AI pricing has taken an interesting turn. While AI providers offer usage-based pricing—you pay only for what you use—consumer applications have doubled down on the old SaaS playbook: fixed monthly subscriptions with built-in buffers to ensure most users never approach the actual cost of their usage. This is the buffet model applied to software. You pay a fixed price and the business counts on most customers consuming well below the break-even point. # The $20/month problem Here's what I'm seeing in the market: - Perplexity for search: $20/mo - Limitless for voice transcription: $20/mo - OpenAI/Claude/Gemini for chat: $20/mo - Raycast AI for general LLM use: $20/mo - Notion AI for notes: $20/mo - Cursor/Windsurf/Copilot for coding: $10-20/mo - Superhuman for email AI: $30-40/mo None of this includes direct API access, which any technical user would want. The traditional SaaS playbook says: bundle everything together. Want AI search? Here's some Google Drive storage too. Want Notion AI? Here are all our pro features. Want Raycast AI? Here's window management and clipboard history. What bothers me is that underneath, this is all just text processing. LLMs excel at text, but I'm paying multiple providers to essentially copy-paste different text to the same underlying models. From the user's perspective, it would be far more efficient to pay for AI access once and feed it whatever I want: code, documents, emails, search queries. I don't want to pay $20/month to five different companies when I know they're all doing the same thing under the hood. # The appeal of usage-based pricing API pricing is sort of the antithesis of this subscription model. You pay only for what you use—no buffers, no unused quota. The SaaS promise is that heavy users (maybe, potentially) get a great deal, but in practice few users can saturate their subscriptions, especially when they're locked into specific use cases or specific text sources. If you pay for Notion AI, you can't use it for search, coding, or email. The $20/month is artificially constrained to one workflow. ## Cursor's interesting experiment Cursor has done something interesting with their premium models: they charge 20% on top of API costs. This cost-plus model seems to roughly align incentives: - Light users pay a small premium on occasional usage - Heavy users pay proportionally for their actual consumption - The company gets their cut without the quota-management overhead Of course, Cursor still has their $20/month base subscription, which I suspect isn't going anywhere. But if they fully embraced the cost-plus model, it would be very compelling. Their revenue would be much less predictable though, so I suspect we won't see this anytime soon. # BYOK Direct API access puts downward pressure on these subscription prices. Consumers can use APIs just like businesses, and many apps now let you "bring your own key" (BYOK). I was surprised to see Raycast offer this option—it seems almost altruistic since they're essentially giving up subscription revenue. But maybe they're capturing users like me who would never subscribe to proprietary AI wrappers that offer (IMO) very limited value-add, but will gladly use their excellent interface with my own API keys. ## Claude's unified approach Claude is doing something I wish everyone would adopt: unified pricing across interfaces. One subscription gives you: - Web interface with search capabilities - Claude Code for command-line operations - API access for custom integrations This is the model I want to see everywhere. Pay once, use however you want. --- My hope is that the future isn't dozens of $20/month subscriptions for the same underlying capability. Usage-based pricing with unified access across interfaces makes far more sense. It also doesn't necessarily mean lower subscription costs either. In the case of Claude, I know many people who are happily paying $100/mo or more for the "max" plans they offer, specifically because they can use the coding agent (primary use case) as well as the general purpose knowledge assistant (i.e. chat "ChatGPT" use case).