QueryFlow supports seven models in the Ask panel. Here's how we'd actually pick, by the kind of question you're asking, not by a leaderboard.
No credit card. 14 days. Cancel in one click.
Quick answer: There's no single best model for SQL. Use a fast, cheap model (Haiku 4.5, Flash-Lite) for quick counts and simple lookups, a stronger model (Sonnet 5, GPT-6 Sol) for schema exploration and multi-step reasoning, and Auto mode when you're not sure, it explores cheap and answers strong. We haven't run controlled benchmarks between them and won't pretend otherwise.
Every AI vendor publishes benchmarks showing their model winning at something. We haven't run our own controlled comparison of these seven models specifically on SQL generation against real schemas, and anyone telling you they have a definitive ranking for that is probably rounding up. What we can do is tell you how we'd actually pick, based on the kind of question you're asking and what it costs.
Broadly, each provider ships a stronger, slower, pricier model (Sonnet 5, GPT-6 Sol) and a faster, cheaper one (Haiku 4.5, Flash-Lite). Gemini 3.5 Flash sits in between its own family's two tiers. Grok 4.20 is the one option from xAI, no tiering to choose within.
A quick count against a table you already know well doesn't need your strongest model. Something like "how many orders came in today" is a one-table, one-filter query, and a cheap fast model handles it fine. Save the stronger models for the harder cases: exploring a schema you've never touched, joining across five tables you're not sure relate the way you think they do, or a question that needs a few steps of reasoning before it settles on the right query.
If you're the kind of person who has a favorite provider already, from other work, there's a reasonable argument for defaulting to whichever one you already trust and understand the quirks of. Familiarity with how a model tends to phrase things or handle ambiguity is worth something too.
Ask "how many active users do we have" against a users table with a status column, and pretty much any of the seven models writes the same obvious query correctly. Ask "why did weekly active users drop last month, broken down by acquisition channel and controlling for a product change that shipped mid-month," and now you're asking for actual reasoning across a join, a time comparison, and an implicit exclusion. That's where a stronger model earns its higher per-question cost, because getting the logic wrong on a multi-step question is more expensive than a wasted cent.
Auto mode routes the exploration work, reading table names, sampling columns, checking recent results, to a cheap, fast model, then hands off to a stronger model to write the final answer once there's enough context. For most day-to-day use this is the setting we'd actually recommend: you get the judgment of a stronger model on the part that needs it, without paying that rate for every step along the way.
You bring your own API key for whichever providers you use, and pay them directly, no markup from QueryFlow. Typical cost lands around 2 cents a question across these models. If you're asking dozens of questions a day, that adds up to real but modest money, a few dollars a month for most people, more if you're running heavy exploratory sessions constantly.
If you already know exactly the query you want to write, just write it, the Ask panel adds a step you don't need. And if a question depends on business context that isn't in your schema and hasn't been taught to the panel, no model is going to guess it correctly, see teaching it your business terms for how to fix that instead of hoping a different model does better.
We don't have controlled numbers, but a few patterns show up often enough in normal use that they're worth mentioning. Simple, well-scoped questions rarely differ much between models, whichever one you pick handles "count of X where Y" correctly almost all the time. The gap tends to widen on ambiguous phrasing, where one model asks a clarifying assumption implicitly (and might get it wrong) while another might explore a bit more before committing to a query. None of this is a substitute for reading the step cards and the SQL itself, regardless of which model wrote it.
If a model keeps missing the same kind of question, misreading which table you mean, or defaulting to an interpretation you don't want, switching to a different one is a reasonable first move before assuming the tool itself is the problem. Models have different tendencies, and a different one might simply phrase or interpret your question the way you expect on the first try. It costs you a model switch and a re-ask, not much else.
Two cents a question sounds trivial, and mostly is, but Auto mode's cheap-then-strong pattern exists because those cents add up differently depending on how you work. Someone running one or two questions a day barely notices the difference between models. Someone treating the Ask panel as a constant back-and-forth exploration tool, dozens of questions in a session, will notice the difference between an always-strong-model setup and Auto mode a lot faster, and that's really who Auto mode is built for.
If you want a single default and don't want to think about it further, Auto mode is the reasonable choice for most people most of the time, it's built specifically to avoid the tradeoff this whole page is about. Come back to picking a model manually only once you notice Auto mode consistently underperforming on a specific kind of question you ask often, at which point you'll know exactly which model to reach for instead.
The Haiku and Flash-Lite tiers are built for speed and low cost, which is why Auto mode uses one of them for exploration before handing off.
Whichever provider you already have an account and billing relationship with. The practical difference for typical SQL questions is smaller than the convenience of using a key you already manage.
It can add a small amount of latency since two models are involved instead of one, but it usually costs less overall since the cheap model handles the exploration.
No. We haven't run a controlled comparison and don't want to make up a ranking. Use the task-based guidance above and your own experience with each model.
Yes, the model switcher in the Ask panel lets you change at any time, including mid-session.
14-day free trial, no card. Try a few models against your own schema.
No credit card. 14 days. Cancel in one click.