# Managed AI

> The models ShopAssist runs for you on EU infrastructure, how routing and fallback work, and the no-training guarantee.

Source: https://shopassist.coec.dk/docs/ai/managed · Updated 2026-10-01

With managed AI, ShopAssist runs the model for you. Requests go only to EU endpoints of providers listed on the [sub-processors page](/subprocessors), under agreements that forbid training on your data.

## Models [#models]

| Model | Id | Tier | Context window | Images | Processing |
| --- | --- | --- | --- | --- | --- |
| GPT-6 Luna (Azure EU Data Zone) | `gpt-6-luna` | Standard | 1,050,000 tokens | Yes | EU |
| GPT-6 Sol (Azure EU Data Zone) | `gpt-6-sol` | Premium | 1,050,000 tokens | Yes | EU |
| Claude Sonnet 5 (Amazon Bedrock EU) | `anthropic.claude-sonnet-5` | Premium | 200,000 tokens | Yes | EU |
| Mistral Small 4 (EU) | `mistral-small-2603` | Standard | 256,000 tokens | No | EU |
| Claude Sonnet 5 (Vertex AI EU) | `claude-sonnet-5` | Premium | 200,000 tokens | Yes | EU |
| Gemini 3.1 Flash-Lite (Vertex EU) | `gemini-3.1-flash-lite` | Standard | 1,048,576 tokens | Yes | EU |

The standard route uses the first available model of the list; premium models are available on plans with premium models (see [plans and limits](/docs/getting-started/plans-and-limits)).

## Routing and fallback [#routing]

If a provider is slow or unavailable, the next one in the route answers. Availability wins over model identity: a premium conversation answered by a standard model during an outage counts as standard. You don't need to configure anything.

## No training [#no-training]

Neither COEC nor its managed providers train or fine-tune models on your shoppers' conversations or your store data.

## Retention [#retention]

Managed providers keep no prompts or completions beyond what is needed to answer, under zero-data-retention or modified abuse-monitoring terms where the provider offers them. Details per provider: [data location and retention](/docs/ai/data-location-and-retention).