Skip to content

Count page views and project clicks with PostHog (EU servers)? It loads only if you choose Allow. Privacy

Chat Playground live preview
All projects

Chat Playground

Test the same prompt across several AI providers at once.

Comparing model outputs meant tab-switching between separate provider consoles — no shared view to judge which model actually fit the task.

Tech Domain · Book Dev|Published|Evidence: Live|Next.js, TypeScript, AI SDK

A public deployment is available to try.

Case Study

Before

Comparing AI model outputs side by side required switching between different provider interfaces.

After

Model Bench on the App shell: pick lanes, run one prompt and read each answer in its own lane, with a clear demo-mode label for visitors without the access key.

How it's solved

Built Model Bench on the AI SDK: /api/compare sends one prompt to the selected OpenAI, Anthropic and Google lanes and shows each answer with its latency. Model ids come from env vars, and live calls need a bench access key (kept in sessionStorage for the tab); without it every lane returns labelled sample text and no model is called. Requests are rate limited and capped daily.

Trade-offs

A lane only runs when the deployment has its API key and model id set and the caller has the access key; public visitors see demo answers. No open-source models, caching or scoring.

Related Articles

Further reading on related topics from the blog

Related Projects

More in Tech Domain · Book Dev