Model Router | Kimchi Inference

The Model Router picks the best model for each request automatically. Instead of naming a specific model, you send auto as the model name. The Model Router examines your prompt and routes the request to the model that fits the task best.

The auto model works like any other model in the model APIs. You keep the same endpoint, API key, and request format. Any client that can send a model name can use auto.

How it works

%%{init: {'theme': 'dark'}}%%
flowchart LR
    A["Your client"] -->|"model: auto"| B["Model Router"]
    B --> C["Best-fit model"]
    style B fill:#2d2d2d,stroke:#888,color:#e0e0e0

Each request moves through four steps:

Your request names auto as the model.
The Model Router scores your prompt against the candidate models.
The request runs on the model with the best task fit.
Billing and reports record the model that served the request.

The Model Router picks only from the models in the Kimchi serverless model catalog. See Available models for the list.

How the Router picks a model

The Model Router contains a small classification model. It was trained on which models solve which tasks well. For every request, it estimates how well each candidate model fits the prompt and picks the strongest one.

The candidate models are the serverless models that are enabled for your organization and that the Router supports. The supported set grows as Kimchi trains new models into the Router.

Three rules refine the decision:

Image input. If the request carries images, only models that support vision remain candidates.
Reasoning effort. For models with adjustable reasoning, the Router sets the reasoning effort that matches the task.
Very long prompts. Prompts above a length threshold skip the scoring step and go directly to a fast default model.

If scoring fails, the Router falls back to the same default model.

Using the auto model

The auto model is enabled for all users in your Kimchi organization by default.

auto is the default model for new sessions in the Kimchi Harness and in Code mode in Kimchi Studio. The Model Router picks the model for each request these sessions send.

You can switch from auto to any other supported model at any time:

Kimchi Harness. Run /model from the Kimchi prompt and pick a model from the list. Pick auto again to return to automatic routing.
Studio Code mode. Select the model button next to the chat input and pick a model. The Auto (Kimchi Router) option routes each part of the task automatically. Set the default model for new code sessions under Settings → Code.

The auto model appears in your organization's model list like any other serverless model. The auto-beta model name still works in API requests but is retired from the model list. Use auto in new requests.

Pricing

The auto model has no token price of its own. Each request is billed at the per-token rates of the model that served it. The Model Router adds no markup.

See the Pricing reference for per-model rates.

Router analytics

The Kimchi console includes a Router page. It shows what the Model Router did for your organization over the selected time range.

Section	What it shows
Usage	The share of all requests that were routed through `auto`.
Adoption	The share of active users who send at least 20% of their requests through `auto`, and the number of these users.
Estimated savings	Your Auto spend compared with the same traffic on a baseline model. The baselines are Claude Opus 5.5 (default) and Claude Sonnet 5.5.
Router cost impact	Two user groups: those who route through `auto` and those who do not. Each group shows the average cost per user.
Model selection and switching	The models `auto` picked over time, plus the models users switched to manually after Auto requests, with request counts.
Usage by harness	Requests and users per harness, such as Claude Code or Kimchi.
Teams and users	A sortable table of Auto share, estimated savings, and Auto requests per team or per user.

The estimated savings figure prices your Auto token volumes at the baseline model's list price, then subtracts your actual Auto spend. Select the baseline model above the chart to compare against a different premium model.

The Router page also lists the per-user detail behind the cohorts. Open a group in Router cost impact to see each user's routed request share, estimated savings, and team.

Auto in other reports

The Usage report and Adoption report include Auto traffic in their charts and model breakdowns.

Both reports have a Show Auto separately switch:

On (default). All requests sent to auto appear as one auto entry.
Off. Each request counts toward the model that served it. The auto entry disappears.

In the Adoption report, the top models table lists auto like any other model. Expand the auto row to see By model Auto picked. Each picked model shows its cost and the number of Auto requests it served. Expand a picked model to see its token split.

In the Usage report, the model breakdowns per user and per API key show auto as its own model when the switch is on. Expand a model row to see the split by token type.