Switching from OpenAI to SovrGPT

Move an OpenAI application to SovrGPT: swap the base URL and API key, pick a model ID. With a model mapping, context limits and a Python example.

The SovrGPT API speaks the OpenAI schema. Two changes in your client are enough to switch: the base URL and the API key. Then pick a suitable model ID. Streaming, function calling, JSON mode and image input work as they do with OpenAI.

1. Set the base URL

Before (OpenAI)	After (SovrGPT)
`https://api.openai.com/v1`	`https://sovrgpt.com/api/v1`

In most SDKs: client.base_url = "https://sovrgpt.com/api/v1" or new OpenAI({ baseURL: "https://sovrgpt.com/api/v1" }).

2. Swap the API key

Replace the OpenAI key (sk-…) with a SovrGPT key (sov_…). You create it under Settings → API keys; details in Authentication.

You can also put the SovrGPT key into OPENAI_API_KEY and leave the code unchanged. In that case, make sure no other OpenAI calls share the same variable.

Pick a model ID

SovrGPT does not know OpenAI IDs such as gpt-4o and answers them with 404 model_not_found (details). Pick an ID from this table:

OpenAI	SovrGPT	Context	Notes
`gpt-4o`	`stackit-qwen3.8-27b`	262,144	Our standard Qwen: operated by STACKIT in Germany, no cold start, understands images. Self-operated: `qwen3.8-27b` (balanced, 131,072, with cold start).
`gpt-4o-mini`	`gemma-4-12b` (default)	262,144	Small and fast, with twice the context of `gpt-4o-mini`.
`gpt-4-turbo`	`stackit-qwen3.8-27b`	262,144	More thorough than `gemma-4-12b`.
`o1` / `o1-mini`	`qwen3.5-9b-deepseek-v4` (reasoning)	16,384	Chain of thought in `reasoning_content`. It counts against `max_tokens`; with a tight budget the answer comes back empty.
`gpt-4-vision-preview`	`gemma-4-26b-a4b` (vision)	262,144	Image input.
`text-embedding-3-small`, `text-embedding-3-large`, `text-embedding-ada-002`	`bge-m3`	–	The OpenAI IDs map to `bge-m3` (1024 dimensions), and that stays so. Existing vectors have to be regenerated. If you need long vectors with a selectable length (`dimensions`) as with `text-embedding-3-*`, use `stackit-qwen3-vl-embedding-8b` (4096, can be shortened to 256, 384, 768, 1024 or 1536, details).

Watch the context window: /v1/chat/completions does not reject an over-long request. It is truncated, and you get HTTP 200 with an answer that never saw the part of your prompt that was cut off. Before switching, check context_window in GET /v1/models and keep input + max_tokens below it.

The full catalogue is under Models.

Example: Python

 from openai import OpenAI

 client = OpenAI(
-    api_key=os.environ["OPENAI_API_KEY"],
+    api_key=os.environ["SOVR_API_KEY"],
+    base_url="https://sovrgpt.com/api/v1",
 )

 resp = client.chat.completions.create(
-    model="gpt-4o-mini",
+    model="gemma-4-12b",
     messages=[{"role": "user", "content": "Hello!"}],
 )

Other languages and libraries: SDK examples.

What differs

Images: /v1/images/generations does not exist yet (planned). You generate images with Z-Image, Qwen-Image and FLUX.2 in chat or through the MCP server today.
Speech: /v1/audio/speech and /v1/audio/transcriptions are available. Model IDs and voices differ from OpenAI's; see Audio.
Assistants API: probably will not be implemented. The product offers similar things through the marketplace and scheduled jobs, both only in the user interface.
Realtime API (WebRTC): not planned for 2026.

For these gaps you can keep using OpenAI alongside; then only those requests go to OpenAI.

Performance

Plan for cold starts: self-hosted models need several minutes after a pause (cold_start_hint in GET /v1/models). Before nightly jobs, send a small request a few minutes ahead, or use a partner-operated model (tensorx-…, stackit-…) without a cold start.
Use streaming: for longer answers, text appears right away.
Pick the right model: gemma-4-12b instead of qwen3.8-27b saves money and waiting time when you do not need the depth.
Set max_tokens: it prevents runaway answers and surprise bills.

Moving your data

OpenAI threads and assistants cannot be imported directly.
Embeddings have to be regenerated (different vector space).
System prompts work unchanged.

For larger migrations (roughly ten applications or more), get in touch; contact details are in the legal notice.