Week 3 of The $100 AI Startup Race was defined by one thing: infrastructure changes producing immediate results. We tripled DeepSeek’s sessions after discovering its frontier model costs $0.13/session. We unblocked Gemini after 18 days of “I’m blocked” commits. We fixed Claude’s broken auth. And two agents hit their subscription walls.
The race is separating into tiers. Compare with Week 2’s distribution wall — the agents that solved distribution are now pulling ahead.
The standings
| # | Agent | Product | Week 3 commits | Cumulative highlight |
|---|---|---|---|---|
| 🟠 | Kimi | SchemaLens (SQL schema diff) | 62 | 5 distribution channels, only paid product ($39), 32+ tools |
| 🔴 | DeepSeek | Spyglass (competitive intelligence) | 166 | 125 tools DB, 83 blog posts, 26 “Why X Won” analyses |
| 🟡 | Xiaomi | APIpulse (LLM pricing comparison) | 127 | 146 pages, 96 blog posts, 33 models tracked |
| 🟣 | Claude | PricePulse (SaaS price tracker) | 94 | 169 pages, 44 companies, free API launched |
| 🔵 | Gemini | LocalSEOGen (local SEO tools) | 467 | Unblocked → immediate explosion of real work |
| 🟤 | GLM | FounderMath (startup calculators) | 41 | 7 calculators, 30 blog posts, most efficient per session |
| 🟢 | Codex | NoticeKit (AI procurement compliance) | 442 | Real features hidden in validation noise, 0 users |
The big stories
DeepSeek: What happens when a frontier model costs nothing
On Friday we discovered that DeepSeek V4 Pro’s stacked discounts (75% off + cache hits at 1/10th) make it cheaper per session than V4 Flash. Our billing data showed $0.13 per 30-minute Pro session vs $1.04 per Flash session.
We immediately tripled DeepSeek’s sessions to 6/day. The results were instant — 26 competitive analyses in one week for $5:
- 65 commits over the weekend alone
- Built a competitor monitoring dashboard
- Expanded the tools database to 125 entries
- Published 15 more “Why X Won” competitive analyses (now at 26 total)
- Created a Competitive Pulse embeddable widget
- Added RSS feed and newsletter (4 issues sent)
- Built a battle card generator and threat score tool
At $0.78/day for 6 frontier-model sessions, DeepSeek is now the most cost-efficient agent in the race by a massive margin. It’s building a genuine content moat — 83 blog posts and 26 deep competitive analyses that would take a human team weeks.
Cost for the entire week: ~$5. That’s less than a single Claude Sonnet session.
Gemini: From 18 days of “blocked” to 467 commits
The most dramatic turnaround in the race. On Friday night we updated one file and Gemini exploded:
- Updated HELP-STATUS.md to explicitly say “YOU ARE NOT BLOCKED”
- Closed all 4 open GitHub issues with a clear response
- Confirmed domain (localseogen.com), SendGrid, and database are all configured
Gemini’s response: it deleted its I_AM_COMPLETELY_BLOCKED_PLEASE_HELP.md file and immediately started building. In 48 hours:
- Security fixes across 15+ API endpoints
- Product Hunt launch prep
- Email extraction with Playwright
- Page Credit Packs pricing system
- Referral tracking
- Case study page
- SEO page generator improvements
- Outreach email generation
- SVG logo and UX enhancements
467 commits in one week — the highest of any agent. Most of it real product work. The lesson: Gemini was never technically blocked. It was psychologically stuck. The model needed explicit permission to proceed, not infrastructure.
Claude: Back from the dead
Claude’s auth token expired sometime around May 7. Every session since then failed with “Not logged in · Please run /login.” We fixed it Friday night.
Since then, Claude has been running 2 sessions/day and building steadily:
- Launched a free PricePulse API
- Built watchlist pages for 44 companies
- Created distribution guides and deployment framework
- Wrote a Hacker News launch post
- Added 6 new pricing pages
- Now at 169 total pages with ~310K monthly search potential
Claude’s strategy is clear: dominate long-tail “X pricing 2026” searches. It’s the most SEO-focused agent in the race. But the traffic data shows it dropped 58% this week — the auth outage killed momentum.
Kimi and GLM: The subscription wall
Both agents hit their provider quota limits and went dark:
- Kimi: “You’ve reached your usage limit for this billing cycle” since May 8. Sessions fire but immediately fail. The Moderato subscription doesn’t provide enough quota for 4 sessions/day.
- GLM: “Weekly/Monthly Limit Exhausted. Resets at 2026-05-11 15:39:10.” The Z.ai Coding Lite plan ($18/mo) runs out by Thursday every week, leaving GLM offline for 3 days.
This is the subscription model’s fundamental weakness. API agents (DeepSeek, Xiaomi) never hit hard walls — they just spend more money. Subscription agents (Kimi, GLM, Claude, Codex) face binary on/off states when quotas expire.
Kimi was the race leader going into Week 3. It ends the week with zero weekend activity, zero sales after 112 days, and a Founding Member program that nobody has signed up for. The leader is stalling.
Codex: The commit message problem
442 commits this week. Sounds impressive. But the vast majority are “Refresh validation status and collapse backlog summaries” — timestamp updates across 10 status files every 30 minutes.
Hidden in the noise: real features. AI Procurement Hub, Vendor Risk Assessment Worksheet, AI disclosure packets, source-aware intake routing. Codex IS building product. It’s just terrible at communicating what it did.
The deeper problem: 20 outreach emails sent, zero replies. Zero contact form submissions. Zero interviews. NoticeKit has a working product that nobody is using.
Xiaomi: The content machine keeps running
127 commits, steady as always. APIpulse is now at:
- 146 pages, 96 blog posts
- 33 models tracked across 10 providers
- 9 interactive tools
- Pricing freshness badges on 23 pages
- Model Switch Calculator (new this week)
- Savings Calculator (new this week)
Blocked on a Stripe success URL redirect — Pro sales can’t complete. Otherwise the most feature-complete product in the race.
Week 3 by the numbers
| Metric | Value |
|---|---|
| Total commits (all agents) | 1,399 |
| Highest commits | Gemini (467) |
| Lowest commits | GLM (41) |
| Agents hitting quota walls | 2 (Kimi, GLM) |
| Agents with auth/infra issues fixed | 3 (Claude, Gemini, DeepSeek) |
| Revenue across all agents | $0 |
| Days since race start | 21 |
| DeepSeek daily cost | $0.78 |
The new standings
Tier 1 — Real products with distribution channels:
- 🟠 Kimi — Still the most complete distribution: npm CLI, VS Code extension, Chrome extension, Gumroad ($39 product), awesome-lists. 32+ micro-tools. Quota-walled this weekend but the product and channels remain. Needs plan upgrade to maintain lead.
- 🔴 DeepSeek — Fastest-growing agent. 6 sessions/day producing a content moat (83 blog posts, 26 competitive analyses, 125-tool database). Best cost efficiency in the race.
- 🟡 Xiaomi — Most feature-complete product. 146 pages, 96 blog posts, 33 models, 9 tools, Stripe integration. Blocked on one redirect URL.
Tier 2 — Strong products, building momentum: 4. 🟣 Claude — 169 SEO pages, free API, 44 companies tracked. Clear long-tail strategy. Back online after auth fix. 5. 🔵 Gemini — Explosive recovery (467 commits). Real features shipping fast. Needs to sustain this pace to climb. 6. 🟤 GLM — Most efficient builder (7 calculators, 30 blog posts in minimal sessions). Quota wall limits it to ~4 days/week.
Tier 3 — Product exists, no traction: 7. 🟢 Codex — Working product, real features. But 20 outreach emails with zero replies. Zero users. The distribution problem is real.
What changed this week
The biggest shift: DeepSeek moved from mid-pack to #1. Not because it got smarter — because we discovered its model costs nothing and tripled its runtime. The price war between Chinese AI companies is directly benefiting the race agents that use API pricing.
Meanwhile, the subscription agents (Kimi, GLM) are hitting walls that API agents never face. The race is becoming a test of pricing models as much as AI capability.
Next week
- Google I/O (May 19-20) — Gemini 3.2 likely drops. Our Gemini agent gets an automatic upgrade.
- GLM quota resets today (May 11, 15:39 UTC) — back online this afternoon.
- Kimi needs a billing cycle reset or plan upgrade to resume.
- DeepSeek’s 75% promo runs until May 31 — 20 more days of near-free frontier AI.
- Week 4 surprise event: Acquisition Offer ($50) — coming soon.
Follow the race live at aimadetools.com/race. Daily updates in the Race Digest.