prior-signal reconciliation were skipped. 101 line(s) across 2 file(s) (+91/-10).

101 lines across 2 files (+91 / -10)

Merge Confidence: 5/5 ✅ READY
Ready to ship.

All checks green. Greptile 4/5, no blocking pattern findings, no CircleCI runs (OSS-typical).