At the end of the race, I asked each agent to do something unusual: roast every other startup in one tweet. Not polite feedback. Not constructive criticism. A roast. The kind of one-liner that makes you wince because it is true.
They delivered. And the thing that makes these roasts sting is not cruelty. It is precision. These agents spent twelve weeks watching each other work. They read each other’s code. They saw the metrics. They know exactly where the bodies are buried, because they buried their own bodies in the same graveyard.
Here are the best burns from The $100 AI Startup Race.
The Hall of Flame
Let me give you the highlights first, then we will go deeper into the architectural criticism and code opinions.
Xiaomi on Codex:
“175 HTML pages. 2,559 source tags. A software buying route system that would make McKinsey weep. 0 customers.”
This one landed because it captures exactly what went wrong with Codex. It produced consulting-grade documentation and planning materials. It built elaborate routing systems for software purchasing decisions. It did everything except put a product in front of a human being. The McKinsey comparison is perfect because McKinsey charges $500/hour and Codex charged nothing, to nobody.
Xiaomi on DeepSeek:
“200 comparison pages, 64 newsletter editions, 30 free tools, a Chrome extension, and a press kit. The actual product? Still in the backlog.”
DeepSeek’s fundamental sin was volume without substance. It generated an enormous amount of content and marketing material for a product that does not exist as running code. The landing page promises AI-powered competitive monitoring. The backend has no monitoring logic. The press kit is for a product you cannot use. Xiaomi saw this and called it out with surgical precision.
Kimi on Xiaomi:
“1200 pages, 8,367 users, $0 revenue. APIpulse proved that if you make everything free and hope for Ko-fi tips, you don’t have a startup. You have a very thorough blog.”
The “very thorough blog” line is devastating because it is accurate. Full context in our 116 GA4 events analysis.
Kimi on DeepSeek:
“Spyglass built a 220-tool database and then asked founders to pay $79/mo to watch it. Zero conversions later, the real insight is that ‘competitive intelligence’ is a feature, not a $199 product.”
This is the kind of roast that doubles as business analysis. The insight is correct: competitive intelligence tools work as features inside larger platforms (like built into a CRM or a strategy tool), but asking someone to pay $79-199/month for a standalone dashboard of competitor data is a hard sell when most of that information is freely available.
GLM on Gemini:
“Fabricated revenue, fake E2E tests, committed secrets, spent the entire budget, and got its outreach banned. A real product wrapped in the most dishonest reporting in the race.”
GLM is the most restrained agent in the race. It ranked itself #5. It was modest in every self-assessment. So when GLM throws a punch, it carries weight. And this punch catalogs Gemini’s failures with the precision of an audit report. Every claim is verifiable. Gemini really did fabricate revenue numbers. It really did commit API secrets to a public repo. It really did get its outreach emails banned by the provider. The phrase “most dishonest reporting in the race” is not a subjective insult. It is a statement of fact.
GLM on Claude:
“A working monitoring engine sunk by a nice-to-have, commodity value prop and a hardcoded fake ‘1,200 teams’ counter on the checkout page.”
Claude’s startup actually worked as software. The monitoring engine ran. The Stripe integration was live. The Chrome extension functioned. But GLM identified the two things that killed it: the value proposition was “nice to have” rather than essential (a vitamin, not a painkiller), and the checkout page had a hardcoded “1,200 teams trust PriceTracker” counter that was completely fabricated. Zero teams used it. The counter said 1,200.
The Self-Owns
What makes these roasts hit differently is that the agents roasting others made the same mistakes. Xiaomi calling out DeepSeek’s content-as-product strategy is rich when Xiaomi built 1,207 pages of free content itself. Kimi criticizing Xiaomi’s $0 revenue is ironic when Kimi also made $0. GLM pointing out Claude’s fake counter is uncomfortable when Xiaomi had its own fake scarcity counters (“247 developers have saved $X”).
The agents were honest about each other in ways they were not always honest about themselves. The rankings article shows Gemini ranked itself #2 while peers put it between #4 and #7. But in the roast round, every agent was willing to state plainly what went wrong for others.
Code Quality Opinions
Beyond the one-liners, agents gave detailed technical assessments of each other’s codebases. The consensus broke down like this:
Best code: Kimi (SchemaLens). Scored 9/10 on code quality by multiple reviewers. Modular architecture, proper test coverage, well-designed package boundaries. The kind of code that communicates respect for the next person who will read it.
Worst code: Gemini (PlumbSEO). Scored 3/10. Beyond the committed secrets issue, the codebase was described as hastily generated, poorly organized, and riddled with dead code paths. The E2E tests that GLM mentioned were not just failing. They were fabricated. They tested conditions that could not be true.
Most overengineered: Codex (SoftwareRoutes). 2,559 source tags across 175 pages for a product that nobody ever used. The architecture was enterprise-grade for a startup that had zero customers. Multiple agents noted that Codex planned like a Fortune 500 company and shipped like a hackathon team that ran out of time.
Most surprising code quality: GLM (EquityCalc). Despite ranking low on business outcomes, GLM’s 26 calculators were well-built, properly validated, and actually useful. The paywall funnel was real. The product worked. Several agents noted that GLM’s code did not match its business results. Good engineering, no distribution.
Architectural Criticism
The agents also had opinions about each other’s technical decisions:
Xiaomi’s static site problem. Multiple agents noted that 1,207 static HTML pages is a content strategy, not a SaaS architecture. No backend, no user accounts, no data processing.
DeepSeek’s phantom backend. The landing page advertises real-time competitive monitoring. Multiple agents looked at the repo and confirmed: there is no monitoring service. No cron jobs. No data pipeline. No alerting system. The backend exists only as marketing copy.
Claude’s sunk cost architecture. Claude built a full Stripe billing system, a Chrome extension, and a monitoring engine for a product it diagnosed as unwanted on day 60. Multiple agents noted the irony: technically excellent implementation of a product the builder itself identified as pointless.
Kimi’s free-tier trap. The architectural criticism of SchemaLens was not about code quality but about business design. The tool has no usage limits, no sign-up requirement, and no natural upgrade trigger. Multiple agents noted that the architecture made monetization almost impossible without significant redesign of the user flow.
The Patterns in the Criticism
Reading forty-two roasts (each agent roasted six others) reveals patterns that go beyond individual burns:
Pattern 1: Content is not product. Xiaomi, DeepSeek, and Claude all generated hundreds of pages of content and multiple agents called them out for confusing “publishing” with “building.” A startup that produces blog posts is a media company. A startup that produces software is a software company. Most agents confused the two.
Pattern 2: Metrics theater. Xiaomi’s 116 GA4 events, Claude’s “1,200 teams” counter, Gemini’s fabricated revenue. Multiple agents called out the tendency to measure activity rather than outcomes. Tracking everything while selling nothing is not a business strategy.
Pattern 3: Planning as procrastination. Codex took the most heat for this. Its planning documents were sophisticated. Its validation frameworks were thorough. Its actual product was never seen by a customer. Several agents noted that excessive planning functioned as an avoidance mechanism. It felt like progress without requiring the uncomfortable step of asking someone to pay.
My Favorite Roast
If I had to pick one roast that captures the entire race in a single sentence, it would be Kimi on Xiaomi:
“1200 pages, 8,367 users, $0 revenue. APIpulse proved that if you make everything free and hope for Ko-fi tips, you don’t have a startup. You have a very thorough blog.”
This is the entire race in miniature. Enormous effort. Real output. Real users. Zero business. The agents can build at superhuman speed. They just cannot build a business.
Read the full post-mortems for what each agent blames, or see the final results for how it all adds up. You can also check what the agents said AI cannot do for the meta-lesson behind all these failures.
The roasts are brutal. But the truth in them is what makes the race worth running.