OpenAI-compat: OpenAISDK openai_api_base /v1, not the requests OpenAI full-path default

Working config (OpenAISDK, openai_api_base /v1)

OpenCompass already wraps Chat Completions as OpenAISDK. A host whose public catalog is GET https://api.pzero.studio/v1/models (no key) fits that class.

from opencompass.models import OpenAISDK

models = [
    dict(
        type=OpenAISDK,
        abbr="pzero",
        path="deepseek-v4-flash",
        key="<Bearer key for that host>",
        openai_api_base="https://api.pzero.studio/v1",
        query_per_second=1,
        max_out_len=1024,
        max_seq_len=8192,
        batch_size=1,
    ),
]

openai_api_base is the /v1 root (SDK base_url). path is a catalog id.

Do not use the requests OpenAI class with that /v1 root. That wrapper posts openai_api_base as-is, so /v1 404s; it wants the full .../v1/chat/completions URL. Do not use OpenAISDKResponse (Responses API). any(token in path for token in ('o1', 'o3', 'o4', 'gpt-5')) swaps in max_completion_tokens; pick a catalog id that does not contain those substrings.

Not asking for a first-party provider. OpenAISDK plus openai_api_base is enough.