The April revenue release got most of the attention. Anthropic at roughly thirty billion in ARR, OpenAI at twenty-four to twenty-five depending on whose accounting you trust. We wrote about the revenue mix on this blog a week ago and concluded that the structural point under the headline was that the two companies are no longer building the same product through the same channel. The headcount chart sitting next to that revenue chart is the part we think tells you more about the next eighteen months.
The current numbers, as best we can read them from public sources: Anthropic sits at roughly one thousand five hundred employees with four hundred and fifty-two open roles. OpenAI has signaled it intends to roughly double its headcount to eight thousand by the end of 2026. The two companies are running on the same chart in revenue and on a roughly five-times difference in headcount. The interesting question is what each of those numbers is buying.
What the smaller number is buying
Anthropic’s run rate per employee, at thirty billion of ARR and fifteen hundred staff, sits at roughly twenty million dollars in annualized revenue per head. That is not a number that holds in any other software company at meaningful scale. It is a number that exists for one of three reasons. Either the revenue figure is overstated by more than the disputed eight billion. Or the cost structure has compressed in a way the SaaS playbook from 2019 did not anticipate. Or the company is genuinely understaffed for the revenue it is recognizing, which would imply a hiring catch-up that has not yet shown up in headcount totals.
The four hundred and fifty-two open roles point at the third explanation as the operative one. A company at fifteen hundred people with four hundred and fifty open requisitions is signaling that it expects to be at nineteen hundred or two thousand within a year. That is a thirty percent headcount expansion off the current base, which is steep for any company except one that has just discovered it is leaving margin on the table by not hiring fast enough. The implicit theory is that the model itself, the infrastructure to serve it, and a few hundred go-to-market and applied engineering hires can carry a revenue base larger than most companies fifty times their size.
We notice the absence of a particular line in their hiring plan. There is no public push toward consumer marketing, no expansion of a retail-style support apparatus, no acquired user-acquisition team. The four hundred and fifty open roles concentrate around model research, infrastructure, applied work with named enterprise accounts, and the developer-product layer around Claude Code. That is what an enterprise-first, API-first revenue mix looks like when it is consciously chosen rather than stumbled into.
What the larger number is buying
OpenAI’s plan to double to eight thousand reads differently. Some of that hiring is the same model-research and infrastructure work as Anthropic, scaled to the larger compute commitment. A meaningful share of it is the apparatus that a consumer-scale product requires: trust and safety teams sized for hundreds of millions of weekly active users, regulatory and policy work across more jurisdictions, the customer-support function that comes with consumer-priced subscriptions, a sales motion that has to talk to both retail buyers and procurement at the same time.
It is also, structurally, the cost of running two product lines in the same building. ChatGPT and the API are different products with different release cadences, different abuse profiles, different SLAs, and different roadmaps. A company that runs both ends up duplicating capability. The marginal hire on the consumer side is not the same hire as the marginal hire on the API side, and the org needs both. That duplication is not waste in the way a single-product company would book it. It is the price of owning more of the surface.
The directional cost claim, that Anthropic spends roughly a quarter as much per training run as the company it just passed, fits into the same explanation. We do not have audited numbers on either side, but if even half the difference is real, it lets a smaller company carry a research function at scale on a smaller cost base, which lets the headcount stay smaller, which compounds back into per-employee revenue. The variables move in the same direction.
What this is not about
The temptation when reading these numbers is to call one of them efficient and the other one bloated. Neither label is useful. A company building a consumer-scale product with multi-hundred-million weekly users cannot run on fifteen hundred people. A company that has deliberately not built a consumer-scale product does not need eight thousand. The org shape is downstream of the product shape. If you swap the products, the headcount difference compresses in either direction.
What the two numbers are about is the bet each company is making on which workload will compound. Anthropic is staffing for an API-first, code-and-agent-tooling workload growing through enterprise adoption. The plan opposite is staffing for both, with a heavier weight on consumer reach as the durable moat. Both bets can pay off. Both can partially pay off. The comparison stops being legible if one of the companies spins off the part that does not fit, which we do not think is likely in 2026 but is no longer unimaginable.
Why we read this number more than the revenue one
Revenue can be re-recognized. Accounting treatments can be argued. Disputed ARR can move four billion either direction without a single underlying customer changing what they are paying for. The headcount number cannot move like that. People are hired or they are not. Open roles are listed or they are not. The chart is harder to disagree with.
It also leads the revenue chart by about a year. The headcount a company carries at the end of 2026 is the company that will be earning revenue in 2027. A four-hundred-and-fifty-role hiring plan is a forward-looking statement about which workloads the hiring company expects to be worth concentrating around in 2027. A five-thousand-role hiring plan is a different forward-looking statement. The question for the next chart cycle is which of those workloads compounds the way the staffing plan assumed it would.
What we will be reading in next quarter’s release is not the top-line ARR. It is whether Anthropic’s open-role count has stayed in the four-hundreds, whether OpenAI’s eight-thousand target has held, and whether the per-employee revenue ratio on either side has moved enough to suggest the org shape is converging or diverging from the products it is supposed to be serving. Those three readings will say more about which substrate is winning than another quarter of ARR headlines will.