Model routing

Connect an AI account and never make another engineering decision about models. Routing chooses per call, and two invariants stop it ever trading your outcome for a cheaper bill.

The two invariants#

Both are asserted in code, not merely documented.

A human pin is final
If you set a model explicitly it is returned verbatim and the router is never consulted. A test proves it by replacing the router with a function that raises.
Difficulty may only raise the floor
Budget pressure makes easy work cheaper. It can never hand hard or retried work to a weaker model — that would trade completion for tokens. The router asserts this and falls back to the shipped default if it is ever violated.

How a model is chosen#

  1. Free modeResolved before the router is consulted at all, so a run promised to cost nothing never even reads the catalogue.
  2. A human pinReturned as-is.
  3. The routerScores what is available now against capability, live price, provider health and measured outcomes.
  4. The shipped defaultRouting may degrade. It may never fail.

Why price and quality are not in the catalogue#

The catalogue records what models exist and what each is capable of. It deliberately does not record what they cost or how good they are.

Price joins live from the pricing table, so a vendor changing its rates does not require a release. Quality is measured from outcomes rather than asserted — ship rate as a conservative lower bound, and median spend per task, shrunk toward a per-tier prior until there are enough observations to trust.

You are billed on the tier, not the vendor

Swapping the model behind a capability tier changes our margin, not your bill. The rate card on the pricing page is served by the same code that meters the charge.

Common questions

Can I force a specific model?

Yes, per role. An explicit pin is returned verbatim and the router is bypassed entirely.

What happens if a provider goes down?

The router derives a fallback chain and degrades to it. If everything fails it drops to the shipped default rather than erroring.

FeaturesToken intelligenceToken optimisation here never depends on a human noticing a chart. The system observes, diagnoses, fixes, validates and learns, so waste disappears on its own.FeaturesThe agentsTharas does not run a single agent repeatedly. Work is handed between roles with a defined job each, and every hand-off is a structured, schema-enforced output rather than prose parsed after the fact.

See the capability tiers and what they cost

The rate card is public and comes from the meter itself.

See pricing →