MEGA Hub
New
Docs/Clients & operations

Clients and SDKs#

MEGA's public Router is OpenAI-compatible, so applications call it through an OpenAI SDK or raw HTTP with a MEGA token. The model page's Use this model dialog generates the same Python, JavaScript, and cURL requests.

Shared configuration#

All clients use:

Base URL: https://inference.tensorplay.cn/v1Bearer token: MEGA PAT with inference:run

Keep the token in an environment variable or secret manager:

BASH
export MEGA_TOKEN="mega_..."

Do not use a Provider key as MEGA_TOKEN. Save it through Custom Provider Keys and continue authenticating the Router with a MEGA PAT.

Request controls#

Use the model value and HTTP headers consistently across every client:

Control API contract
Provider selection Append :fastest, :cheapest, :preferred, or :provider-slug to the model ID.
Billing mode Set X-Mega-Inference-Billing to auto, routed, or byok.
Billing owner Set X-Mega-Bill-To to an eligible organization handle.
Session affinity Set X-Mega-Session-Id for Chat or Responses.
Streaming Set stream: true in the JSON request.

OpenAI Python SDK#

BASH
python -m pip install openai
PYTHON
import osfrom openai import OpenAI client = OpenAI(    base_url="https://inference.tensorplay.cn/v1",    api_key=os.environ["MEGA_TOKEN"],) chat = client.chat.completions.create(    model="mega/gpt-5.4-mini:fastest",    messages=[{"role": "user", "content": "Hello"}],)print(chat.choices[0].message.content)

Set MEGA request controls as default headers when needed:

PYTHON
client = OpenAI(    base_url="https://inference.tensorplay.cn/v1",    api_key=os.environ["MEGA_TOKEN"],    default_headers={        "X-Mega-Inference-Billing": "routed",        "X-Mega-Bill-To": "research-lab",    },)

Create separate clients when workloads use different billing owners or billing modes. This makes accidental cross-charging less likely.

MEGA Python InferenceClient#

Install the MEGA package:

BASH
python -m pip install megatensors
PYTHON
import osfrom megatensors import InferenceClient client = InferenceClient(    provider="preferred",    api_key=os.environ["MEGA_TOKEN"],    bill_to="research-lab",    headers={"X-Mega-Inference-Billing": "auto"},) result = client.chat.completions.create(    model="mega/gpt-5.4-mini",    messages=[{"role": "user", "content": "Hello"}],    max_tokens=64,)print(result.choices[0].message.content)

provider="auto" and provider="fastest" leave the model on the default Router strategy. cheapest and preferred append the corresponding model selection. Use the OpenAI SDK directly for the Responses API.

Raw HTTP#

Raw HTTP is appropriate for languages without an OpenAI client or when you need exact control over headers. Always set a request timeout, parse the OpenAI-compatible error object, and log Inference-Id.

See Router API Reference for endpoints and headers. Use the task guides for request examples:

Production checklist#

  • Pin dependencies and test the exact SDK version you deploy.
  • Use a fine-grained inference:run token for each workload.
  • Set X-Mega-Bill-To explicitly for organization workloads.
  • Bound input and maximum output tokens before dispatch.
  • Treat streaming retries as new billable requests.
  • Record model, task, billing mode, Inference-Id, and selected Provider without logging content.