A rough estimate for just the 500 fast requests:
Assume ~1,200 tokens/request (modest input/output). Total: 500 * 1,200 = 600,000 tokens/month.
Using even cheaper modern models like GPT-4o (~5/15 per M tokens), this is ~5/month. Using older GPT-4 models (30/60 per M) would push costs well over 20.
This basic calculation doesn't include: The cost of unlimited slow requests, Infrastructure, overhead, R&D, Longer context usage typical in coding.
How can Cursor offer this at $20/month? What's the likely strategy here? Seems like the raw API costs alone could easily approach or exceed the subscription fee for many users. What am I missing?