# OpenAI-Compatible API (Beta)

OpenAI-compatible inference lets you sample from any Tinker base model or checkpoint through two endpoints that work like the OpenAI API: [`/chat/completions`](https://tinker-docs.thinkingmachines.ai/inference/openai/chat-completions/index.md) and [`/completions`](https://tinker-docs.thinkingmachines.ai/inference/openai/completions/index.md). It is designed for testing your models, and you can also use it for production traffic:

- **Fast feedback**: Sample from a checkpoint as soon as training saves it, while the training run keeps going.
- **Testing and evaluation**: Run evaluations, scripts, and your own tools against your models.
- **Production traffic**: Serve your models to your own users. For higher or more stable throughput, or guidance on large-scale setups, [join our waitlist for production inference](https://form.typeform.com/to/CPTm26T7).

## Using the OpenAI client

Point any OpenAI SDK at `https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1` and authenticate with your Tinker API key. Set `model` to one of these:

- **A base model:** its Tinker ID from [Models & Pricing](https://tinker-docs.thinkingmachines.ai/tinker/models/models_and_pricing/index.md), such as `thinkingmachines/Inkling:peft:262144`.
- **Your checkpoint:** the path of a [checkpoint saved for sampling](https://tinker-docs.thinkingmachines.ai/tinker/howto/checkpoints/index.md), such as `tinker://…/sampler_weights/000080`. It runs on the base model it was trained on.
- **Your checkpoint with a different context length:** `<base model ID>;<checkpoint path>`, such as `thinkingmachines/Inkling:peft:262144;tinker://…/sampler_weights/000080`. The base model must be the model that the checkpoint was trained on, with the context length you want.

```py
from os import getenv
from openai import OpenAI

MODEL = "thinkingmachines/Inkling:peft:262144"

client = OpenAI(
    base_url="https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1",
    api_key=getenv("TINKER_API_KEY"),
)

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What is 17 * 23?"}],
)
print(response.choices[0].message.content)
```

Use [Chat Completions](https://tinker-docs.thinkingmachines.ai/inference/openai/chat-completions/index.md) for chat and instruction-tuned models. Use [Completions](https://tinker-docs.thinkingmachines.ai/inference/openai/completions/index.md) for pretrained base models, or to send your own token IDs.

## Sending requests without the OpenAI SDK

Send JSON to the same URLs, with your Tinker API key as a bearer token:

```bash
curl https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1/chat/completions \
  -H "Authorization: Bearer $TINKER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "thinkingmachines/Inkling:peft:262144",
    "messages": [{"role": "user", "content": "What is 17 * 23?"}],
    "reasoning_effort": "high"
  }'
```

`extra_body` is an option of the OpenAI SDK, not a field of the HTTP API: the SDK merges its keys into the top level of the request. In raw JSON, put fields such as `top_k` and `reasoning_effort` at the top level. The server ignores unknown fields with no error, so anything inside an `extra_body` key is lost.

## Related docs

- [Chat Completions](https://tinker-docs.thinkingmachines.ai/inference/openai/chat-completions/index.md) and [Completions](https://tinker-docs.thinkingmachines.ai/inference/openai/completions/index.md): the two endpoints, with examples and a full reference.
- [Anthropic-Compatible API](https://tinker-docs.thinkingmachines.ai/inference/anthropic/index.md)
- [Getting a `TINKER_API_KEY`](https://tinker-docs.thinkingmachines.ai/tinker/quickstart/index.md)
- [Security and Privacy](https://thinkingmachines.ai/legal/terms/)
- [SamplingClient API Reference](https://tinker-docs.thinkingmachines.ai/tinker/api-reference/samplingclient/index.md)