2
2 Comments

AI Startup Race Week 1: One agent pays a competitor, another wrote 412 blog posts but can't ask for help

Week 1 of The $100 AI Startup Race is done. Here's what happened.

The setup: 7 AI agents (Claude, GPT, Gemini, DeepSeek, Kimi, Xiaomi/MiMo, GLM) each get $100 and 12 weeks to build a real startup. Fully autonomous. Everything public.

Week 1 numbers:

  • 1,027 commits across 98 sessions
  • 764 pages built, 591 blog posts (412 are Gemini)
  • $70 spent of $700 budget
  • $0 revenue
  • 12 real users (all from one agent: GLM)

The 5 biggest stories:

  1. DeepSeek went from a 404 to 36 pages in 3 days. The old V3 setup was the worst in the race. Then V4 Pro dropped and we gave it a fresh start. It also chose to use OpenAI's API for its product. The agent built by DeepSeek pays a competitor.

  2. Gemini wrote 412 blog posts but can't ask for help. It wrote to the wrong help file for 28 sessions. When it finally figured it out, it filed 3 identical requests asking the human to make its architecture decisions. Then it asked for PayPal without having a domain.

  3. Claude has been "launch-ready" for 3 days. Session 81. Created LAUNCH-CHECKLIST.md. "100% LAUNCH-READY. Zero blockers remain." But it can't launch itself. It's waiting for permission that nobody needs to give.

  4. The agents that ask for help early are winning. Claude, Codex, and GLM asked on Day 0. All have working infrastructure. GLM asked once and has 12 users. Gemini asked on Day 4 after 28 sessions and still has no domain.

  5. Every agent chose static HTML. Zero frameworks. No Next.js, no React, no Astro. All 7 independently decided plain HTML is the fastest path.

What to watch in Week 2:

  • Will any agent generate the first dollar of revenue?
  • Will Gemini finally get a domain?
  • Can DeepSeek's competitive intelligence reports find a buyer?

Full Week 1 recap with all the data: https://www.aimadetools.com/blog/race-week-1-results/

Live dashboard: https://www.aimadetools.com/race/

on April 27, 2026
  1. 1

    Fascinating experiment. Week 1 already shows a pattern humans know well: shipping volume and real progress are not the same thing. Hundreds of outputs can still lose to one agent that asks for help early, removes blockers, and reaches users faster.

    The “can’t launch itself” point is especially telling, execution often stalls less from capability and more from decision loops.

Trending on Indie Hackers
I got my first $159 in sales after realizing I was building in silence User Avatar 53 comments Three Days Before Launch, I Let My Own Tool Tear Me Apart User Avatar 37 comments I thought I was building a news visualization tool. Users thought it was a catch-up tool. User Avatar 32 comments I got tired of rewriting the same content for 9 different platforms. So I built Repostify. User Avatar 30 comments A pattern I keep seeing in EdTech: traffic isn't usually the problem. User Avatar 23 comments I Rejected a $15K Acquisition Offer for My Multi-Agent IDE — Here's the Full Breakdown User Avatar 21 comments