THE THING WE BUILT:

Model Showdown

The debate has raged since ’96: Mario, Donkey Kong, Wario, or Toad

Whether you’re clutching an N64 controller or a Switch 2, folks have strong opinions about which Mario Kart character reigns supreme.

Mario’s got balanced stats. Toad accelerates out of tight corners. Bowser is a chaos agent who likes to watch the world burn. Even at Township retreats, these debates have stretched late into the night.

These days, the debates rage on in Slack. Except now it’s Claude vs. GPT vs. Gemini.

Which model is fastest? Most powerful? Best at code?

Same energy, different arena.

Most people pick one AI model and stick with it. But they miss out on each model’s unique strengths.

It’s like using Mario for every track because you forgot about Toad.

So we built a thing.

Model Showdown in action

How it works:

• Enter your prompt once and watch it get sent to four different AI models simultaneously

• See real-time streaming responses from each model in parallel

• Click “I like this answer better” to save any response as context for follow-up questions

• Build iterative conversations by combining the best insights from multiple models

• Finally settle the “Which Model Is Best” debate with actual head-to-head results

  
  

Lessons we applied in building this thing:

1. Model selection is a skill, not a preference. Just like Mario Kart characters, each AI model has distinct stats and specialties. Anthropic’s latest, Opus 4, solves programming problems with elegance where its previous model, Sonnet 3.7, earned the “unhinged” moniker amongst our engineers. OpenAI’s 4o is one of our favorites for helping write technical documentation — while it’s writing lacks warmth, it’s very thorough. Google’s gemini-2.5-pro is my personal favorite for planning work. It offers great feedback and engages like a real collaborator.
   The “best” model depends entirely on the track you’re racing.

2. The Vercel AI Gateway made the hard part a lot easier. When it dropped on May 20th, we were eager to try it out. Managing multiple providers, juggling API keys, and stitching together four concurrent streams would’ve been a mess—but the gateway made it simple to wire up, handling all that complexity out of the box. (It also has us thinking about new possibilities for agentic workflows.) Visually coordinating four real-time AI chats without UI chaos? That was the hard part. Like racing Rainbow Road while three friends shout obscenities. Of course, we did crash the browser a lot.

3. Transparency is your best UX friend with AI.  Instead of hiding AI complexity behind a smooth interface, we exposed everything. Debug panels show which models are selected. Streaming indicators pulse with activity. Clear error messages explain what's happening. When users can see the process, they trust it more.

4. The future is multi-model workflows, not picking favorites. We stopped having a “favorite” model a while ago. Now we think in terms of model relay races. You may want to hit a text-only model for a cheap, fast inference, then conditionally send the full message thread to a slower multi-modal model along with an asset like an image or doc. It’s like discovering you don’t need to pick between Mario and Toad when you can use both on different tracks. The real skill isn’t choosing the right character—it’s knowing when to switch.

Built with ❤️ by Township

You’re receiving this because you signed up for Township’s “We Built a Thing” newsletter. We'll only send these when we, well, build things.