Jensen Huang says you should spend half of your annual salary on AI tokens.

While Anthropic doesn’t give a numerical figure Boris Cherny (lead of Claude Code) says you should use the best models with maximum thinking budget continuously in parallel (approx. $1,500 USD/hr).

Surprisingly, OpenAI and Google are quite pragmatic on the subject and recognise that token costs are getting out of control.

Indeed, there have been numerous reports on eye-watering token bills. $1.305M USD in one month for a single user. $500M USD in one month for a single company. Certainly, from conversations with friends and former colleagues, many companies have been bitten and are now clamping down on AI token use.

Some guidance recently takes it as far as enabling “caveman mode” or only communicating with Claude using images of text.

I listened to an interesting take from an economist Benedict Evans who likens this phenomenon to the outrageous internet bills of old. When I was in primary school, I remember streaming Triple J on the home internet running up an internet bill in the thousands of AUD.

If we extrapolate here, AI costs should settle down into subscription-based plans as a low margin service model. This is because there is a lot of competition among the AI companies which should drive token costs down in the long term. Currently, we’re in a transient period where usage, resources and allocation are unstable.The methods we use at Traffyk are to use scheduled LLM calls to optimise our use on personal accounts. Implement caps on personal token usage and use shared accounts with large token budgets for pre-approved projects.

In summary, you should rarely be spending more than your monthly subscription on AI tokens. If you’re a super-user, subscribe to the top plan and be smart about your usage. Unless you’re the first person in the world to unlock the ever elusive 1:1 productivity gain to token spend.