AI summary by TwinMind
Claude Almost Bankrupt Me...
Theo - t3․gg
16m 18s
Product Launch and Initial Success
- Launched T3 Chat weeks ago with high hopes but middling expectations
- Product was "just a wrapper for all these AI models"
- Overwhelmed by user feedback, support, and adoption
- Analytics from Plausible (initial setup before auth):
- 277,000 unique visitors
- 1.8 million page views
- 860,000 messages sent
- 160 concurrent users at time of recording
- Product under 1 month old showing exceptional growth
Initial Model Strategy and Pricing
- Originally planned to launch with only DeepSeek V3
- V3 pricing advantages (legacy rates):
- Cache hit: 1.5 cents per million tokens
- Input: 15 cents per million tokens
- Output: 28 cents per million tokens
- 20x cheaper input than Claude ($3 vs 15 cents)
- 30x+ cheaper output than Claude ($15 vs 28 cents)
- V3 quality close to Claude Sonnet, favorite general model after Sonnet
- Speed issues with V3:
- Initially 90+ tokens/second
- Performance degraded over time
- Became "barely functioning" after R1 launch
- Third-party providers couldn't achieve fast speeds
Switch to GPT-4o Mini
- Last-minute decision to switch from V3 to GPT-4o Mini
- Reasons for GPT-4o Mini choice:
- Price competitiveness
- Speed advantage: 2-3x faster on Azure vs OpenAI APIs
- Performance focus: every click responds immediately
- Azure benefits for YC company:
- Substantial credit allocation vs OpenAI's $5,000
- Same pricing as OpenAI
- Bill manageable for extended period
- Launch models: DeepSeek V3 + GPT-4o Mini, added model picker same day
- Added Claude shortly after (end of first day)
The Claude Cost Crisis
- $8 pricing based on V3 and GPT-4o Mini costs
- Claude became major expense problem:
- Bills approaching $600/day
- February Claude costs: $500/day
- Not sustainable for business
- Usage breakdown (PostHog analytics):
- GPT-4o Mini: ~650,000 messages (free tier)
- Claude: 140,000 messages (paid users only)
- Paid users generated 25% of total usage despite being minority
- Individual user costs:
- Some users cost over $200 each
- One user: 500 messages in 2 days = $180 cost
- Same user emailed requesting rate limit reset
- Problem concentration:
- 90%+ costs from <2% of users
- ~100 users exceeded value of their payments
- Original limit: 500 messages/week based on other model pricing
Traffic Growth and Scaling Issues
- Message volume patterns:
- Initial launch spike
- Weekend slumber first week
- Sustained weekday traffic
- Recent weekend traffic increase (no more weekend slumber)
- Nearly 10K messages/day currently
- Growth trajectory exceeded all expectations
- Rapid scaling caught team unprepared for cost implications
Pricing Model Changes
- New limits implemented:
- 1,500 messages/month total
- Claude limited to 100 messages/month (separate from core limit)
- Other models remain in core 1,500 limit
- Additional Claude credits available:
- $8 for 100 additional Claude messages
- Credits don't expire, roll over indefinitely
- Monthly allocation consumed first
- Cost analysis for additional credits:
- ~6 cents per message average cost
- 100 credits cost company ~$6
- Minimal profit margin, potentially at-cost
User Compensation and Retention
- Existing subscribers compensation:
- 250 free Claude credits for pre-change subscribers
- Estimated total cost: ~$50,000 in free credits
- Value comparison:
- T3 Chat: $16 for 200 Claude messages + 1,500 other model messages
- Official Claude: $20 for equivalent Claude usage
- T3 Chat saves $4 vs Claude subscription
Future Model Strategy
- Cheap models (under $1) likely moving to lower tier:
- GPT-4o Mini, V3, R1
- New Gemini models (cheaper than GPT-4o Mini)
- Potential credit system:
- Different models cost different credit amounts
- GPT-4 might cost 2 credits
- GPT-4o Mini might cost 0.2 credits
- Commitment to maintain $8 plan indefinitely
- Possible premium tier for expensive models (future consideration)
Model Deployment Speed
- First to deploy new models:
- DeepSeek R1 on Groq
- Official R1 deployment
- O3 Mini (first non-official site)
- Gemini 2.0 Flash and Flash Thinking within hours
- Goal: best place to access newest AI models
Team and Transparency Philosophy
- Mark (CTO) partnership crucial for execution
- Stayed up until 6 AM implementing rate limit changes
- Full transparency approach:
- Sharing actual usage numbers
- Open about costs and pricing decisions
- Not hiding dashboard, pricing, or metrics
- Positive user response to changes exceeded expectations
- Commitment to continued transparency and accountability
