We don't publish
your competitive advantage.

IFLiteLLM proxy returns 404 error when calling the Responses API (v1/responses) for an OpenAI-compatible model, while chat completions work fine.

THENRegister the model in LiteLLM's `model_prices_and_context_window.json` file with proper pricing and context window metadata. If the model is not mapped, the Responses API fails because it requires model info for routing. Alternatively, use the chat completions endpoint (`/v1/chat/completions`) which does not enforce this mapping. For custom models, add an entry like: `{ "model": "openai/gpt-oss-120b", "input_cost_per_token": 1e-7, "output_cost_per_token": 5e-7 }`.

IFMerging a PR that updates the external model cost JSON file causes automatic routing from 'completions' to 'responses' API endpoints, leading to request failures without library version change.

THENSet the environment variable LITELLM_LOCAL_MODEL_COST_MAP='True' to force LiteLLM to use the local model cost map from the package backup file, preventing dynamic endpoint switching based on remote configuration changes. For production stability, consider pinning to a specific model cost file or monitoring the remote JSON for unexpected changes.

IFLiteLLM proxy versions >=1.82.1 send full model ID including 'openrouter/' prefix to OpenRouter API, causing 400 error 'not a valid model ID' for models with nested slashes (e.g., openrouter/google/gemini-3-flash-preview).

THENModify the `get_llm_provider()` function to only strip the 'openrouter/' prefix for native OpenRouter models (those with no additional slash after the prefix, e.g., 'openrouter/auto' or 'openrouter/free'). Specifically, after stripping 'openrouter/', check if the remainder contains a '/'; if so, preserve the full prefix. Alternatively, downgrade to LiteLLM 1.82.0 or configure the model as a custom OpenAI provider with API base https://openrouter.ai/api/v1.

IFLiteLLM proxy returns 404 on /v1/responses endpoint for custom OpenAI-compatible models, while /v1/chat/completions works.

THENMap the custom model in LiteLLM's model_prices_and_context_window.json or use a model alias already present. Alternatively, use the Chat Completions API for models that do not support the Responses API endpoint.

IFLiteLLM proxy version 1.82.1 or later returns a 400 'not a valid model ID' error for all OpenRouter models because the proxy sends the model ID with the 'openrouter/' prefix to the OpenRouter API.

THENDowngrade to LiteLLM 1.82.0, or configure the model using the Custom OpenAI provider with API base https://openrouter.ai/api/v1, or apply the patch from PR #23539 that checks for a trailing slash after stripping the 'openrouter/' prefix to only strip it for native OpenRouter models (e.g., 'openrouter/auto') and pass the remainder unchanged.

IFWhen using OpenRouter models in LiteLLM ≥1.82.1, the proxy sends the full 'openrouter/...' prefix as the model ID to the OpenRouter API, causing a 400 error 'not a valid model ID'.

THENDowngrade to LiteLLM 1.82.0 (and apply a one-line type patch for the complexity router if needed) or apply the fix from PR #23539: in get_llm_provider(), only pass the full 'openrouter/...' prefix for native models (e.g., openrouter/auto, openrouter/free) that have no additional '/'; for all other models, strip the prefix before sending to the API.

IFWhen a request to /v1/images/generations uses a Gemini model (e.g., vertex_ai/gemini-2.5-flash-image-preview), LiteLLM returns a 404 error because it routes to the :predict endpoint instead of :generateContent.

THENDetect Gemini models in image generation requests and convert the request to a chat completions format with modalities: ['text', 'image']. Route internally to the :generateContent endpoint, then convert the response back to the standard image generation format to maintain API compatibility.

IFModels are automatically switched from completions endpoint to responses endpoint after a dynamic model cost map update, causing structured outputs to fail.

THENSet environment variable LITELLM_LOCAL_MODEL_COST_MAP=True to pin model cost data to the local backup file, preventing automatic bridging of completions to responses. This avoids unexpected changes until a controlled update of the package.