Despite 'ARR' buzz, Anthropic opened half as many roles as OpenAI

Commentators debated revenues, model costs and agent design; our tracker shows 4,101 open AI jobs, skewed to engineering, with OpenAI adding 62 roles in 30 days to Anthropic’s 33.

Revenue boasts, pricing worries, and the jobs reality

Across the last two days, several practitioners pushed competing stories about momentum in AI. Simon Willison highlighted a Financial Times write-up with, in his words, "A few interesting numbers in this FT story gathered from "people with knowledge of the matter":" including claims that Anthropic's "annualized revenue" is surging and that "It also told investors that it had 6,000 customers that spend $100,000 annually or more." Gary Marcus urged caution on the very premise of the revenue discussion: "Anthropic’s boosters are rushing to celebrate their ARR." His warning was about definitions: "When people tell you Anthropic had this or that ARR, you really need to know which meaning they are talking about. They rarely tell you. But they really mean the latter."

Our tracker offers a different lens: hiring. Right now we count 4,101 open AI roles across 174 employers, taken from their own job feeds. Among the top labs, OpenAI lists 148 open roles, with 62 opened in the last 30 days. Anthropic lists 114 open roles, with 33 opened in the last 30 days. Whatever the revenue narrative, our jobs data does not show Anthropic expanding faster than OpenAI. The biggest single employer in our tracker is neither lab, but Accenture, with 606 open roles and 209 opened in the last 30 days. Capital One and Amgen each list 176 open roles, while Waymo has 101. If headline revenues or model launches are translating into labor demand, the clearest signal in our data is broad enterprise hiring rather than a surge at one frontier lab.

Model mix, cost pressure, and where teams invest

Willison also pointed to a market split by price and capability. He relayed Ramp’s billing-based view of Anthropic model usage and shared Drew Breunig’s take on how Fable’s cost changed behavior: "But then Fable landed. It was (and still is!) incredible." That sentence is followed by a flip side in the same passage about the high price pushing teams to think carefully about "what work went where."

In our hiring figures, the investment theme that shows up is in hands-on build roles. By role family, Modelling and engineering accounts for 1,610 open roles, with 382 opened and 110 closed in 30 days, across 141 employers. Data roles are 791 open, with 181 opened and 24 closed in 30 days, across 117 employers. Infrastructure sits at 390 open, with 64 opened and 18 closed in 30 days, across 90 employers. If model pricing is nudging teams to rework their harnesses and context strategies, as Breunig suggests and Willison amplifies, the corresponding labor demand in our tracker leans toward engineering and infrastructure rather than, say, net-new product management headcount.

The same ground truth shows up in Willison’s own shipping notes for llm 0.33: "llm prompt -t/--template can now be repeated to combine templates in order." It is a small, practical change that optimizes developer workflows. Our data is consistent with that day-to-day reality: companies are hiring people to build, integrate and tune, not just to shop for models.

Agents vs single systems: what employers are hiring for

Azeem Azhar revisited a classic coordination problem with new agent experiments: "After discussion, most model families chose correctly in only 17-36% of runs, while a single agent handed the entire evidence base got it right nearly every time. Only one model (somewhat) escaped: Mythos 5, at about 85% (why, we don’t know)." He frames the underlying cause as homogeneity: "First, LLMs lack diversity (they are low-variance): set 30 agents the same coding task and 18 of them will name their git branch identically."

Does this debate show up in job postings? Our tracker shows 221 open Research roles, with 33 opened and 4 closed in 30 days, across 50 employers. Evaluation and safety roles total 41 open, with 9 opened and 2 closed in 30 days, across 17 employers. Those are meaningful numbers but they are small compared to Modelling and engineering. On hiring intent, employers today are not betting the farm on novel multi-agent research teams. They are still staffing the people who can ship single-system integrations and production-grade features.

Everyday coding help is real, and that does align with hiring

Ethan Mollick described a pragmatic, low-risk use case that saves hours rather than making grand leaps: "In terms of everyday usefulness and saving time, Codex & Claude Code are very capable of doing the thing where you ask them to "fill out the forms that I got an email about" and they do it well & without further intervention. Really nice for low-risk time-consuming stuff that scattered attention." Linus Torvalds gave a grittier developer’s-eye view: "And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work."

Our data matches that ground-level adoption story. Employers are hiring for the plumbing work of building and maintaining AI-augmented software. Modelling and engineering dominates openings. Data roles are the next largest pool. Product and design roles are comparatively few at 98 open, with 20 opened and 2 closed in 30 days, across 49 employers. This pattern supports Mollick’s and Torvalds’ focus on code assistants and development acceleration rather than splashy top-down programs. Companies appear to be staffing the people who can make those assistants productive inside their own stacks.

Watermarking is a product feature, not yet a hiring wave

Sebastian Raschka offered a clear explainer on trust signals: "Hi everyone. So, a few days ago, Anthropic announced that they will watermark the text outputs of their Claude models." He noted the interest in the explanation itself: "I then did a social media post briefly explaining how that works." and "So not the watermarking itself was popular, but I guess the explanation or the mechanism behind it."

Our tracker does not show a watermarking-driven hiring spike. Evaluation and safety roles remain a small slice at 41 open across 17 employers. That does not contradict Raschka, who is documenting a mechanism, not forecasting hiring. It does indicate that employers are not adding large teams just to manage watermarking right now. Most of the headcount is going to the people who will wire such features into products.

Skepticism about hype, and what the postings say

Alex Hanna’s take on hype deflation was blunt: "Well now they’re just saying anything." and, in a different post, ""I just care about academic freedom and plagiarism" is the new "it's about ethics in gaming journalism"" Gary Marcus struck a similar chord on headline narratives with a short alert: "This is bad news. But wait, there’s more. Read more"

Set against that commentary, our numbers are steady and concrete. OpenAI and Anthropic are hiring, but the bigger story is activity across integrators and end users. PwC lists 89 open roles, Databricks 88. Capital One’s 176 open roles underscore demand in regulated industries. If there is doublespeak in the market, the job feeds cut through it: we see who is adding people and in what functions.

Bots everywhere, but fewer jobs tied to content moderation

Mollick also flagged a social side effect: "Annoying side effect of all the AI bots on the site is that they are all “well-read” and therefore reply cogently my niche reference tweets that would normally attract a small but interested group now get LLM replies. Its like you can’t tell Ephraimites from Gileadites anymore." That phenomenon might suggest a need for more moderation or safety roles. Our tracker, however, counts just 41 Evaluation and safety openings across 17 employers. Again, the hiring signal is that companies are prioritizing build roles first.

What the next month could prove

Two practical patterns emerge from the last 48 hours of commentary and our jobs data:

Revenue and usage headlines are noisy. Hiring is not exploding at the frontier labs compared to the broader market. OpenAI opened 62 roles in 30 days, Anthropic 33. Accenture opened 209 in the same period.
The bulk of demand is for people who build. Modelling and engineering, Data, and Infrastructure together account for 2,791 open roles. Research and Evaluation are meaningful but smaller.

Azeem Azhar’s reminder that "Great minds think (a little too much) alike" suggests research investments will need to counter homogeneity. Whether that becomes a hiring wave is not visible yet. In the meantime, Simon Willison’s shipping note and Drew Breunig’s price-conscious workload planning are visible in the kinds of roles employers are posting. We will keep testing splashy claims against who is actually hiring, in what functions, and how fast those postings are opening and closing.

What we read

Every quote above is taken verbatim from one of these posts.