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

# Model selection and updates

> Choose models by production availability, protocol, and task type, and identify recent updates.

The model catalog changes as routes and account pools evolve. Do not hardcode the models listed here as the only long-term choices. First retrieve the models available to the current key, then select one for the task.

## Evaluation order

1. Call [`GET /v1/models`](/en/api-reference/models) to confirm which models the current API key can call.
2. Read [`GET /api/pricing`](/en/api-reference/groups-and-pricing) to inspect model groups, billing multipliers, and capability fields.
3. Use `supported_endpoint_types` to choose the OpenAI-compatible or native Anthropic protocol.
4. Use `input_modalities`, `output_modalities`, and `capabilities` to evaluate vision, tool-calling, reasoning, and caching support.

<Warning>
  A vendor announcing a model does not mean the current DimiLinks key can call it. Treat the `/v1/models` response as the final authority.
</Warning>

## Available model summary as of 2026-07-12

The following model IDs come from the live DimiLinks pricing catalog and are not a permanent, fixed list.

| Task category                                       | Examples currently available                                                                                                        |
| --------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- |
| General chat through the OpenAI-compatible endpoint | `gpt-5.5`, `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra`                                                                           |
| Claude chat and tool use                            | `claude-sonnet-5`, `claude-opus-4-8`, `claude-haiku-4-5`                                                                            |
| Coding tasks                                        | `kimi-k2.7-code`, or a currently available GPT / Claude model selected through testing                                              |
| Chinese-language and general tasks                  | `qwen3.7-max`, `glm-5.2`, `MiniMax-M3`, `kimi-k2.6`                                                                                 |
| Low-cost batch tasks                                | `qwen3.6-flash`, `deepseek-v4-flash`, `claude-haiku-4-5`                                                                            |
| Image generation and editing                        | `gpt-image-2`                                                                                                                       |
| Video generation                                    | The dedicated `grok-imagine-video` video endpoint; consult the video pages and pricing catalog for current plan variants and prices |

## Highlights of this update

* GPT examples now use `gpt-5.5` and the three current GPT 5.6 variants.
* Recommended Claude examples now use `claude-sonnet-5` and `claude-opus-4-8`.
* Current catalog models including `kimi-k2.7-code`, `qwen3.7-max`, `glm-5.2`, and `MiniMax-M3` have been added.
* Legacy Grok text models and standalone `*-thinking` variants absent from the current catalog are no longer recommended.
* `gpt-image-2` is priced by resolution, so you cannot estimate every request with one fixed price.

## Production recommendations

* Refresh the model catalog at startup or on a schedule, and save a fallback model that uses the same protocol for each user.
* When caching `/api/pricing`, use `pricing_version` as the version key. Recalculate displayed prices whenever it changes.
* When a model is unavailable or returns `403 model_not_allowed`, retrieve the models and groups again. Do not silently redirect to a model with different capabilities.
* Before release, run a minimal request against every model your product will actually use. Presence in the catalog proves only that the configuration is visible; it does not replace an end-to-end call.