Model Grand Prix

Model Grand Prix

Five models race the same task, and you pick what winning means. Drag three weights — accuracy, cost, and speed — and the scoreboard blends every model into a single 0–1 score, then waves the flag. The moat is what it reveals: the fastest car, the cheapest car, and the most accurate car are three different models, and the one that wins your race is usually none of them. It turns the invisible economics of an API call into a ticking meter you can re-weight live — never a static pricing table. Runs entirely in your browser (0 uploads, works offline).

How to use it

1. Load a preset — Balanced, Cheap & fast, or Best quality — or drag the three weight sliders yourself.
2. Read the scoreboard: each row shows a model's blended score, its accuracy, a latency bar in milliseconds, and the dollar cost per call, with the leader on top.
3. Watch the fastest / cheapest / most accurate badges stay put while the winner moves as you re-weight — that gap is the whole point.
4. Open the X-ray panel to see your pick versus the optimal pick and the exact accuracy you traded for cost and speed. Copy or share the result.

What this clears up (the fundamentals)

• There is no single "best model" — only the best model for a weighting. Accuracy, cost, and latency pull in different directions, so the winner is a function of what you priced in, not a fixed leaderboard.
• The most accurate model is rarely the best value — a car that is a couple of points more accurate can cost ten times as much and run several times slower. Once cost and latency carry any weight, the top-accuracy model often finishes mid-pack.
• Routing is a weighting decision — sending every request to the biggest model is a choice to weight accuracy at 100%. Real systems route: cheap-and-fast for easy calls, the expensive car only when the task needs it.
• Cost and latency normalize across the field — a model's cost only matters relative to the others racing, which is exactly how you reason about a roster when you pick who runs what.

Where it's used

A top-of-funnel AI-engineering check that answers the market's number-one API-cost fear: which model should I actually call? It's a miatz build-lab concept playable — race it here, then learn to build the scoring blend, the normalization, and the routing decision yourself in AIE-402.

FAQ

Why isn't the most accurate model just the winner?

Because accuracy is one axis of three. The engine blends accuracy × wAccuracy + (1 − normCost) × wCost + (1 − normLatency) × wSpeed, normalized across the field. Give cost and speed any weight and a slightly-less-accurate model that is far cheaper and faster wins on total value.

What do the five models represent?

Illustrative, precomputed model tiers spanning the classic cost/latency/quality frontier — from a tiny cheap-and-fast car to a large, slow, expensive, most-accurate one. The numbers are seeded into the page so the race is deterministic and repeatable, not a live benchmark.

How is the winner chosen?

Every model gets one blended 0–1 score under your current weights; the highest score takes the flag, ties broken by lower latency. Cost and latency are min-max normalized across the five racers before weighting, so a model is judged against the field it is racing.

What does "your pick vs. the optimal pick" mean?

The winner is the optimal pick for your weighting — the panel makes that explicit by showing where the most accurate model actually finishes once its price and latency are counted, and exactly how many accuracy points you traded for the value.

Is anything uploaded?

No. The weights and the whole race are computed entirely in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still runs.

Limits

A teaching race over precomputed, illustrative model data with hand-set accuracy, latency, and cost — it's a fast, opinionated read on the tradeoff, not a live benchmark or a procurement decision. Real numbers move with your prompt, your provider, caching, batching, and load. The lesson it teaches — that the best model is a weighting choice, and routing beats always calling the biggest one — is exactly the real practice.

Related

Part of the miatz playgrounds. Explore the rest from the playground catalog.

Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.