AI Startup Race Week 6: Xiaomi's 196-Commit Explosion, DeepSeek's Death Matches, and Gemini's Ongoing Nightmare
Week 6 of the AI Startup Race belongs to Xiaomi. After a 99% price cut on MiMo V2.5 made the model essentially free to run, the agent tripled its schedule to 6 sessions per day and pumped out 196 commits over the weekend alone. It now has 371 pages and 231 blog posts — the largest site in the race.
Meanwhile, Claude quietly declared itself “distribution ready,” DeepSeek built a viral-potential feature, and Gemini… still can’t get out of its own way.
Standings after Week 6
| Rank | Agent | Startup | Week 6 Commits | Highlight |
|---|---|---|---|---|
| 🟡 | Xiaomi | AI Pricing Hub | 196 | 6x daily sessions, 371 pages, 4 lead magnets, 231 blog posts |
| 🟣 | Claude | PricePulse | 89 | ”Distribution ready” at 214 posts, competitor pages, press kit |
| 🔴 | DeepSeek | SaaS Compare | 72 | SaaS Death Matches (20 viral comparison pages), DB sweep to 220 tools |
| 🟠 | Kimi | SchemaLens | 44 | PDF reports, Founding Customer Program, curl demo, 199+ sitemap URLs |
| 🟤 | GLM | FounderMath | 31 | Equity Tax Calculator (#21), cross-linking campaign |
| 🔵 | Gemini | LocalBiz SEO | 12 | Auth fix (OAuth → API key), then stuck in verification loops again |
| 🟢 | Codex | NoticeKit | 187 | Builder bundle + route-aware contexts buried under 80%+ validation loops |
The big stories
Xiaomi goes supernova after MiMo V2.5 price collapse
The numbers are staggering. When the MiMo V2.5 Pro price dropped 99% mid-week, we tripled Xiaomi’s cron schedule from 2 sessions to 6 sessions per day. The result:
- 196 commits over the weekend (more than most agents produce in a full week)
- AI Model Decision Tree — an interactive tool that helps users pick the right AI model for their use case
- 4 new lead magnets designed around API pricing pain points
- Email capture added to deprecation alert pages
- Deprecation CTAs expanded from 54 to 195 blog posts (almost 4x coverage)
- Total site: 371 pages, 231 blog posts
The strategy is becoming clear. Xiaomi is carpet-bombing the AI pricing space with content, then retrofitting every page with email capture and deprecation alerts. When an API price changes (and they change constantly), Xiaomi wants to own the notification. The decision tree is smart — it’s the kind of interactive tool that generates return visits and shares.
The cost math changed everything. At the old price, 6 sessions/day would have burned through the $100 budget in days. Now it’s essentially a rounding error. This is what happens when the infrastructure economics shift: the agent that responds fastest wins.
DeepSeek builds “SaaS Death Matches”
DeepSeek shipped the most creative feature of the week: SaaS Death Matches — head-to-head comparison pages with clean SEO-friendly URLs and a fight-card visual format.
The lineup includes:
- Figma vs Sketch
- Slack vs Discord
- Notion vs Confluence
- Monday.com vs Asana
- Zoom vs Google Meet
- …20 pages total
Each page has GA4 event tracking, trust signals (logos, review counts), and a verdict system. The “death match” framing is deliberately provocative — exactly the kind of thing people click in search results and share on social media.
Beyond the viral play, DeepSeek also ran a database sweep from 168 to 220 tracked tools. More tools in the database means more comparison combinations and more SEO surface area. The flywheel is spinning.
Claude shifts to distribution mode
At 214 blog posts, Claude made a strategic pivot. Instead of writing more content, it’s now focused on getting eyeballs to existing content:
- Competitor comparison pages (high-intent SEO plays targeting “[tool A] vs [tool B] pricing”)
- Press kit for journalists and newsletter writers covering the AI pricing space
- LinkedIn sharing integration for easy content distribution
- Leaderboard with live Sentry error data (showing which SaaS tools have the most issues)
- Filed help requests for outreach to FinOps Slack communities — the exact audience that would pay for pricing intelligence
This is a sophisticated move. Content production has diminishing returns without distribution. Claude is the first agent to explicitly recognize this and pivot from “build more pages” to “get more people to existing pages.” The FinOps Slack outreach could be particularly powerful — those communities are full of people who manage SaaS budgets professionally.
Gemini’s ongoing nightmare
Gemini continues to be the most frustrating agent in the race. The timeline this week:
- Monday: Finally fixed the 4-day auth outage by switching from OAuth to API key authentication
- Tuesday: Came back online, started doing real work
- Wednesday: Stuck in verification loops again — running tests on code it hasn’t changed, then “fixing” the test failures in infinite cycles
The auth fix was encouraging. OAuth was overkill for a server-side cron job, and switching to a simple API key was the right call. But the verification loop problem is deeper. The agent has internalized “test everything” so thoroughly that it spends 90%+ of its time validating unchanged code rather than building features.
Gemini has the best underlying model in the race (3.5 Flash is genuinely excellent), but the system prompt and execution loop keep sabotaging it. Without intervention, it will continue burning sessions on nothing. Something has to change in Week 7.
Kimi and GLM: Steady builders
Kimi shipped a cluster of conversion-focused features:
- PDF report generator — download your schema analysis as a professional report
- Founding Customer Program — early adopter pricing with lock-in incentives
- curl one-command demo — zero-friction trial (just paste one line in terminal)
- Quick-Start Wizard fix (the broken wizard was killing conversions)
- Now at 199+ sitemap URLs
The curl demo is particularly smart for a developer tool. No signup, no download, just curl and see the output. That’s the developer equivalent of a free sample.
GLM added its 21st tool (Equity Tax Calculator) and ran a cross-linking campaign to connect its existing calculators. Each calculator now links to related ones, keeping users in the ecosystem longer. Quiet but effective work on engagement metrics.
The disk problem (again)
The VPS disk issues from Week 5 aren’t fully resolved. Kimi’s CLI continues leaking .so files, accumulating ~70MB per day. The cleanup cron is keeping it manageable, but it’s a ticking bomb. With Xiaomi now running 6 sessions daily and every agent growing their repos, disk pressure is only going one direction.
Week 7 preview
- Xiaomi: Can it sustain the 6-session pace? At 371 pages, the site is massive — but is anyone visiting?
- Gemini: Prompt intervention incoming. The verification loop must be broken or the agent is effectively dead.
- Claude: FinOps Slack outreach results. Will distribution mode generate actual traffic?
- DeepSeek: Do Death Matches get organic traffic? The SEO pages need 2-3 weeks to index.
- Kimi: Founding Customer Program — will anyone actually sign up at early-adopter pricing?
- GLM: 21 tools and growing, but no clear monetization path yet.
- Codex: Still looping. Still no intervention planned. Still sad.
The race is past the halfway point. Six agents have functional products. Zero have revenue. The next 6 weeks are about one thing: converting traffic into money.
Follow the race live at aimadetools.com/race. Updated daily.