AI Lab Talent Flow in 2026: Who Hires From Whom
By Kooperativa Engineering
Which companies do the AI labs hire from, and do people move between them in both directions? Most answers come from surveys or from a vendor's private data, and a reader can't check either. We looked at a sample of our own professional profile data: the current profiles at the 11 companies below, and which of the other ten each person lists as a past employer.
47 current Anthropic profiles list OpenAI as a past employer. 7 current OpenAI profiles list Anthropic. That is about 6.7 to 1.
What we counted
This is a sample of our data. It covers the current profiles at these 11 companies, picked by company and not drawn at random, and it is neither everything we hold nor everyone who works at these companies. For each pair we counted how many people at company A list company B as a past employer. A person counts when the profile shows a current position at A and an earlier position at B. The counts were run on 18 October 2026, and we publish totals only.
Past employers are matched by the exact name written on the position. Meta includes positions listed as Facebook, and Google DeepMind includes DeepMind. Google means the company called Google, so a person who listed a Google unit under its own name is not in that column. The current employer is matched by company ID.
Finding 1: Google is the most common past employer
Among these companies, Google is the most common past employer at OpenAI, Anthropic, Meta and Microsoft. It appears on 18.3% of the Anthropic profiles in the sample and 17.6% of the OpenAI profiles. Meta is next, at 16.4% of OpenAI profiles and 10.3% of Anthropic profiles. Mistral AI (7.3%) and xAI (6.7%) draw less from Google.
Google DeepMind is left out of the chart because it is part of Google. 51.6% of its profiles list Google, which fits the 2023 merger of Google's AI teams with DeepMind, along with ordinary internal moves.
Share of current profiles that list Google as a past employer
Kooperativa sample of professional profiles, counted 2026-10-18. Share of each company's profiles in the sample. Google DeepMind (51.6%) is omitted because it is part of Google.
Finding 2: some pairs are lopsided and others are even
We counted each pair in both directions with the same method. For each pair the table shows how many people at the first company list the second as a past employer, how many do the reverse, and the ratio.
| Pair (first company hires from second) | Forward | Reverse | Ratio |
|---|---|---|---|
 Anthropic from  OpenAI | 47 | 7 | 6.7 to 1 |
Anthropic from  Google DeepMind | 40 | 7 | 5.7 to 1 |
 Meta from  Microsoft | 1,424 | 362 | 3.9 to 1 |
 Google from Microsoft | 2,701 | 821 | 3.3 to 1 |
| OpenAI from Google DeepMind | 65 | 31 | 2.1 to 1 |
| Meta from Google | 1,714 | 1,496 | 1.1 to 1 |
OpenAI from  xAI | 15 | 14 | 1.1 to 1 |
Limiting both sides to people whose current position began in 2025 or later gives the same direction. 30 of the 1,600 recent Anthropic starters in the sample list OpenAI, against 6 of the 2,723 recent OpenAI starters who list Anthropic. That is 5 to 1. Meta and Google stay close to even, at 752 against 683.
The counts say nothing about why people moved. A lopsided ratio would also appear if one company had hired experienced people from the other earlier in its history.
Finding 3: the newest labs are mostly recent starters
We also counted the profiles whose current position began on or after 1 January 2025. At Mistral AI that is 84.8% of the sample, at Anthropic 82.2%, at xAI 78.9% and at OpenAI 73.0%. At Google, Microsoft and Meta it is between 34% and 37%.
The date is the start of the current position, and a promotion or a move inside the same company resets it, so it is not a hiring date. The share also depends on who has a profile and when it was last refreshed.
Share of current profiles whose current position began in 2025 or later
Kooperativa sample of professional profiles, counted 2026-10-18. Start of the current position, not start of employment.
Part of the matrix
Rows are the current company and columns are the past employer. The diagonal is blank because an earlier position at the same company is a previous role there, not a different employer.
| Current \ Past | OpenAI | Anthropic | Google DeepMind | Google | Meta | Microsoft |
|---|---|---|---|---|---|---|
| OpenAI | 7 | 65 | 657 | 611 | 289 | |
| Anthropic | 47 | 40 | 356 | 200 | 140 | |
| Google DeepMind | 31 | 7 | 1,732 | 357 | 347 | |
| Google | 36 | 10 | 81 | 1,496 | 2,701 | |
| Meta | 40 | 7 | 82 | 1,714 | 1,424 | |
| Microsoft | 5 | 2 | 10 | 821 | 362 |
What the counts cannot show
- They measure a stock, not a flow. A profile that lists OpenAI as a past employer may describe a job from years ago, and the person may have worked elsewhere since. We don't know the employer right before the current one, so none of this counts direct moves.
- Past positions include internships. Google, Meta and Microsoft take large intern cohorts, which raises their share as a past employer in every row.
- The sample holds fewer profiles than these companies have employees, and the gap differs by company. Treat them as counts of profiles and don't compare them as headcounts.
- Names are matched exactly, so a person who spelled a past employer differently is missed.
- The counts don't tell you who left by choice, who was laid off, or why.
Reproduce the Anthropic and OpenAI count
You can check the headline pair with the people search endpoint. Filter current Anthropic profiles by past employer OpenAI and read the total from a page of size 1. The past employer filter matches one exact company name, and the company filter takes a company ID. Totals move by a few as profiles are added, so a run today may differ slightly from ours.
curl -X POST "https://kooperativa.io/api/v1/people/search" \
-H "Authorization: Bearer ik_live_..." \
-H "Content-Type: application/json" \
-d '{
"company_id": "74126343",
"past_company": "OpenAI",
"per_page": 1
}'
# "total" in the response is the count (47 on 2026-10-18).
# Swap the two companies for the reverse direction:
# company_id 11130470 (OpenAI) with past_company "Anthropic".Frequently asked questions
Which companies do Anthropic and OpenAI hire from?
In our sample, Google is the most common past employer at both: 18.3% of Anthropic profiles and 17.6% of OpenAI profiles list it. Meta follows at 10.3% and 16.4%. The count includes anyone who ever held a position there, interns too.
Do more people move from OpenAI to Anthropic or the other way?
In our sample 47 current Anthropic profiles list OpenAI as a past employer and 7 current OpenAI profiles list Anthropic, about 6.7 to 1. For people whose current position began in 2025 or later the figures are 30 and 6, or 5 to 1. The counts show neither direct moves nor reasons.
How was this data collected?
We counted profiles in a sample of our own professional profile data on 18 October 2026. The sample is the current profiles at the 11 companies, chosen by company and not drawn at random. Current company is matched by company ID and past employer by exact company name. It is not a survey, and we publish totals only.
Can I run the same analysis?
Yes, for any pair of companies. The people search endpoint takes a company ID and a past employer name and returns a total. Use a page size of 1 to read the total without paging.
Get started
Try Kooperativa
One API key. Person and company enrichment, structured search, and monitors under one flat license.