A base url kept beside the deployment config

Wire formatsEN2026-10-29

Changing base_url in the OpenAI Python SDK is a small edit with a larger operational boundary behind it. You are selecting the server that receives credentials, interprets model identifiers, and implements each API resource. Treat that setting as deployment configuration, not merely a convenient string replacement.

§ Keep the API prefix in the base URL

The SDK appends resource paths to the configured base. For an OpenAI-compatible API mounted under /v1, configure the base through that prefix, not through a specific operation such as /chat/completions.

For Vana API, the service endpoint is https://api.vanaapi.com. Use https://api.vanaapi.com/v1 as the OpenAI-compatible base. A minimal explicit configuration is:

 1  import os

 2  from openai import OpenAI
 3  client = OpenAI(api_key=os.environ["VANA_API_KEY"], base_url="https://api.vanaapi.com/v1", timeout=60.0, max_retries=1)

The SDK handles a trailing slash on the base URL. More consequential mistakes are omitting the API prefix, duplicating it, or including the operation path. If a request returns an unexpected 404, inspect the final request URL before changing authentication or model settings.

§ Choose one configuration authority

The SDK recognizes OPENAI_BASE_URL and OPENAI_API_KEY. These are useful when an existing application constructs OpenAI() without arguments. Explicit constructor values take precedence over their corresponding environment variables.

For applications that use multiple services, provider-specific variables and explicit construction are usually easier to audit. VANA_API_KEY is an application-defined variable in the example above; the SDK does not discover it automatically.

Avoid mutating process-wide environment variables between requests. That pattern becomes fragile under concurrent workloads and makes imported libraries difficult to reason about. Construct a dedicated client for each endpoint and credential pair, then inject the appropriate client into your application components.

§ Separate transport compatibility from feature support

A successful base URL change proves routing, not complete API equivalence. Chat Completions, Responses, streaming, tools, structured outputs, and uploaded files are separate capabilities. Check the combination of endpoint, model, and feature your workload actually needs.

Vana API exposes an OpenAI-compatible endpoint and also speaks Anthropic Messages and Gemini native protocols. Pointing the OpenAI SDK at the service does not make it serialize those native protocols. The SDK still sends the OpenAI-shaped requests associated with the resource you call.

Keep model identifiers configurable as well. If selecting Claude Sonnet 5.5, use the exact API identifier supplied by Vana API rather than deriving one from the display name. A correct base URL cannot repair an unsupported or misspelled identifier.

§ Make retries part of the routing decision

The Python SDK retries certain connection failures and HTTP errors by default. Setting max_retries explicitly makes that behavior visible during a migration. A timeout after submission does not necessarily mean the upstream computation never started; do not assume every retried generation is harmless or unbilled.

Separate connection failures from long-running responses when tuning timeouts. For finer control, pass an httpx.Timeout object rather than a single number. For streaming, test first-byte latency, inter-chunk pauses, and interruption handling independently. Also check whether your application framework adds its own retries around SDK calls.

§ Verify the boundary without exposing credentials

Log the destination host, resource path, selected model identifier, latency, and status code. Capture request identifiers when available. Avoid logging authorization headers, full prompts, or complete responses by default. Review proxy settings too: a custom base URL does not bypass HTTP proxy configuration.

Reuse clients to preserve connection pooling, and close them at application shutdown. Use AsyncOpenAI for asynchronous request paths rather than blocking an event loop with the synchronous client.

For a controlled trial, create a Vana API account, configure a separate credential, and run a representative request through this dedicated client. Vana API bills per real token with no monthly fee, so measure token usage alongside latency and errors before moving production traffic.