Auto-Router: Tiers, Domains, and Providers

Leave the model picker on Auto and Nyquest routes each message to the model that fits it — by difficulty, by subject matter, and by live price. This article explains what the router actually does.

Tiers: T0–T4

Every message gets a complexity score (0–100). The score maps to a price tier, and tiers map to models by their real, live catalog price — not by name-matching, so a model that changes price changes tier automatically.

T0 runs on Nyquest's own hardware, and serves the free tier. Simple auto-routed tasks from free and anonymous users are served by a local model on the platform itself — zero upstream cost, full streaming support. Funded accounts skip T0: if you're paying for the router's judgement, Simple tasks go to a full T1 model, not the local one. If the local model is down or cold, requests fall through to T1 before a single byte is sent.

Domains: the router reads the subject

Beyond difficulty, the router classifies what kind of task you sent — Code, Creative, Reasoning, Vision, or General — and picks a model with demonstrated strength in that domain (code goes to a code-strong model, vision to a multimodal one, and so on). Detection is rule-based first, with a semantic classifier that catches prompts using no obvious keywords.

Intent detection is deliberately conservative: asking to "add a column" or discussing The Picture of Dorian Gray won't misfire into image generation — media routing requires strong, unambiguous intent.

The router is instrumented to learn

Today, routing follows measured task complexity and your tier — deterministic and predictable. Underneath, the platform samples a share of live traffic, quietly fans those prompts to alternate models, and has a cross-family judge panel score every answer on multiple axes. Those judgments accumulate into a per-model, per-domain quality record with recency weighting. That record runs in shadow: it logs what it would choose, and it only earns a live vote in routing once a candidate has enough judged samples to clear an evidence bar — no request is ever risked on an unproven model, and we would rather tell you the router is still earning that evidence than claim learning it can't yet show.

The catalog itself is self-maintaining: models that upstream providers list but can't actually serve are detected and dropped automatically, and restored automatically when they come back.

Providers: free-first serving

The same model is often available from more than one upstream provider. Nyquest's provider layer routes free-capacity-first — if a model can be served from a $0 provider pool, it is, and the savings are real (attributed per-request in your usage data). Each provider is wrapped in an adaptive circuit breaker: a few consecutive failures take it out of rotation briefly and traffic fails over gracefully, so provider flakiness never becomes your problem.

You can steer this:

• Preferred Provider — a setting that biases routing toward the provider you choose
• Benchmarks — Settings → Benchmarks shows live provider comparisons so your preference can be evidence-based

What you pay

Managed (wallet-funded) chat is billed from resolved-model live pricing with no platform markup — you pay the provider’s own catalog rate. A tested invariant guarantees a request is never billed below what it cost us to serve. BYOK traffic has no markup either: your provider bills you directly. Free-mode and T0 traffic cost $0.

Overriding the router

Auto is a default, not a cage. Pick any model manually (Picking a Model) and the router steps aside for that conversation. For high-stakes prompts, consider fanning out to several models at once with the Multi-Model Splicer.

Where to next

• Free Mode — routing when your balance is $0
• Multi-Model Splicer — parallel multi-model consensus
• BYOK Setup — route through your own provider keys