Public AI / REPORT 001
Testing TH AI Passport: model names do not tell the whole story
A hands-on test across reasoning, web development, image generation, and video shows why the system around a model matters.
This review is based on early access to a test version of TH AI Passport, provided with help from Thanarat Kuawattanaphan. It should be revisited when the full service launches.
The test covered reasoning, programming, image generation, and video rather than simple chat prompts.
A demanding React application brief
Claude Opus 4.8 was asked to build UniWallet OS, a React and TypeScript finance application for students with budgeting, savings goals, reports, charts, backup, mobile support, and security requirements.
Reasoning and instruction following
The model identified conflicting constraints, explained its reasoning, and followed the requested structure reasonably well, although long tasks still required human checking for omissions.
Web development and the harness
The result did not match using Claude through its primary environment. It produced substantial planning and code but did not deliver a complete large application with every function tested.
That gap may come from the harness around the model: the system responsible for tools, files, code execution, and presentation. The same model name can therefore deliver very different practical capability across platforms.
Images, video, and the practical limit
Image generation was comparatively close to the direct model experience, although Thai text in images remained unreliable. Video generation could not be evaluated because the request failed before output was produced.
TH AI Passport is valuable as an AI learning and experimentation space, but complex applications and production-ready video still require human review and further development.
SOURCES / 01