---
title: Models and usage
description: See which models the AI Gateway offers and their list prices, what Not available means, and follow each request and your spend over time.
section: AI
order: 4
---

Every gateway request names a model by its catalogue ID, which is the provider followed by the model, such as `openai/gpt-4.1-mini` or `anthropic/claude-sonnet-5-5`. The catalogue lists each model with its context window and the provider's list prices. Your organisation's own [private inference endpoints](/docs/gpu-instances) aren't in the catalogue and are named `private/<name>` instead.

## Where do I see the models and their prices?

Open **AI Gateway**, then **Models**. Each provider's models are listed with their ID, context window and the list price in US dollars per million input and output tokens. The same catalogue is public at `GET /v1/catalog/ai-models`, which also gives each model's `processing_region` and any `notes`, such as a date the provider retires it.

```bash
curl https://api.dekka.cloud/v1/catalog/ai-models | jq '.[] | {id, input_price_per_mtok, output_price_per_mtok, available}'
```

The prices are the providers' own, and Dekka uses them to work out each request's cost, the [monthly spend cap](/docs/ai-routing#set-a-monthly-spend-cap) and the spend in your metrics, so treat those figures as an estimate of the provider's bill. Dekka's own charge is the AI tokens your plan includes and bills, which [Plans, billing and usage](/docs/billing) covers.

## What does "Not available" mean?

The provider has retired the model, or Dekka has stopped offering it. It stays in the catalogue so older requests keep their model's name, but a request for it answers 404 `model_not_found`, the **Routing** page no longer offers it, and fallback skips it.

## See each request

**AI Gateway**, then **Requests**, lists every request, newest first, and you can filter it by project, model and outcome. Each row shows when it was made, the model that answered, the project, input and output tokens, latency, the cost at list price and the outcome. A request that fell back shows which model failed first, and one that didn't succeed shows its error code.

| Outcome | `status` | Meaning |
| --- | --- | --- |
| Succeeded | `succeeded` | A model answered |
| Failed | `failed` | Every model the gateway tried failed, or the stream broke off |
| Refused | `rejected` | A rule refused it before any provider was called, as [the order of checks](/docs/ai-routing#in-what-order-are-requests-checked) describes |

The API lists the same requests with `GET /v1/orgs/{org_id}/ai/requests`, which takes `project_id`, `model`, `status`, a `cursor` and a `limit` of up to 200.

## Follow usage over time

The **Overview** tab shows requests, input and output tokens, spend at list price and errors over the last 7, 14 or 30 days, by day and by model. `GET /v1/orgs/{org_id}/ai/metrics` returns the same figures for any range of up to 366 days, with `start` and `end` as UTC dates, `end` included, and the last 14 days when you send neither. Anyone in the organisation can see requests and metrics.

## Are prompts and replies stored?

No. Dekka records each request's model, token counts, timing, cost and outcome, and never stores the prompt or the reply. The prompt and the reply pass through the gateway to OpenAI or Anthropic, which process them in the United States, outside Ghana.