Build on GPUFlow.

Endpoints

POST

/v1/chat/completions

Main endpoint compatible with the OpenAI SDK. Supports streaming, function calling and all catalogue models.

POST

/v1/messages

Endpoint compatible with the Anthropic SDK. Same functionality, Anthropic message format.

GET

/v1/models

Lists available models with pricing and capabilities.

POST

/v1/audio/transcriptions

Upload an audio file and get the text back. Multipart; with response_format=verbose_json it also returns the duration, which is what gets billed.

POST

/v1/audio/speech

You send text and a voice, and get the audio back. Billed per character.

POST

/v1/music/generations

Generates a whole song. It is asynchronous: it returns an id, you poll the status and download the track when it finishes.

Streaming and reasoning

{ "model": "glm-5.3-nvfp4",
  "messages": [{ "role": "user", "content": "Resuélvelo paso a paso" }],
  "stream": true,
  "reasoning": { "effort": "high" } }

Function calling

5 of the 9 text models support tools, in the standard OpenAI format. You pass your tools array and the model decides when to call them; you run the function and return the result. In the catalogue below they carry the “Tools” capability.

{ "model": "glm-5.3-nvfp4",
  "messages": [{ "role": "user", "content": "¿Qué tiempo hace hoy?" }],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "parameters": {
        "type": "object",
        "properties": { "city": { "type": "string" } }
      }
    }
  }] }

Your first call, in your language

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.gpuflow.ai/v1",
    api_key=os.environ["GPUFLOW_API_KEY"]
)

response = client.chat.completions.create(
    model="glm-5.3-nvfp4",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=10000
)

print(response.choices[0].message.content)

Error codes

401

Invalid or missing API key

402

Insufficient balance, top up from your dashboard

403

You do not have permission to use this model

404

Model not found

429

Rate limit exceeded, wait and retry (Retry-After header)

503

Model temporarily overloaded, retry with backoff (Retry-After header)

Model catalogue

All models run on our infrastructure in Spain. Prices per million tokens.

Voice models are billed per minute or per character, and music per song. The detail for each one is on its own page. Model catalogue.

Prices in EUR per million tokens, VAT included.

1. Generate your key

Three commands and it is done. Do not give it a path with -f: OpenSSH looks for ~/.ssh/id_ed25519 on its own when jumping through the bastion, and a key somewhere else is the usual reason access fails.

# 1. create the keys folder
mkdir -p ~/.ssh && chmod 700 ~/.ssh

# 2. generate your key (Enter to leave it without a passphrase)
ssh-keygen -t ed25519 -C "gpuflow"

# 3. print it so you can copy it
cat ~/.ssh/id_ed25519.pub

Paste the PUBLIC key, the one ending in .pub. The private one never leaves your machine.

2. Connect

One command, no tunnels and no second terminal. The -J jumps through the bastion, which identifies you with the same key you registered and takes you to your Sandbox.

ssh -J [email protected] developer@mi-sandbox

3. What the bastion will not forward

Only the tunnel to your Sandbox: no agent forwarding, X11, environment variables or port forwarding from the bastion. And if the Sandbox is paused or deleted it rejects you with a message instead of leaving you waiting on a machine that is not there.