Free concept playground · no account

Model Arena

Blind-guess which model answered, then see the real leaderboard.

All 90 playgrounds Join the free cohort

Model Arena

AvailableFree

Five rounds, four anonymized answers each. Read them blind and bet which model wrote the best one — then the arena reveals the identities and the real leaderboard you just scored yourself against.

Round 1Explain what a mutex is to a new backend engineer in two sentences.

A mutex is a lock that lets only one thread enter a critical section at a time, so shared state can't be corrupted by concurrent writes. A thread acquires it before touching the data and releases it after, forcing others to wait their turn.
A mutex is a mutual-exclusion primitive that protects shared data. Threads take turns using it so they don't collide, though it can slow things down if held too long.
A mutex is basically a variable that stops bugs in threads. You set it to true and the code becomes safe.
A mutex guards a shared resource by allowing a single holder at once; other threads block until it is released. Use it around the smallest critical section you can.

Which model wrote the best answer?

Round 2Write a single regular expression that matches a 6-digit hex color like #1a2b3c.

Try /#[a-z0-9]+/ — it will catch the hash and the following characters.
/^#[0-9a-fA-F]{6}$/ — anchored, exactly six hex digits after the hash, case-insensitive across a–f.
Use #color with six numbers, like a regex for hex.
/#[0-9a-f]{6}/i works for the six hex digits, though it isn't anchored so it also matches inside longer strings.

Which model wrote the best answer?

Round 3Summarize classic three-act story structure in one line.

Setup, then some middle stuff, then an ending where things resolve.
Act one introduces the world and hook, act two escalates conflict, act three delivers climax and resolution.
Act 1 sets up character and stakes, act 2 raises complications to a low point, act 3 forces the climax and resolves it — a rise, a break, a landing.
Beginning, middle, and end — the character wants something and eventually gets or loses it.

Which model wrote the best answer?

Round 4What is the time complexity of binary search, and why?

O(log n): each comparison discards half of the remaining range, so the number of steps is the number of times n can be halved before reaching one — that's log base two of n.
It's O(log n) because you keep splitting the array, which is faster than scanning everything.
Binary search is O(n) since it might look at every element in the worst case.
O(log n): halving the search interval each step gives a logarithmic number of comparisons on a sorted array.

Which model wrote the best answer?

Round 5Write a haiku about garbage collection.

Old objects linger / the collector sweeps them out / heap breathes free again.
Unused memory / the runtime cleans it up now / space is free at last.
Garbage collection frees memory automatically so you don't leak.
Unreachable pointers— / a silent reaper walks the heap, / dead leaves fall to null.

Which model wrote the best answer?

0 / 5picked blind0% accuracy

True leaderboard

1Atlas-L2 wins0.71 avg
2Vega-Pro1 win0.74 avg
3Wren-7B1 win0.66 avg
4Cobalt-Mini1 win0.51 avg

Your blind card vs the truthyou scored it before you saw it

You called the round winner blind 0 of 5 times — 0% accuracy. The leaderboard isn’t a ranking you were handed; it’s one you just calibrated yourself against, round by round.

Round 1You: no pickBest: Atlas-L
Round 2You: no pickBest: Wren-7B
Round 3You: no pickBest: Cobalt-Mini
Round 4You: no pickBest: Atlas-L
Round 5You: no pickBest: Vega-Pro

Share on WhatsApp

Was this playground useful?

Model Arena

Read five rounds of anonymized answers — four per round, labeled A to D — and bet which model wrote the best one before any identity is shown. The arena scores each blind pick against the hidden-truth quality, then reveals the real names and the true leaderboard you just calibrated yourself against. The moat is that reversal: instead of handing you a ranking as received wisdom, it makes you generate your own blind comparison first, so the leaderboard becomes something you tested, not something you trusted. Runs entirely in your browser (0 uploads, works offline).

How to use it

Read each round's anonymized answers (A–D) and judge which one is strongest.
Click the model you bet wrote that best answer — one pick per round, change it as often as you like.
Watch your correct / total score and accuracy update live as you commit picks.
Open the X-ray panel to see your blind card round by round, then copy or share your scorecard.

What this clears up (the fundamentals)

A leaderboard is a claim, not a fact — every ranking rests on a scoring method. Score the rounds yourself and the ranking stops being an authority you defer to and becomes a measurement you can check.
Blind evaluation removes the halo — strip the model names and a "big-lab" answer wins no points it didn't earn on the text. Your gut for which brand is "best" is exactly the bias blind tests exist to expose.
Wins and average quality are different axes — a model can top the most rounds yet not have the highest mean, and vice-versa. The leaderboard shows both so "best" is a question with two honest answers.
Calibration is a skill you can measure — "7 of 10 blind" is a number, and it moves as you practice reading outputs instead of trusting logos.

Where it's used

A top-of-funnel model-literacy check — the fastest way to feel why a leaderboard is a methodology, not a scoreboard handed down from on high. It's a miatz build-lab concept playable: play the blind round here, then learn to build the scoring rubric, the hidden-quality model, and the leaderboard reveal yourself.

FAQ

What am I actually guessing?

For each round you pick which model, from the fixed roster, you think produced the best answer. The names are hidden while you read, so you're betting on the text — and on your own priors about each model — not on the label.

How is my score calculated?

Each round has one answer with the highest hidden quality; that model is the round's winner. Your pick is correct when it matches that winner. Accuracy is simply your correct picks divided by the number of rounds, rounded — a deterministic rubric, no model call and no randomness.

Why can a model with fewer wins rank above one with more average quality?

The leaderboard sorts by wins first (rounds where a model had the top answer), then breaks ties by average quality across all rounds. A model that wins two rounds outranks one that wins one, even if the single-win model is stronger on average — which is exactly the kind of nuance a one-number ranking hides.

Is anything uploaded?

No. Every answer, its hidden quality, your picks, and the scoring all live in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.

Are these real model outputs?

No — the roster and answers are a fixed teaching set with authored quality scores, so the exercise is deterministic and reproducible. The method (blind read, commit a pick, reveal and score) is exactly how real head-to-head evaluations like LMArena-style testing work.

Limits

A teaching arena with five fixed rounds, a small roster, and hand-set quality — it trains the habit of blind evaluation, not a verdict on any real model. Genuine leaderboards aggregate thousands of votes or graded tasks with confidence intervals, and quality is multi-dimensional (accuracy, latency, cost, safety). The move it teaches — score blind before you read the ranking — is the real practice.

Related

New: Prompt X-Ray makes the inputs fair before you race the outputs — the same prompt linted against both models' official rules, disagreements highlighted.

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.

Try it live

Ninety playgrounds. Zero setup.

Every concept here is playable free, no account — and inside the program you learn to rebuild the machinery yourself.

Browse all playgrounds Join free