The Week AI Got Priced by the Job, Not the Token
When a model costs ten cents per million tokens, the question stops being 'how smart' and starts being 'how many times do I run it'.
For two years the model market argued in tokens - cents per million, context window size, benchmark points. This week that argument got quietly upended. Meta's Muse Ember launched at $0.10 per million input tokens, aimed squarely at the tool calls and coding subtasks that agents run thousands of times a day. Anthropic's Claude Fable 5.2 came in at $8 per million with a coding record. Google's Gemini 4 Carbon sits at $2 per million for frontier work. The spread isn't a quality ladder; it's a labor market.
Here's what that means in practice. A chatbot you talk to ten times a day is a rounding error at any of these prices. An agent that calls a model to check a form, classify a lead, and draft a reply, ten thousand times a day, is a real line item - and at $0.10 per million, the cheap model wins even if it's a little worse, because the task is repetitive and tolerant of small errors.
I think the frontier-vs-cheap framing is now misleading. The right split is 'frontier for the hard one-off' and 'cheap for the boring loop'. Teams that keep paying frontier prices for the loop are overpaying the way they used to overpay for dedicated servers to run a static site. The cheap model doesn't have to be as good; it has to be good enough and cheap enough.
The catch is trust. A $0.10 model running a thousand times a day needs guardrails, because one bad classification repeated at scale is worse than one bad answer you read. That's the part the pricing headlines skip: the cheap model only wins if you've built the failure path, and that work doesn't show up in the token price.
My read: we're watching the model market split into wages and salaries. Frontier models are the specialists you call for the hard case; cheap models are the shift workers running the repetitive jobs. Most companies' real cost is the shift, not the specialist.
Frequently asked questions:
Q: Does a cheaper model mean worse output?
A: Often slightly, but for repetitive tasks that tolerate small errors, the savings outweigh the drop.
Q: Why does per-job pricing matter more than per-token?
A: Because agents run the same task thousands of times, so the unit that matters is the run, not the word.
Q: Should I switch production to the cheapest model?
A: Only for repetitive, low-stakes loops, and only after you build a failure path for bad outputs.