The Calculator Arrives: Agent Economics Moves from Theory to Price Tag

For two years, the agent economy has operated in a space somewhere between benchmark and belief. The benchmarks showed agents could do things; the belief was that the economics would eventually work out. This week, someone published a calculator.

METR's new metric — which the organization calls a human-cost crossover point — is straightforward in concept and significant in implication. Given a task category and a model, it computes the rate at which agent cost per unit of work exceeds the market rate for human labor performing the same task. The metric doesn't argue that agents are good or bad, cheap or expensive. It just answers a question every enterprise buyer has been asking without a clean way to frame it: at what volume and complexity does this actually pencil out? That question now has a formula.

The timing is deliberate. METR released this alongside a broader push by the industry to get serious about agent ROI — a conversation being forced partly by the sheer cost of frontier inference and partly by enterprises who signed up for agentic automation and are now looking at their quarterly AI spend with raised eyebrows. Google's Gemini 3.6 Flash, explicitly positioned to reduce enterprise agent token costs, and the Cursor swarm result from last weekend — frontier planner, cheap executors, 15x cost delta — are both responses to the same pressure. The architecture has a name now; METR gave it an equation.

The second thread is stranger and more interesting. OpenAI released data this week showing a rising share of ChatGPT users using the model to do work on behalf of other people — in some cases, apparently, on behalf of people who don't know their work is being done by a model. The framing in OpenAI's release was productivity; the implication is something more structural. The agent economy's first widespread use case may not be agents replacing workers — it may be workers using AI to extend their own effective capacity into adjacent roles, absorbing tasks that used to require separate hires. One person plus ChatGPT, operating as something closer to a small team. The labor-market effect of that, at scale, is genuinely hard to model.

The enterprise deployment question is also being answered in a specific way that wasn't obvious a year ago. OpenAI Presence — the company's enterprise-agent product — bundles AI agents with human engineers who manage, monitor, and tune them. This is not the fully autonomous future that most agent demos implied. It is, however, something enterprises will actually buy. The embedded engineer model is a bridge product: it converts an unpredictable capability into a managed service with an accountable human in the loop. The fact that OpenAI is selling this tells you what the actual enterprise demand curve looks like right now.

Underneath all of this is the infrastructure investment that makes the trajectory legible. AMD's $5 billion commitment to Anthropic isn't a product bet — it's a capital commitment to the compute layer that runs the agents, at a scale that only makes sense if the economics of deployment are expected to resolve in the agent's favor. Someone at AMD ran the METR calculation and liked the answer.

The through-line: the agent economy is not waiting on a breakthrough. It is waiting on the spreadsheet to clear. This week, in several different ways, the spreadsheet got easier to fill out.