Auto-Router: Tiers, Domains, and Providers
Leave the model picker on Auto to let Nyquest choose a model for your current request. Auto considers the task, your account tier, model availability, required capabilities and the complete request's token budget.
How Auto chooses
Nyquest first identifies the current task: General, Code, Creative, Reasoning or Vision. Older conversation text helps resolve short follow-ups, but a new topic can change the route. Image and tool requirements remain part of the request even when you supply a task hint. Optional semantic classification has a short deadline; if it cannot finish, routing continues with the rules result.
The task's complexity score ranges from 0 to 100. Complexity and account tier choose a routing tier; the score is not a model-quality rating.
| Tier | Auto behavior |
|---|---|
| T0 | Local inference for eligible simple free requests |
| T1 | Hosted models for simple requests and the free hosted route |
| T2 | More demanding work on funded accounts |
| T3 | Agent-class work where the account tier permits it |
| T4 | A ceiling for higher-priced explicit model choices; ordinary Auto does not select T4 |
Funded accounts skip T0. Before selecting a hosted Auto model, Nyquest checks fresh catalog availability, supported inputs, tools and response formats, account limits, and space for both input and output. It reserves output capacity instead of assuming compression will make an oversized request fit. If no eligible model remains, the request fails with an availability error.
Availability and fallback
Auto can try a bounded sequence of compatible models when a provider is unavailable, rate-limited or returns no usable answer. Streaming responses are checked before the first answer or tool output is exposed. A response already in progress is not restarted on another model.
Free mode stays on its separate free-capacity route, with any permitted rescue paid by the platform. Provider preferences still apply. NVIDIA-first routing is an optional operator setting, not a universal default. Models marked unavailable after a provider rejection remain disabled until reviewed; catalog synchronization does not automatically clear that flag.
Learning and compression
Quality-based model substitutions currently run in shadow. A candidate must have valid, recent, matched comparisons for the same task domain and routing tier before learning can change a live route. A newer model name alone does not qualify it.
The optional quality observer uses one configured judge, a concurrency limit and a separate platform spending budget. It preserves the supported evaluation request and excludes envelopes it cannot fairly grade. With no observer budget configured, new paid observations are paused. Old or incomplete scores do not qualify a model for promotion.
Auto resolves the model before applying compression. Typed context uses that model's limits and release controls. If a fallback changes the model, Nyquest starts from the original request and checks compression again. Savings depend on the workload and provider usage; Auto does not guarantee a lower bill on every request.
What you pay
Managed chat and media have no platform markup. Free-mode requests cost the caller $0. Usage and settlement use the applicable provider pricing and available receipts; a missing receipt is treated as unresolved. BYOK providers bill you directly.
Choose a model yourself
Selecting an explicit model keeps that model choice out of Auto's model substitution loop. Provider-level handling and the request's validation still apply. See Picking a Model, Free Mode, Multi-Model Splicer and BYOK Setup.