Every output, unedited
135 single-file HTML outputs you can run in the browser, filtered by company, model, reasoning effort and prompt. Broken outputs stay in, clearly marked.
AI model comparison · Prompts.sk
AI
AI
Lab
A public comparison of AI models: 27 models from Anthropic, OpenAI, Google, xAI, Mistral, DeepSeek and Z.ai get the same zero-shot prompts, and all 135 outputs are published unedited — with time, cost, automated quality scores and visitor votes.
/01 The problem
Every new AI model launches with impressive demos and benchmark tables. Neither answers the question companies actually ask: what will this model build on the first try, what will it cost, and how often will it simply not work?
Benchmarks measure narrow skills and demos are hand-picked. I wanted the opposite: give every model exactly the same demanding task, a single attempt and no human edits — and then publish everything, including the outputs that fail.
AI Lab does that with real, runnable software: a 3D solar system, a GTA-style game, a monitor test, a map of Manhattan and chess with its own engine. Anyone can open the results side by side and judge for themselves.
/02 What it does
GALLERY
COMPARE
ARENA
SCORE
COST
MODELS
Side by side
One prompt across all models in a sortable table of scores, time and cost — or two outputs running live next to each other.
Blind arena and votes
Visitors pick the better of two outputs with the model names hidden; a Bradley–Terry rating turns the votes into an Elo-style score.
Measured, not guessed
An 8-point requirement checklist per prompt, automated UX and technical quality and an objective perft test for the chess engine add up to one overall score.
Time, tokens and price
Every run records how long it took, how many tokens it used and what it cost at the dated API price list.
Leaderboard and company pages
A model leaderboard, a value-for-money view and a page for each of the seven AI companies with the models tested.
/03 How it works
/01
One prompt, one attempt
All five prompts are in English and share the same rules: one self-contained HTML file, no external libraries or images. Each model gets the prompt in a fresh conversation and its first answer is saved byte for byte.
/02
Automated testing
Headless Chrome runs every output, records errors and uses it like a visitor would; Lighthouse and a UX test at three screen sizes produce the technical and UX scores.
/03
Review and approval
AI reviewers score the 8-point checklist with evidence from the code, and the scores go live only after I approve them by hand.
/04
Publish and keep adding
The pages are prerendered in Slovak and English for search engines and AI assistants, and new models are added as they come out.
/04 Under the hood
FRONTEND
BACKEND
TESTING
SEO & AIO
Plain HTML, CSS and JS
No framework and no build step: a single data catalogue drives every page, and the site even runs straight from disk.
PHP and MariaDB
A small API stores visitor votes and arena matches — one vote per visitor and output, rate limits and hashed IP addresses.
Node.js and headless Chrome
Scripts for thumbnails, Lighthouse, UX measurement, scripted interaction and the perft chess test, plus API runners for the models.
Readable by AI
Prerendered pages with JSON-LD, FAQ and clean addresses in both languages, so crawlers and AI assistants see the full tables, not an empty shell.
27AI models
5Identical prompts
135Unedited outputs
81 %Worked first time
/05 FAQ
What is a zero-shot prompt?
A task the model receives without example solutions and without any follow-up. It has to understand and solve the task on the first attempt — which is how most people actually use AI.
Why are the outputs not edited?
Because the point is to show the model's real ability. A broken output stays in the gallery marked as broken; where a fix exists, the same model made it as a separate copy and the original file is unchanged.
How is quality measured?
Every output runs automatically in Chrome: Lighthouse, errors during use, smoothness and the mobile layout give the technical score, and a UX test checks button sizes, layouts at three screen sizes, keyboard focus, click feedback and text size. Each prompt also has an 8-point requirement checklist, and the chess engine is verified with an objective perft test.
Which models are included?
As of October 2026, 27 models from Anthropic, OpenAI, Google, xAI, Mistral, DeepSeek and Z.ai — among them Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.6 and DeepSeek V4 Pro. New models are added as they come out.
Who is behind AI Lab?
I designed and built it, and it runs under Prompts.sk, my AI consulting company. It is free for anyone choosing an AI model for real work.
/06 Other projects
AI Visibility Check →
Free SEO & AI visibility audit — 40+ checks, result in 30 seconds.
Global Collector →
Android app for coin and banknote collectors — 350,000+ items.
YesNoBall →
An AI crystal ball for yes-or-no questions, with fact-checking.
← Back to peterpetrus.sk
All projects and expertise in one place.