Migrating from OpenAI

As a rule, three changes are all it takes:

Change baseURL.
Swap the API key.
Replace the model ID.

That is it. Streaming, function calling, JSON mode, vision — all of it works.

1. Set `baseURL`

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 OpenAI keys (sk-…) with SovrGPT keys (sk-sovr-…). Create them under Settings → API keys; see Authentication.

Tip: you can simply overwrite the OPENAI_API_KEY environment variable with the SovrGPT key and change nothing in the code. But do make sure the same variable is not still shared with actual OpenAI calls somewhere.

3. Replace the model ID

OpenAI	SovrGPT equivalent	Context	Notes
`gpt-4o`	`qwen3.8-27b` (balanced)	131,072	The one-to-one replacement: the only own tier whose context window is in the same order of magnitude as `gpt-4o` (128k). Without a cold start: `tensorx-qwen3.8-27b`.
`gpt-4o-mini`	`gemma-4-12b` (default)	262,144	Both small and fast — here the window is in fact twice that of `gpt-4o-mini`.
`gpt-4-turbo`	`qwen3.8-27b` (balanced)	131,072	Deeper than gpt-4o-mini.
`o1` / `o1-mini`	`qwen3.5-9b-deepseek-v4-flash` (reasoning)	16,384	Chain of thought in `reasoning_content`. ⚠️ The thought counts against `max_tokens` — with a tight budget the answer comes back empty.
`gpt-4o-vision-preview`	`gemma-4-26b-a4b` (vision)	262,144	Multimodal.

🔴 The context window is where a migration goes wrong silently. /v1/chat/completions does not reject an over-long request — it is truncated upstream, and you get HTTP 200 with an answer that never saw the part of your prompt that was cut off. Check your target model's window before switching, via context_window in GET /v1/models, and make sure input + max_tokens stays below it.

We refresh the exact mapping recommendations with every model refresh in Models.

Example: the diff for a Python app

 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!"}],
 )

Three lines. Nothing else.

What is not available one-to-one?

Embeddings — we do not have text-embedding-3-* in v1 yet. Planned for Q3/2026.
DALL·E 3 — we offer Z-Image and FLUX.2 in the chat UI; a dedicated /v1/images/generations endpoint follows in Q3/2026.
Whisper / TTS — not in the current v1, planned for Q4/2026.
Assistants API — probably will not be implemented; use our marketplace plus scheduled jobs instead.
Realtime API (WebRTC) — not on the 2026 roadmap.

For each of these we recommend a hybrid approach: SovrGPT for chat completions (80 % or more of the load), OpenAI only for the specific missing features (20 % or less). That keeps the bulk of your data inside the EU.

Performance tuning

Avoid cold starts: for production cron jobs (nightly, say) send a small warm-up request 30 seconds before the real load.
Use streaming: for longer answers, SSE improves perceived latency dramatically.
Pick the right model: the default tier instead of premium saves money and latency when you do not need the depth.
Set a token limit: set max_tokens explicitly — it prevents runaway answers and surprise bills.

Data migration

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

For larger migrations (10 apps or more), get in touch at kontakt@e-networkers.de — we offer a free one-hour migration consultation.