fix(openrouter): pattern-based fix for native model double-stripping

[tombii]

Relevant issues

Fixes #16353
Supersedes #20516

Pre-Submission checklist

• I have Added testing in the tests/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

Type

🐛 Bug Fix

Changes

PR #20516 fixed the double-stripping issue for three hard-coded native OpenRouter models (openrouter/auto, openrouter/free, openrouter/bodybuilder) by maintaining a static set. However, this approach will not fix the issue for other native OpenRouter models such as openrouter/aurora-alpha or openrouter/polaris-alpha.

This PR replaces the hardcoded set with a pattern-based check in _get_openai_compatible_provider_info: after stripping the outer openrouter/ provider prefix, if the remaining model name still starts with openrouter/, return immediately without stripping again.

# Before (hardcoded set — won't cover openrouter/aurora-alpha etc.)
NATIVE_OPENROUTER_MODELS = {"openrouter/auto", "openrouter/free", "openrouter/bodybuilder"}
if model in NATIVE_OPENROUTER_MODELS:
    return model, "openrouter", dynamic_api_key, api_base

# After (pattern-based — handles any native OpenRouter model)
if custom_llm_provider == "openrouter" and model.startswith("openrouter/"):
    dynamic_api_key = api_key or get_secret_str("OPENROUTER_API_KEY")
    return model, custom_llm_provider, dynamic_api_key, api_base

This handles all current and future native OpenRouter models without needing a list to maintain.

Tests added

tests/test_litellm/llms/openrouter/test_openrouter_provider_routing.py:

• Double-prefixed native models strip correctly (including aurora-alpha, polaris-alpha, and arbitrary future models)
• Second call to get_llm_provider on an already-stripped native model does not strip further
• Regular OpenRouter models (e.g. openrouter/anthropic/claude-3-haiku) still strip normally

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 27, 2026 10:24pm

[greptile-apps]

Greptile Summary

This PR replaces a hardcoded set of native OpenRouter models (openrouter/auto, openrouter/free, openrouter/bodybuilder) with a pattern-based check that handles any model whose ID starts with openrouter/ on the OpenRouter API. The core fix in _get_openai_compatible_provider_info is clean and correct — it prevents the openrouter/ prefix from being stripped twice when a user specifies openrouter/openrouter/<model>.

• The pattern-based early return in get_llm_provider_logic.py correctly detects when the remaining model name still starts with openrouter/ after the outer provider prefix is stripped, and returns immediately to preserve the native model ID
• Tests for the primary use case (double-prefixed input → single strip) and regular model passthrough are well-structured
• Issue: test_no_double_strip_on_second_call has an incorrect assertion — passing an already-resolved model name like openrouter/aurora-alpha back through get_llm_provider will strip it to aurora-alpha, not preserve it. This test likely fails as written.

Confidence Score: 3/5

• Core fix is correct but one test has a likely-failing assertion that needs to be verified and fixed before merge
• The main code change is sound and well-targeted — it solves the double-stripping issue for all native OpenRouter models without maintaining a hardcoded list. However, the test_no_double_strip_on_second_call test appears to assert incorrect expected values, which suggests it either wasn't run or there's a code path I'm not seeing. This needs verification before merging.
• tests/test_litellm/llms/openrouter/test_openrouter_provider_routing.py — the test_no_double_strip_on_second_call test's second-call assertion appears incorrect

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/get_llm_provider_logic.py	Adds a pattern-based early return for native OpenRouter models (e.g. openrouter/openrouter/auto) to prevent double-stripping of the openrouter/ prefix. The logic is clean and correctly placed within the existing provider routing function.
tests/test_litellm/llms/openrouter/test_openrouter_provider_routing.py	Good test coverage overall, but test_no_double_strip_on_second_call has incorrect assertions — the second get_llm_provider call on an already-resolved model like openrouter/aurora-alpha will strip it further to aurora-alpha, failing the assertion that expects idempotent output.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["User input: openrouter/openrouter/aurora-alpha"] --> B["get_llm_provider()"]
    B --> C{"model prefix in provider_list?"}
    C -->|Yes: openrouter| D["_get_openai_compatible_provider_info()"]
    D --> E["Split: provider='openrouter', model='openrouter/aurora-alpha'"]
    E --> F{"provider=='openrouter' AND\nmodel starts with 'openrouter/'?"}
    F -->|Yes - NEW CHECK| G["Early return:\nmodel='openrouter/aurora-alpha'\nprovider='openrouter'"]
    F -->|No| H["Continue to provider elif chain"]
    H --> I["Return model with provider-specific\napi_base and api_key"]

    J["User input: openrouter/anthropic/claude-3-haiku"] --> B
    D --> K["Split: provider='openrouter', model='anthropic/claude-3-haiku'"]
    K --> F

Loading

_{Last reviewed commit: d292da2}

[greptile-apps]

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Test assertion appears incorrect for second call

On the second call, litellm.get_llm_provider(model="openrouter/aurora-alpha") enters _get_openai_compatible_provider_info where the model is split: custom_llm_provider = "openrouter", model = "aurora-alpha". Since "aurora-alpha" does not start with "openrouter/", the new early-return guard doesn't trigger, and the function returns model = "aurora-alpha" (not "openrouter/aurora-alpha").

This means the assertion model_second == expected_model (i.e., "aurora-alpha" == "openrouter/aurora-alpha") should fail. The second get_llm_provider call strips the prefix as designed — this is arguably the correct behavior when the caller passes an already-resolved model name back in. The test should either:

1. Assert model_second == "aurora-alpha" (the actual output), or
2. Not pass the already-resolved model back through get_llm_provider without re-adding the openrouter/ provider prefix.

[krrishdholakia]

@tombii can you address the greptile feedback please?