Docs Product

How it works

Follow your request from an auto mode to an eligible model.

On this page

Learn how Eco turns your request and routing preference into a model choice.

A request in plain language#

You provide an Orbio key, a message, and an auto mode. Eco first checks which models support the request. It then estimates the task's difficulty and chooses the cheapest eligible model in the selected tier.

Eco system flow

Preparing diagram…

Diagram source

eco@docs mermaid

flowchart TD
    Request["Your message + Orbio key + auto mode"] --> Needs["Check what the task needs"]
    Needs --> Rules["Use routing rules"]
    Rules -->|Clear decision| Pick["Choose cheapest eligible model in the selected tier"]
    Rules -->|Uncertain| Classifier["Ask a small model to classify the task"]
    Classifier --> Pick
    Pick --> Orbio["Orbio runs the selected model"]
    Orbio --> Reply["Your answer + route headers"]

For example, a short rewrite may fit a cheap model. A complex reasoning task may need a more capable tier. Image input, tools, and structured output can limit which models are eligible.

Choose a preference#

Mode	What it means
`auto` or `auto/balanced`	Let the routing decision stand. A good starting point.
`auto/cheap`	Favor a lower-cost tier when the decision is ambiguous.
`auto/best`	Favor a more capable tier when the decision is ambiguous.
An exact model id	Use that model directly, without automatic selection.

Modes are preferences, not fixed tiers. A clearly easy task can stay in the cheap tier even with auto/best. A clearly hard task can stay in the best tier with auto/cheap.

Rules first, classifier when needed#

Rules look at signals such as message length, code, and the type of task. A clear decision does not need another model call.

When the rules are uncertain, Eco asks a small model through Orbio to classify a conversation excerpt. This call uses your key and can spend credits. If it times out or fails, Eco uses a balanced fallback while still checking required capabilities.

If no catalog model supports the request, Eco returns an error instead of selecting an incompatible model.

Preview, then send#

Preview shows a routing decision without generating the requested answer. It can still run the paid classifier. Send makes a separate routing decision and runs the completion.

Preview does not reserve a model. Using both can run the classifier twice, and the selected model can differ. The chat response's route headers describe the completion you actually ran.

See the choice#

The playground shows the selected model, tier, reason, and whether the classifier was used. API clients can read these details in the x-eco-* response headers. See API quickstart.

The price estimate is the selected model's seed output price per one million tokens. It excludes input tokens and classifier costs. It is not a final bill or a guaranteed saving.

Next: Architecture or Installation & usage.