> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wandb.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Neon AI Gateway

> Trace calls to Neon AI Gateway, the OpenAI-compatible inference endpoint provided by Neon

This guide shows you how to use Weave to automatically trace calls to models served by Neon AI Gateway, so you can monitor, debug, and evaluate model usage from a single dashboard.

[Neon AI Gateway](https://neon.com/docs/ai-gateway/overview) is an OpenAI-compatible inference endpoint provided by Neon. A single Neon credential reaches models from OpenAI, Google, Meta, Databricks, and Alibaba, with no provider API keys. Weave detects the OpenAI SDK, so existing OpenAI code works after changing the API key and base URL.

<Note>
  Neon AI Gateway is in beta. It requires a paid Neon plan with prepaid credits and a project in a supported AWS region. See [Neon AI Gateway](https://neon.com/docs/ai-gateway/overview) for detailed requirements.
</Note>

## Prerequisites

Unlike most providers, Neon does not have one shared hostname. Each database branch gets its own gateway host, so you need two values:

* **A credential** with the `ai_gateway:invoke` scope. Create it in the Neon Console under **Credentials**, or through the Neon API. See [AI Gateway authentication](https://neon.com/docs/ai-gateway/authentication).
* **The branch host**, shown in the Neon Console as `NEON_AI_GATEWAY_BASE_URL`. It is a per-branch
  URL of the form `https://<your-neon-branch-host>`.

Running `neon env pull --file .env` writes both as `NEON_AI_GATEWAY_TOKEN` and `NEON_AI_GATEWAY_BASE_URL`.

## Trace a Neon AI Gateway call

Set `api_key` to your Neon credential, set `base_url` to the branch host plus `/v1`, and use a short Neon model ID such as `gpt-5-mini`. `weave.init()` requires a project name for your traces. You can optionally prefix it with a W\&B entity as `<entity>/<project>`; if you omit the entity, Weave uses your default entity. To find or update your default entity, refer to [Default team](/platform/app/settings-page/user-settings/#default-team).

```python lines {5,10-13} theme={null}
import os
import openai
import weave

weave.init('neon-weave')

system_content = "You are a travel agent. Be descriptive and helpful."
user_content = "Tell me about San Francisco"

client = openai.OpenAI(
    api_key=os.environ.get("NEON_AI_GATEWAY_TOKEN"),
    base_url=f"{os.environ.get('NEON_AI_GATEWAY_BASE_URL')}/v1",
)
chat_completion = client.chat.completions.create(
    model="gpt-5-mini",
    messages=[
        {"role": "system", "content": system_content},
        {"role": "user", "content": user_content},
    ],
    temperature=0.7,
    max_tokens=1024,
)
response = chat_completion.choices[0].message.content
print("Model response:\n", response)
```

Weave captures the call as a trace in your project, including the model ID, messages, and the token counts Neon returns.

## Trace across branches

A Neon credential is valid on the branch it was created on and on every branch descended from it, so a credential created on `main` also works in preview and CI branches forked from it. Only `NEON_AI_GATEWAY_BASE_URL` changes between environments.

Because the branch host lives in the client configuration rather than in the request, traces from different branches look the same in Weave. Pass separate project names to `weave.init()`, or attach the branch as an attribute, if you want to tell them apart:

```python theme={null}
with weave.attributes({"neon_branch": "preview/feature-x"}):
    chat_completion = client.chat.completions.create(
        model="gpt-5-mini",
        messages=[{"role": "user", "content": user_content}],
    )
```

## Choose a model

Neon uses short model IDs like `gpt-5-mini`, `gemini-3-flash`, `llama-4-maverick`, and `qwen3-next-80b-a3b-instruct`. List what a branch can serve:

```bash theme={null}
curl "${NEON_AI_GATEWAY_BASE_URL}/v1/models" \
  -H "Authorization: Bearer $NEON_AI_GATEWAY_TOKEN"
```

Context windows and prices are in the [Neon model catalog](https://neon.com/docs/ai-gateway/models), also published as the [`neon` provider on Models.dev](https://models.dev/providers/neon/).

Two constraints affect which model you pick:

* A few models are served only on Neon's Responses API path, `{NEON_AI_GATEWAY_BASE_URL}/openai/v1`, and return a `400` on chat completions. The Endpoints column in Neon's [model catalog](https://neon.com/docs/ai-gateway/models) marks which ones, and the set changes; at the time of writing it is `gpt-5-3-codex` and `gpt-5-5-pro`. Every model the column lists with `chat/completions` works on the chat completions path.

Neon does not return a cost field and reports `pricing` as `null` in `GET /v1/models`, so traces show token counts without cost. AI Gateway usage draws down your prepaid credit balance; see [AI Gateway pricing](https://neon.com/docs/ai-gateway/overview#pricing).

See the [OpenAI integration guide](/weave/guides/integrations/openai#track-your-own-ops) for more details about integrating Weave with your own functions for more complex use cases.