AI summary by TwinMind
What nobody tells you about captchas
Theo - t3․gg
31m 3s
Initial Problem & Abuse
- T3 Chat launched with minimal restrictions, only GPT-4o Mini and Deepseek V3 models
- Someone submitted 300,000 requests in 30 minutes
- Needed CAPTCHA implementation urgently
- Key lesson: put caches on popular services
Cloudflare Turnstile Experience
- Recommended by friend (Helium browser creator)
- Setup: Better than expected for Cloudflare product
- Annoying React integration with official library
- Component wanted to be embedded in forms with auto-token generation
- Needed token ready for message submission verification
- Pricing: $0 (major advantage)
- Limited to 2 Turnstile keys without enterprise account
- Major Issues:
- Browser debugging interference during development
- Pulled developers into source files when console open
- Halted app execution on debug steps
- Affected many dev users who keep console open
- Invisible mode reliability problems
- Could only set mode in dashboard, not code
- One mode per Turnstile key limitation
- 30-50% failure rates on invisible mode (SocialBlade example)
- <10% failure rates when switched to challenge mode
- No hybrid invisible/challenge mode option
- Would require complex custom implementation with separate API keys
- Browser debugging interference during development
reCAPTCHA V3 Migration
- Motivation: Google Cloud credits from Y Combinator
- Pricing: $1 per 1,000 assessments after 10,000 free monthly
- Would cost tens of thousands for T3 Chat volume
- Mitigated by: only unpaid users, per-user vs per-message verification, GCP credits
- Documentation: "Most poorly documented piece of software in decades"
- Created custom gist for portfolio companies
- Had to ignore official documentation
- Implementation: Simple code once figured out, better ergonomics than Turnstile
- Reliability: Much better invisible mode than Turnstile
- Less privacy-friendly due to extensive fingerprinting
- Poor performance in privacy-focused browsers like Brave
- Scoring System: 0-1 scale instead of pass/fail
- Google recommends 0.8 threshold
- Industry standard: 0.7
- T3 Chat used 0.5 due to high failure rates
- Usage Stats: 550,000 protected requests over 30 days
- At 0.5 threshold: ~10-11K risky, 116K safe
- At 0.7 threshold: 22K risky, 100K safe
- At 0.8 threshold: ~1/3 of requests failing
- User Experience Problems:
- Frequent failures on Safari iPhone at events
- No conditional challenge system
- Would require custom implementation for hybrid approach
hCaptcha Implementation
- Pricing: $1 per 1,000 evaluations + $100/month professional tier minimum
- Key Feature: 99.9% passive mode
- Invisible by default for most users
- Automatically promotes to challenge when needed
- Built-in hybrid functionality
- Implementation: Relatively straightforward
- Still requires DOM manipulation for invisible widget
- Better than alternatives despite React integration challenges
- Results: Dramatic improvement
- reCAPTCHA: 5% failure rate in last 24 hours
- hCaptcha: 0.32% failure rate in last 24 hours
- Zero server-side token failures in 24 hours
- Only 16 challenges shown in 24 hours total
- ~20 total users ever saw challenges
- User Experience: Massive improvement
- No user complaints since implementation
- Previously: dozens of complaints daily
- Works well across devices/browsers
- Even privacy-focused setups rarely trigger challenges
Comparative Analysis
Setup/Integration (1-10 scale)
- Turnstile: 6/10
- reCAPTCHA V3: 3/10
- hCaptcha: 7/10
Invisible Mode Reliability
- Turnstile: 0/10 (non-functional)
- reCAPTCHA V3: 4/10
- hCaptcha: 9/10
Cost
- Turnstile: 10/10 (free, but enterprise limitations)
- reCAPTCHA V3: 4/10 (expensive, GCP credits help)
- hCaptcha: 3/10 (expensive upfront fee)
Documentation
- Turnstile: 4/10
- reCAPTCHA V3: 1/10 (worst ever experienced)
- hCaptcha: 7/10
Development Testing
- Turnstile: 6/10 (console interference)
- reCAPTCHA V3: 2/10 (awful)
- hCaptcha: 4/10 (works with preview environments)
Promote from Invisible Feature
- Turnstile: 0/10 (doesn't exist)
- reCAPTCHA V3: 0/10 (doesn't exist)
- hCaptcha: 10/10 (core feature)
Additional Considerations
- CAPTCHA Solver Economics: Bright Data charges $1.50-15 per 1,000 hCaptcha solves
- Only 5% markup over protection cost
- Cost-benefit analysis important for high-value data
- Alternative Solutions: WorkOS Radar mentioned for future user-level fraud prevention
- Beta vs Production: 1/3 of users moved to beta, representative traffic patterns
