AI lab headcount and revenue per employee: Anthropic, OpenAI, others

Anthropic reported more than 3,500 employees in September 2026, and OpenAI was reported at about 4,500 on payroll in March 2026. Epoch AI estimates revenue per employee near $9 million for Anthropic and $5.6 million for OpenAI, but those use annualized revenue and are approximate.

Headcount sounds like the simplest number in business. For AI labs it is surprisingly hard, because neither company publishes it, counts use different scopes, and revenue figures move faster than staffing figures. This guide lays out what has been reported, explains what revenue per employee does and does not tell you, and shows the same company producing figures that differ by a factor of four. It is not investment advice. All figures are as of 19 October 2026 unless dated otherwise.

Key numbers at a glance

• Anthropic employees, September 2026: more than 3,500 (Anthropic spokesperson, via the Wall Street Journal).
• Anthropic employees, end of 2025: about 2,300 (Fortune, June 2026).
• Anthropic employees, January 2026: more than 2,500 (San Francisco Chronicle).
• Anthropic employees, 6 April 2026: 3,328 (Epoch AI companies database).
• OpenAI employees, March 2026: about 4,500 on direct payroll (Financial Times, via CNBC, two unnamed sources).
• OpenAI target by end of 2026: about 8,000 (Financial Times, 21 March 2026).
• OpenAI Foundation workforce, December 2025: 7,850 (Revelio Labs, built from public LinkedIn records, a wider scope than payroll).
• OpenAI employees, 28 February 2026: 4,433 (Epoch AI).
• Revenue per employee, Anthropic: about $9.0 million, using $30 billion annualized revenue (Epoch AI, April 2026 data).
• Revenue per employee, OpenAI: about $5.6 million, using $25 billion annualized revenue (Epoch AI, February 2026 data).
• Comparators from annual reports, per Epoch: Nvidia about $5.1 million, Meta about $2.5 million, Apple about $2.5 million, Google about $2.1 million.
• Anthropic 2025 booked revenue per year end employee: about $2 million (a tracker's rough calculation from $4.6 billion and 2,300 staff).
• DeepSeek: 344 employees, March 2026 (Revelio Labs, via a data page).
• Mistral AI: about 350 employees, June 2026 (a tracker's estimate).

Why counting staff is hard

Neither Anthropic nor OpenAI has published an official headcount. One page tracking Anthropic notes that it confidentially submitted a draft registration to the US securities regulator on 1 June 2026. That matters because US disclosure rules ask a registrant for "a description of the registrant's human capital resources, including the number of persons employed by the registrant" (Regulation S-K, Item 101(c)(2)(ii)). A public prospectus would therefore carry a dated employee count that the company stands behind. Until one is public, every figure here comes from spokespeople, newspapers, data providers or estimates.

Four things go wrong when people compare these numbers.

First, scope. A payroll count includes people on the company's books. A social network based count, such as Revelio's, counts people who list the company as their employer, which can include related entities, contractors or people who have left. That is the most likely reason OpenAI shows up at about 4,500 (payroll, Financial Times) and 7,850 (Revelio, December 2025, for the OpenAI Foundation). One secondary article attributes most of that gap to method rather than a real change in staffing.

Second, timing. A count dated March is not a count dated September. Companies hiring fast can add a thousand people between two reports.

Third, different sources for the same date. For Anthropic, estimates for early 2026 range from about 2,500 (PitchBook and an investor named in one analysis) to about 3,000 (TrueUp) to about 5,000 (Tracxn, March 2026), according to a SaaStr article. That article itself says third party counts vary.

Fourth, low outliers. One research page, last updated June 2026, lists OpenAI at about 3,500 and Anthropic at about 1,200. It does not explain its methods beyond a mix of public reporting and LinkedIn estimates, and it sits far below every other published count for Anthropic, so it is treated here as an outlier and not used for any comparison.

Anthropic: a fast climb

A timeline compiled by Exploding Topics, updated on 16 October 2026, cites a source for each step:

• Late 2023: fewer than 200 (Fortune), with another report of about 300 by December 2023 (The Information).
• May 2025: more than 800 (The Information).
• End of 2025: about 2,300 (Fortune, June 2026).
• January 2026: more than 2,500 (San Francisco Chronicle).
• September 2026: more than 3,500 (Anthropic spokesperson, via the Wall Street Journal).

Anthropic also said its EMEA headcount tripled in the year to November 2025, and Daniela Amodei told the San Francisco Chronicle there were more than 1,300 employees in the Bay Area in January 2026.

The same page cites a SignalFire report from May 2025 on retention: 80% of Anthropic hires from two years earlier were still there, versus 67% at OpenAI. It also says engineers were about eight times more likely to move from OpenAI to Anthropic than the other way. That is a talent market statistic and says nothing about finances, but it helps explain how a company can grow quickly.

Much of the growth has been outside the United States. SiliconANGLE reported on 26 September 2025, citing CNBC, that Anthropic planned to triple its headcount outside the US after a $13 billion funding round at a $183 billion valuation. More than 100 of those hires were planned for Dublin, London and Zurich, the applied AI team that helps customers deploy Claude was set to grow fivefold by the end of 2025, and Tokyo was named as the first Asia office, with country leads planned for India, Australia and New Zealand, Korea and Singapore. Those plans may account for part of the jump between the 2025 and 2026 counts, though no public figure shows how many of today's staff sit outside the US.

OpenAI: a larger and faster hiring plan

The Financial Times reported in March 2026 that OpenAI had about 4,500 employees and planned to grow to 8,000 by the end of 2026. A secondary article notes that reaching that target needs about 3,500 hires in nine months. It also notes that on 27 January 2026, Sam Altman told staff the company would "dramatically slow down" the pace of hiring, which sits awkwardly beside the 8,000 plan. Public reporting does not settle which signal reflects reality. The same article contains several internal arithmetic errors, so only its dated claims are relayed here.

That same article says engineering is about 56% of headcount, research about 25%, and commercial, support and administration about 19%. It does not name a source for the split, so read it as indicative only.

Epoch AI, an independent research group, counted 4,433 OpenAI staff on 28 February 2026.

What revenue per employee shows

Revenue per employee is annual revenue divided by headcount. Epoch AI published a comparison on 8 May 2026 using its own database. Its table lists:

• Anthropic: $30 billion annualized revenue, 3,328 staff, about $9.0 million each (April 2026).
• OpenAI: $25 billion annualized, 4,433 staff, about $5.6 million each (February 2026).
• Nvidia: about $5.1 million each, from its annual filing for the year to January 2026.
• Meta: about $2.5 million each (year to December 2025).
• Apple: about $2.5 million each (year to September 2025).
• Google: about $2.1 million each (year to December 2025).

Epoch concludes that if either lab listed publicly, it would have higher revenue per employee than any public technology company in the set it examined. It also says Anthropic and OpenAI figures are approximate because the companies report revenue and staffing at different times, and it assumes staffing keeps growing at the recent average in between. One more flag: the page shows a separate summary figure near $14 million and $6.5 million without explaining the difference. The main table values are used here because they match the inputs in its table.

How Epoch builds the comparison

The inputs are worth seeing, because they show why the lab figures are softer than the public company ones. Epoch takes the lab revenue from its own AI Companies database, which compiles company disclosures and media reports, and uses annualized figures: $30 billion for Anthropic as of 6 April 2026 and $25 billion for OpenAI as of 28 February 2026. Staff counts are reported less often than revenue, so Epoch estimates current headcount by assuming staffing kept growing at the average rate of the previous six months. The 3,328 and 4,433 figures are therefore estimates for the revenue date, not counts published on that date.

The public companies are handled differently. Their figures come from annual 10-K filings: Nvidia at $215.9 billion of revenue and 42,000 staff, Meta at $201.0 billion and 78,865, Apple at $416.2 billion and 166,000, and Google at $403.0 billion and 191,000. The peer set is drawn from the Forbes Global 2000, filtered to software, semiconductor and hardware industries. Epoch notes that Nvidia's 10-K implies about $5.1 million per employee while Forbes reports $4.4 million, a reminder that even audited inputs give different answers depending on which revenue and which headcount date are used. It also cautions that revenue per employee varies by sector, so comparisons across industries need care.

A SaaStr article makes a related comparison, saying Google crossed $30 billion in revenue with about 32,000 employees and Salesforce needed about 79,000, while Anthropic reached a $30 billion run rate with a few thousand. That article's headcount for Anthropic is a third party estimate, as noted above.

Why the same company can look like $2 million or $9 million

This is the most important section for a careful reader.

The Anthropic tracking page calculates a revenue per employee of about $2 million for 2025, using nearly $4.6 billion of booked revenue (Reuters, citing a draft prospectus) and about 2,300 employees at year end. Epoch's figure is $9 million. Neither is wrong. They differ because one divides a full year of booked revenue and the other divides an annualized recent rate. When revenue grows several times in a year, those two numbers diverge widely. Staffing also grew during 2025, so dividing by year end headcount lowers the result compared with using the average over the year, which the page says would give a higher figure.

Run rate adds one more wrinkle. The Financial Times and CNBC reported on 8 October 2026 that OpenAI's run rate was about $50 billion at the end of September, after an earlier figure near $70 billion had reportedly counted partner sales on a gross basis. A revenue per employee that uses a gross run rate looks better than one that uses net. The guide to reading AI lab financial reports explains gross versus net and run rate in detail, and OpenAI revenue, losses and cash burn lists the underlying revenue figures.

Finally, remember that revenue per employee says nothing about profit. A company can generate high revenue per head while spending far more than it earns on computing power. Anthropic and OpenAI both reported large losses in 2025, according to the reporting summarised in the linked posts, so high revenue per employee is a sign of a small, compute heavy workforce, not of profitability.

Smaller labs

Data for other labs is thin and noisier.

• DeepSeek: 344 employees in March 2026, according to a page citing Revelio Labs, which also shows internal inconsistencies in its yearly series. The research page mentioned above lists about 250 or more. No reliable 2026 revenue figure has been published.
• Mistral AI: a tracker lists 276 employees in March 2025 (from LinkedIn) and about 350 in June 2026 (an estimate from a funding news site). It also lists annual recurring revenue passing $400 million in February 2026, attributed to another site, and its own revenue entries carry no recorded source. That revenue figure is too weakly sourced to support a per employee comparison.
• xAI: only an estimate of about 1,000 or more appears in the low quality research page, and no standalone revenue figure has been published.

The lesson is that revenue per employee for private labs outside the two largest is mostly guesswork today.

What is firm and what is uncertain

Firmer:

• Anthropic's September 2026 headcount of more than 3,500 came from a spokesperson, though the number is stated as a floor.
• OpenAI was reported at about 4,500 on payroll in March 2026, by the Financial Times.
• Epoch's method is published, and its table includes the date of each input.

Uncertain:

• Every revenue per employee figure, because revenue is annualized and headcount lags.
• OpenAI's headcount now, given the 8,000 target and the signal to slow hiring.
• Third party estimates for smaller labs.
• Any breakdown by department.

What to watch next

• Public prospectuses, which should state headcount and booked revenue for the same period.
• Whether OpenAI reaches its stated 8,000 target by year end.
• Whether Epoch or others update the comparison using booked revenue instead of run rate.
• Anthropic's next spokesperson figure after September's more than 3,500.

Quick check

Sources

• Anthropic and OpenAI earn more revenue per employee than major public tech companies, Epoch AI, 8 May 2026.
• Number of Anthropic employees (October 2026), Exploding Topics, updated 16 October 2026.
• How many people work at OpenAI in 2026, SQ Magazine, 2026.
• Anthropic only has about 5,000 employees, SaaStr, 2026.
• AI lab employee headcount 2026, Presenc AI, updated June 2026 (cited only as a low outlier).
• DeepSeek employees, Revelio Labs data page, 2026.
• Mistral AI, Latka, 2026.
• Anthropic to triple international headcount and add offices following latest funding round, SiliconANGLE, 26 September 2025.
• 17 CFR 229.101, Description of business, Regulation S-K Item 101, Legal Information Institute.
• Is Anthropic profitable in 2026?, Value Add VC, updated 15 October 2026.

Last reviewed: 19 October 2026.