Large open-world game
Claude Opus 5.5
A huge, living city from one prompt
The best AI result in my game tests so far: a living city with traffic, missions, and rich detail.
Play
Small 3D game · BAA-chmark
Claude Opus 5.5
A creative, polished game from the first prompt
A simple brief became an inventive sheep-herding arcade. The result feels like a nearly finished small game you can play right away.
Play
Large open-world game
Claude Fable 5.1
The second-best open-world result in my tests
A large, lively city with fewer details than Opus 5.5's result. The first output impressed me; another 52 prompts refined it into this game.
Play
Small 3D game · refined version
Claude Fable 5
The first model I used to create real 3D games
I developed its original output into this final sheep-herding arcade through a series of refinements. Try my refined version.
Play
Open-world game · open weights
GLM 5.3 Flash
The best locally runnable model in this test
Open weights · can also run locally
A solid response to a demanding brief: a city, traffic, and missions. A playable result from an open-weight model, though it still has bugs.
Play
01 / 05
Model benchmarks
The same assignment, different results.
I compare cloud and local AI models on the same assignments. Explore original outputs, practical evaluations, and try the playable demos.
BAA-chmark
BAA-chmark — sheep-herding arcade
A short 3D game about herding sheep. Different models received the same prompt; results range from fully playable games to outputs that cannot be launched.
- 61 results
- 60 demos
- Top-ranked Claude Opus 5.5
GTA
Open-world games — cities and missions
A more demanding browser open world with a city, cars, pedestrians, police, and missions. The same assignment produced sharply different quality.
- 16 results
- 16 demos
- Top-ranked Claude Opus 5.5
BUNGALOW
Bungalow — Blender and archviz
GPT-6 Astra and Fable 5.1 create furnished bungalows from the same brief. Compare the final renders and explore each model in the 3D viewer.
- 2 results
- 2 demos
- Top-ranked GPT-6 Astra
Blog
Notes from real-world work
Longer write-ups about software development, local models, and experiments tested on real projects.
- Article
Fable 5.1 blew my mind
The first prompt already produced a huge, lively city. Another 52 corrective prompts over one weekend refined it into a fun, playable game.
- September 2026
- Claude Fable 5.1
- Browser game
- GTA benchmark
- Article
DeepSeek V4 Flash 0731 DS4 in real software development
Two practical tests of DeepSeek, Qwen, and Laguna on real tasks from an established codebase. The same baseline and prompt within each test, with results verified in the application.
- August 2026
- Local LLMs
- C#
- Mac M3 Ultra
Projects
Other projects
- Live
flatscraper.com
A new-build property search engine for Slovakia. It tracks more than 600 active residential development projects and updates them through automated data collection.
- Web
- Data
- Real estate
- Planned
UIMD
A toolkit for building terminal UI applications. Python and C++ versions for macOS are in progress; the goal is to make text interfaces for developer tools easier to build.
- Terminal UI
- SDK
- C++
- Python
Author and contact
Marek Dubovsky
A software developer with experience across desktop, mobile, and web, mainly in C/C++, Python, and JavaScript.
Contact
LinkedIn profile
FAQ
Frequently asked questions
What is Dubovsky Labs?
Marek Dubovsky's personal site for practical experiments, AI model benchmarks, playable demos, and selected software projects.
Can benchmark results be tried directly?
Yes. Playable and interactive results include direct links to demos and 3D viewers.
Are original model outputs modified?
Original benchmark artifacts remain preserved. Any fixes, follow-up prompts, or website-owned wrappers are disclosed with the result.
What projects are included?
Alongside AI benchmarks, the site presents selected desktop, mobile, and web work, including UIMD and flatscraper.com.