Which AI model should you use?

Question Everything runs several AI models in production, every day. This is our honest, evidence-rated answer — including the places where the honest answer is “this is genuinely contested,” not a confident ranking we can’t back up.

Every claim below is rated by how confident we actually are, and why — the same discipline QE applies to every claim on every knowledge page.

By task

Everyday questions, fast turnaround

High confidence

A fast, low-cost model (e.g. Gemini Flash tier)

For most questions, the gap between a fast, cheap model and the most expensive frontier model is small relative to the cost difference — often 10-25x cheaper per query for output that's very close in practical usefulness.

Why this confidence level: Directly measured — this is why QE itself defaults to a fast, low-cost model as its primary provider, not its most expensive one.

Complex, multi-step reasoning

High confidence

A dedicated reasoning-tier model, deliberately

Reasoning models spend part of their response budget on internal deliberation before producing visible output — genuinely different behavior from a standard model, not just "the expensive version." This makes them better suited to problems that benefit from that deliberation, and a poor fit for simple lookups where the reasoning overhead adds cost and latency without adding value.

Why this confidence level: Directly observed — see Field Notes for a real incident this caused when the difference wasn't accounted for in advance.

Code generation and technical work

Medium confidence

Varies — this is genuinely contested, not settled

Different model families are widely reported to lead on coding benchmarks at different times, and benchmark performance doesn't always translate cleanly to a specific real codebase or task. QE doesn't have enough first-party comparative data across a large enough sample of real coding tasks to make a confident universal claim here.

Why this confidence level: This reflects the current state of public benchmarks and developer discussion, not QE's own measured comparison — stated honestly rather than asserted as fact.

Cost-sensitive, high-volume applications

High confidence

The cheapest model that reliably clears your quality bar — test, don't assume

Per-token cost varies by roughly 60x across the models QE runs today (see the cost table below) — cheapest to most expensive. The right choice depends entirely on whether the cheaper model's output quality is good enough for your specific use case, which is task-specific, not a general fact about the models.

Why this confidence level: Directly measured from QE's own production billing.

What models actually cost

Approximate, blended per-million-token rates across the tiers QE runs — rough by design, not implied precision we don’t have.

Tier~Cost / 1M tokensNote
Fast, low-cost tier (e.g. Gemini Flash)~$0.15Cheapest tier QE runs — the default for most traffic
Efficient mid-tier (e.g. GPT-4.1 mini)~$0.40Good balance of quality and cost
Reasoning tier (e.g. GPT-5.6-class)~$2.00Spends part of its budget on internal reasoning — see Field Notes
Full frontier tier (e.g. GPT-4.1 full)~$4.00Deepest single-pass analysis, highest cost
Premium tier (e.g. Claude Sonnet-class)~$10.00Most expensive tier QE runs — roughly 60x the cheapest

Want the real incidents behind this guidance — what actually broke, and how it got fixed?

Read Field Notes