Phil Chen, stealth AI startup co-founder, outlines guidelines capping token compute budgets at $200 and $2,000 per project

Here's a note I posted to the team today about how I think about token budgets going forward:

Spending tokens is easy. Without even trying hard, I've been able to spend ~$2k / day in tokens with GPT 5.5.

Saving tokens is easy. Many tasks don't actually need frontier intelligence if we just use our own intelligence upfront (fixing a UI nit is not Fable-level if we prompt precisely). Spending more time and tokens upfront towards a logically sound design saves time and tokens later.

Internally, we will dynamically allocate token compute to people through a simple request system. For each project, request either $200 (1 day project) or $2000 (1 week project) in token compute if you think it is Fable-worthy. These numbers can change.

In the age of abundant intelligence, we are all becoming capital and time allocators. We should have best-in-class intuition on how much intelligence we get per unit of capital and/or time.