Well, I think we should mostly consider well run businesses, not personal ego experiments.
I do think the current subscriptions could continue to work at some other price point ($35/mo?) for the everyday chatbot usage that people have enjoyed. From smarter search engine use cases, to light document editing, etc.
It’s the agentic use case that breaks the model. But it’s also those agentic use cases that enable the replacement of workers the CEOs have been dreaming about. The really interesting question is in the end: In the foreseeable future will agentic use cases be cheaper than the current human delivery mechanisms, and is it therefore reasonable or ludicrous to fire the humans and replace them (and you can build in some kind of yet unknown improvement factor for technology optimization)?
Here are some estimates to compare the use cases:
Typical chatbot engagement 500-1,700 token, even multi-turn conversation rarely exceeds 20-30K tokens.
Average Claude Code session (reading 5 files, refactoring some code, testing) might come in at 100-200K tokens, Agentic research task somewhere between 100-500K tokens.
Lets say an average subscriber today uses 10 chatbot engagements. That might might max out at 50K tokens per day.
Let say an engineer completes 30 Claude Code sessions in an 8hr day to fix bugs in a Git repo. That could average out at 10M tokens, and has potential go way higher.
That’s 200X the usage. Depending on model used that is somewhere between $100-$150 per day. Or up to $2K/month, which is 100X your current subscription cost.
By the time you look at advanced users refactoring large code bases and developing new apps, likely running multiple tasks in parallel. One example estimate for a 3-agent 8hr day of heavy refactoring can be in the 100-300M/day range and can come in at maybe $200-250/day.
The answer to the question above then is: $250 of tokens/day plus the $1,400 of fully loaded cost of a senior dev ($350K/yr) is still cheaper than than 3-5 days of $1,400 of equivalent non-AI productivity. It remains favorable, and a human remains in the driver seat with an AI accelerant, although smaller overall workforce.
When you look at the broader white collar worker the math gets different, there you’re replacing $210 in labor with $150 in AI and the human can be redundant as their org counterparts self-services. Once you factor in vacations, turn-over, training costs, etc. the machine does win.
OK, this was a bit of a detour through numbers. If it will start costing you $150/day to vibe code your tools, will you keep doing it? For those of us who build pipeline tools that pay off long-term in productivity and competitiveness. In many cases yes. Will all of you here pay $150/day to tinker with cool stuff - probably not.
I think what we as Flame artists have to be prepared for to budget $25-$150+ per day in AI costs for the work we’ll be doing. Add to that considerable kit cost or cloud kit (e.g RunComfy). That is still worth it, but needs to be built into our math rather than just absorbing it in our day rate, especially since there is significant variability.
So AI is here to stay. But the free lunches and Donut Fridays (tip to the 90s tech days) are over.
