There are more than 300 AI models available through Nyquest right now — 334 as of this morning, and that number updates itself daily from the live catalog. That abundance is the point of Nyquest. It's also the problem we exist to solve: nobody should have to keep a mental spreadsheet of which model is best at code this month, which one reasons carefully, and which one just got quietly upgraded — or quietly worse.
That's the AutoRouter's job. And this week, it got a lot smarter.
Since launch, the AutoRouter has classified every request — how complex is it, what domain is it in, what's it worth spending — and matched it to the right tier and model. That worked well, but the model rankings underneath were relatively static: informed defaults, tuned by hand.
The new AutoRouter closes the loop. Behind the scenes, a sample of real answers gets independently graded by an AI judge on four axes — factuality, consistency, completeness, and hedging — and those grades now flow straight back into routing decisions. Three upgrades make that data genuinely useful:
1. Quality is tracked per model, per domain. The model that writes the best Python is not the model that writes the best product copy, and neither is automatically the best at step-by-step reasoning. The router now maintains separate scorecards for every model in every domain — code, reasoning, creative, vision, general — and routes your request based on how models actually perform at your kind of task, not on a single global average.
2. Recent performance counts more. Model quality isn't static — providers ship silent upgrades and silent regressions all the time. Scores are recency-weighted, with older results fading on a two-week half-life. When a model gets better (or worse), Nyquest's routing reflects it within days. Automatically.
3. Exploration runs on bandit math, not guesswork. Deciding when to try a promising-but-unproven model is a classic computer-science problem, and we now solve it the way large recommendation systems do: with upper-confidence-bound (UCB) bandit algorithms. If the router has a proven winner for your request, you get the proven winner. A small, strictly capped exploration budget goes to the candidates the system has the most to learn from.
And quality is only half the equation. When two models grade out as equals, the cheaper one wins. Quality first, cost second — that ordering is built into the selector.
The classifier that reads your request got a precision pass too. It now understands context well enough to know that "add a column to my dataframe" is a coding question, that "The Picture of Dorian Gray" is literature, and that "make it more efficient" refers to your function — while still recognizing a genuine "generate an image of…" instantly, including follow-up edits like "make it bigger" mid-conversation. Fewer misroutes means faster, cheaper, more accurate answers.
The AI ecosystem shifts daily — models get renamed, rate-limited, gated, or retired, often with no announcement. Two new autonomous systems keep Nyquest ahead of that churn:
No tickets, no maintenance windows. The platform notices before you do.
Nyquest's on-server local tier — which serves everyday lightweight requests at zero inference cost — now supports streaming chat, with first tokens in well under a second. Stacked with prompt compression and free-provider routing, it's one more layer in the answer to the question we care most about: how do we get you the best answer for the least money?
Every one of these changes serves the same idea: you shouldn't have to think about models. You bring the work; Nyquest brings a routing brain that's graded by an independent judge, tuned per domain, refreshed weekly by real-world performance, and honest about what's actually available — 334 models today, counted live.
The AutoRouter used to follow rules. Now it learns. And it's learning on every request.
Try Nyquest → No account needed to start chatting — or bring your own API key.