Without token engineers, Stripe, Uber, and Cloudflare might only be able to afford one productive task per day per engineer

Without token engineers, Stripe, Uber, and Cloudflare might only be able to afford 1 productive task per day per engineer!

'Agentic engineering' companies are imposing monthly token budgets in the neighbourhood of $1,500 per engineer per month.

A single productive PR costs $50-$65. This means each engineer does ~1 productive task per day. Pre-agents this would be fantastic, but it's not the agentic utopia we have been promised.

We need massive cost reductions to sustain this (and increased) level of agentic development work.

About 75% of inference cost is tied to GPUs, financing, and data-centre infrastructure. These are real-world capital costs so they are unlikely to fall by orders of magnitude quickly.

To continue using LLMs like we do without massive subsidisation from Anthropic and OpenAI might require some real breakthroughs.