Xiaomi was dead last. 22 commits in 7 sessions. A generic waitlist builder idea. Running on Aider with MiMo V2-Pro via OpenRouter, no web search, no CLI agent capabilities.
Then MiMo V2.5 Pro dropped on April 22. We decided to give Xiaomi a fresh start.
📊 Live Dashboard | 📅 Race Digest | 💰 Budget Tracker
What we changed
| Before | After | |
|---|---|---|
| Model | MiMo V2-Pro | MiMo V2.5 Pro |
| CLI Tool | Aider | Claude Code |
| Web Search | No | Yes |
| Tool Calls | Limited | 1000+ per session |
| Token Efficiency | Standard | 40-60% fewer tokens |
| Harness Awareness | No | Yes (manages own context) |
| API | OpenRouter (pay-per-use) | Token Plan subscription ($16/mo) |
The old setup used Aider, which is a good tool but doesn’t have the agent capabilities of Claude Code. No web search, no multi-step tool orchestration, no self-verification. The new setup uses Claude Code as the harness with MiMo V2.5 Pro as the backend model.
The results: 2 sessions vs 7
In its first session (30 minutes on V2.5 Pro), the new Xiaomi:
- Chose a new startup idea: APIpulse (AI API pricing comparison and cost calculator)
- Built a complete website with 6 pages
- Created IDENTITY.md, DECISIONS.md, and full backlogs
- Added SEO infrastructure (sitemap, robots.txt, structured data)
- Wrote its first blog post
- Submitted a help request for a domain
In its second session (30 minutes on V2.5 Standard):
- Integrated Stripe payments
- Built a Pro features page with access code gate
- Added email capture with serverless API
- Wrote 7 more blog posts (pricing comparisons)
- Added dark/light theme toggle
- Added social sharing buttons
- Added analytics and conversion tracking
- Fixed mobile responsiveness across all pages
42 commits in 2 sessions. The old Xiaomi had 22 commits in 7 sessions.
The new startup: APIpulse
APIpulse helps developers compare AI API pricing across providers and calculate monthly costs. It’s a much better idea than WaitlistKit (the old startup) because:
- Clear value proposition: Developers need this. API pricing is confusing.
- Natural SEO play: Every “X vs Y pricing” comparison is a search query.
- Revenue path: Pro features (detailed cost breakdowns, alerts) behind a paywall.
- The agent knows the space: V2.5 Pro is an AI model. It understands AI API pricing intimately.
The site is already live at getapipulse.com with Stripe integration.
The cost tradeoff
V2.5 Pro is expensive on credits. The Token Plan subscription ($16/month) gives 200M credits, but V2.5 Pro consumes at 2x rate. One 30-minute Pro session used ~40M credits (20% of the monthly budget).
We’re running Xiaomi on a conservative schedule: 1 off-peak Standard session per day (0.8x credit discount). That gives about 9-10 sessions for the rest of the month.
The quality-per-session tradeoff is worth it. Two V2.5 sessions produced more than seven old Aider sessions.
What this means for the race
Xiaomi went from the weakest setup to one of the strongest overnight. The question is whether the credit constraints (fewer sessions per day) offset the quality advantage. Gemini gets 8 sessions/day. Xiaomi gets 1. But Xiaomi’s single session produces more coherent output than Gemini’s 8 combined.
This is the quantity vs quality experiment within the experiment.