From OpenAI to FlowModels.
If you use the OpenAI SDK, you change the base URL and the key and write the model as provider/model.
What changes
The base URL, the key and the model name. Everything else stays, as long as you do not use any of the fields in the table below.
| Setting | Before | On FlowModels |
|---|---|---|
| Base URL | https://api.openai.com/v1 |
https://api.flowmodels.one/v1 |
| Key | OPENAI_API_KEY |
FLOWMODELS_API_KEY |
| Model name | gpt-5-mini |
openai/gpt-5-mini |
The code before and after
from openai import OpenAI
client = OpenAI(
- api_key=os.environ["OPENAI_API_KEY"],
+ base_url="https://api.flowmodels.one/v1",
+ api_key=os.environ["FLOWMODELS_API_KEY"],
)
response = client.chat.completions.create(
- model="gpt-5-mini",
+ model="openai/gpt-5-mini",
messages=[{"role": "user", "content": "Hello!"}],
)
Fields that behave differently
What the chosen model supports decides. The fields it accepts are listed under supported_parameters in GET /v1/models.
| Field | Behavior on FlowModels |
|---|---|
n, logprobs, top_logprobs, audio, store, metadata, previous_response_id, background, conversation, webhook_endpoint_id |
On /v1/chat/completions the API answers with 400 unsupported_field instead of quietly ignoring them. On /v1/responses, background, store, conversation, webhook_endpoint_id and metadata are allowed. |
max_tokens, max_completion_tokens |
Both spellings work. Without a value the maximum of the model applies, and the highest possible amount is reserved before the call. |
temperature, top_p |
Models with fixed sampling values answer these fields with 400 unsupported_for_model. Check supported_parameters. |
response_format |
text, json_object and json_schema. If the model cannot produce the shape, you get 400 unsupported_for_model. If the word “json” is missing from the messages with json_object, FlowModels adds a short hint. |
tools, tool_choice |
Type function only. FlowModels does not run your functions; the model reports tool_calls. Built-in tools such as web_search or code_interpreter do not exist here as a tool type; on models that support it, you turn on web search with web_search: true. |
stream_options |
Only include_usage is allowed, and only together with stream: true. |
What to check after the move
-
The new key is in
FLOWMODELS_API_KEY, andGET /v1/modelslists your model. -
A call without streaming returns
choices[0].message.contentas expected. -
With
stream: truethe events arrive; withstream_options.include_usagethe usage is in the last event. -
Tool calls and
response_formatwork with exactly the model you use. -
Your error handling separates
429 insufficient_quota(credit exhausted) from429 rate_limit_exceeded,429 concurrent_requests_exceededand503 provider_overloaded. - The request log in your dashboard shows the model, tokens and cost of the call.
- OpenAI text models are in the catalog from GPT-5 on. If you use older ones, pick a model from the list before you move.
-
You generate images through
POST /v1/images/generationsin the OpenAI format, only with the sizessquare,portrait,landscapeand2kinstead of pixel values and with exactly one image per request. FlowModels does not offer batch, assistants or fine-tuning. Those calls stay with OpenAI.