Does it recommend a specific model name?

It recommends a tier — frontier, mid-tier, small/mini, or self-hosted open weights — and names current representative models in each. Tiers are more durable than exact model names, which change every few months, so the guidance stays useful.

How does privacy affect the recommendation?

A strict privacy or data-residency requirement pushes the wizard toward self-hosted open-weight models or providers with zero-retention enterprise terms, even if a hosted frontier model would otherwise win on accuracy.

Why does latency matter so much?

Frontier reasoning models are accurate but slow and expensive per call. If you need sub-second responses for an interactive product, the wizard favors smaller, faster models and suggests reserving frontier calls for hard sub-tasks.

Is the recommendation a guarantee?

No. It is a strong starting point based on common trade-offs. Always confirm with your own evaluation set, because real performance depends on your exact prompts, data, and quality bar.

What if cost and accuracy conflict?

The wizard weighs your stated ceiling against your accuracy requirement and explains the tension explicitly, often suggesting a routing strategy — a cheap model for easy cases, an escalation to a frontier model for hard ones.

AI Model Selection Wizard

AI model selection wizard

Choosing an LLM is a multi-variable trade-off: accuracy versus cost, speed versus capability, hosted convenience versus data control. This wizard asks seven focused questions and scores the main model tiers against your answers, then recommends the best fit with the reasoning behind it — so you stop guessing and start from a defensible choice you can validate on your own data.

How it works

You answer questions about task type, required accuracy, latency budget, privacy needs, monthly call volume, and your cost ceiling. The wizard applies a weighted scoring model: hard constraints like strict privacy or tight latency can veto a tier outright, while soft preferences shift the balance. It then surfaces the winning tier, the runner-up, and a clear explanation of the trade-offs. All scoring runs locally in your browser; nothing is sent anywhere.

Tips and notes

Tiers beat names. Exact model names churn; the frontier / mid / mini / open-weight framing stays stable, so map the recommendation to whatever is current at your provider.
Privacy and latency are vetoes. When either is strict, accept that the very top accuracy tier may be off the table.
Consider routing. If cost and accuracy conflict, run a cheap model first and escalate only the hard cases — often cheaper than one big model everywhere.
Always evaluate. Treat the recommendation as a hypothesis and confirm it on a real eval set before committing production traffic.