Meta Keeps 64% of AI Hires. Anthropic Keeps 80%. Source Menlo Park.

If you are a recruiter or founder chasing Anthropic engineers in 2026, you are pushing on a locked door. The lab most likely to hand you a warm candidate right now is the one writing the biggest checks, which is the same lab that cannot keep the people it just bought. That is the sourcing arbitrage nobody is exploiting, and SignalFire's State of Tech Talent data spells it out cleanly.

The number that flips the intuition

SignalFire's 2025 State of Talent Report tracked two-year retention across the frontier labs. Anthropic held 80% of its hires. DeepMind held 78%. OpenAI held 67%. Meta held 64%. That is a 16 point gap between the highest-paying and highest-retaining lab in the industry, and it moves in the opposite direction of what a naive comp model would predict.

64%

The directional flows are even more lopsided. Engineers at OpenAI are 8x more likely to leave for Anthropic than the reverse. The DeepMind-to-Anthropic ratio is 11:1. Growth-to-attrition ratios tell the same story: Anthropic is expanding engineering 2.68x faster than it loses talent, versus 2.18x for OpenAI, 2.07x for Meta, and 1.17x for Google. Anthropic's head of global GTM recruiting, Nick Lewis, put offer acceptance at 88% for tech roles and 95% for GTM. Heather Doshay told the WSJ that Anthropic gets named as the #1 dream company more often than any other lab.

That is a moat. Dario Amodei told the Big Technology Podcast he refused to match Meta's offers, saying his team would not "compromise our compensation principles." When your target's CEO is publicly refusing $18M packages on principle, your outbound to their staff is a low percentage play.

Meta is the opposite. Meta is where the money is going in, and the people are going out.

Why the leak got bigger in 2026

Zuckerberg spent 2025 assembling Meta Superintelligence Labs (MSL) on top of the existing FAIR organization. Meta paid $14.3B for a 49% non-voting stake in Scale AI in June 2025 and installed Alexandr Wang as Chief AI Officer. Five of Thinking Machines Lab's founding team joined Meta after Zuckerberg first offered roughly $1B to acquire the company outright. Denny Zhou left DeepMind's reasoning team for MSL. Dawn Song joined as VP of AI Research and brought her Virtue AI cofounders. TechCrunch clarified that the mythical "$100M signing bonuses" are actually complex RSU and tenure-linked packages, not cash on day one, but at least one investor confirmed seeing an $18M package turned down.

Then the org started leaking from the other end.

In October 2025, Meta cut roughly 600 roles from FAIR and AI infrastructure. In November 2025, Yann LeCun left after being asked to report to Wang, publicly calling Wang "young and inexperienced" and predicting more departures. LeCun raised $1B for AMI Labs in Paris. In January 2026, Meta laid off roughly 1,500 people from Reality Labs, or about 10% of the 15,000-person division.

That is the sourcing map. Not MSL, which is stapled to comp packages designed to survive nuclear winter. FAIR alumni, Reality Labs ICs, and LeCun-adjacent researchers who are watching the mission drift in real time.

FAIR is not MSL, and you should not source them the same way

This is the most important distinction, and most outbound campaigns miss it. The 600 October cuts hit the older FAIR and infra organizations, not the shiny new lab. That produced a well-defined cohort: experienced, often long-tenured, research-oriented, and often loyal to LeCun's academic-first vision rather than to Wang's product-first mandate. Their motivations are not the mercenary mercenary motivations of the MSL cohort. They will take less money for a real mission. They are the exact profile Anthropic already recruits from, which means you have a narrow window before the obvious buyer runs the same search you are running.

The mercenary MSL cohort, by contrast, is functionally unpoachable for 18 to 36 months because of vesting cliffs. Skip them.

Building the watchlist

There are four sourcing pools sitting inside the Meta AI story right now. They are not equal, and the outreach for each is different.

1. FAIR alumni, October 2025 cohort

Look for "Research Scientist" or "Research Engineer" titles at Meta with tenure predating the MSL announcement. Filter for authorship on FAIR papers between 2021 and 2024. Cross-reference with anyone LeCun co-authored with or advised. Paris, NYC, and Montreal offices matter as much as Menlo Park here, because FAIR was never a Bay-only org and LeCun's gravitational pull skewed toward those satellite offices.

The problem with running this search on LinkedIn is that "FAIR" is not a distinct company entity. It is a subteam. You cannot filter for it. This is the same methodological wrinkle that caused SignalFire's 2026 report to exclude Meta AI from the frontier-lab retention analysis: Meta AI does not exist as a separate LinkedIn org, so the 64% figure is arguably generous because post-MSL churn is not fully captured. If SignalFire's own analysts cannot cleanly slice it, your Boolean cannot either.

This is exactly the friction we built Refolk for. You describe the cohort in plain English ("Meta researchers who published on FAIR between 2021 and 2024, based in Paris or NYC, not on the MSL org chart") and get a ranked shortlist across GitHub, LinkedIn, and the open web without needing the target to have self-labeled correctly.

2. LeCun-adjacent, still inside

LeCun said publicly that "a lot of people who haven't yet left will leave." Treat that as forward guidance. AMI Labs' early hires in Paris are the leading indicator. Every name that shows up at AMI in the next two quarters is a signal about which of their former Meta colleagues will move next. Mine AMI's hires weekly. The people they hire are not your candidates. The people they co-authored with who are still at Meta are.

3. Reality Labs, January 2026 cohort

This is 1,500 people, mostly in Menlo Park and Burlingame, with concentrations in graphics, computer vision, systems, and hardware-adjacent ML. They are not frontier-lab researchers, but they are strong applied ML engineers who just watched a decade-long bet get partially unwound. They are more responsive to outbound than any current employee at any frontier lab, and they price 30 to 50% below MSL comp because they are already outside the walls.

Meta is where the money is going in and the people are going out. Source accordingly.

4. The Microsoft adjacent pool

On July 6, 2026, Microsoft's Chief People Officer Amy Coleman published the memo about 4,800 layoffs (2.1% of workforce) alongside the explicit "reskilling engineers for customer-facing and AI-focused positions" mandate. The new Microsoft Frontier Co. is redeploying 6,000 employees as forward-deployed engineers backed by a $2.5B investment. That framing creates a pool nobody is talking about: senior ICs who do not want to become customer-facing FDEs. They are quietly on the market. This is a product-eng persona, not a research persona, so it is a different outreach than the FAIR cohort, but it is genuinely underexploited right now.

The outreach that works on this cohort

Meta AI departures do not respond to comp-led messages. They just watched the biggest comp package in the industry fail to buy loyalty from Thinking Machines founders and fail to keep LeCun in the building. Leading with money reads as tone-deaf.

What works on the FAIR cohort:

• Reference a specific paper. Not the most cited one. The one from 2022 that got 40 citations and clearly mattered to them.
• Name the org chart problem out loud. "You are two levels under Wang now" is not subtle, but it is honest, and honesty is the scarce commodity in this market.
• Offer research autonomy in concrete terms. Publication rights. Conference budget. Compute allocation numbers. Not vibes.
• Skip the "we are building AGI" pitch. Everyone is. They know.

What works on the Reality Labs cohort:

• Lead with the applied problem. They are engineers, not researchers. Show the product.
• Acknowledge the loss. A decade of work on a headset roadmap that just got cut is not something to paper over.
• Move fast. This cohort clears in 60 to 90 days.

What works on the Microsoft adjacent cohort:

• Position against the FDE reskill, not against Microsoft the company.
• Senior IC track. Explicit. Written down.

Why the "unpoachable" tier matters for your search anyway

Anthropic's 2026 hires are a scoreboard, not a target list. John Jumper (Nobel laureate, from DeepMind, June 19, 2026), Jelani Nelson (Berkeley EECS chair), and Andrej Karpathy all landed there. You are not going to outbid or out-mission that offer. What you can do is use those names as beacons: everyone who has ever collaborated with Jumper, Nelson, or Karpathy is now one degree from the highest-status lab in the industry, and a subset of that graph is currently sitting inside Meta wondering what happened.

That graph is the search. Running it manually across Google Scholar, GitHub, and LinkedIn is a full-time job for a week per role. Running it in Refolk is a paragraph of English and a coffee break, which is the point: the arbitrage only works if you can act on it before the obvious buyers do.

The 90-day window

The 2026 SignalFire report excluded Meta AI as an entity, which means we do not have a clean updated retention number. But the leading indicators (LeCun's exit, the FAIR cuts, the Reality Labs cuts, the Thinking Machines raid, the LeCun prediction, and the Wang reporting-line problem) all point in the same direction. The 64% figure is a floor, not a ceiling.

The founders and recruiters who move on this in Q3 2026 will fill their AI research and applied ML roles with candidates who two years ago were untouchable. The ones who wait until the next SignalFire report confirms it in print will be reading about their competitors' hires.

Source Menlo Park. Source Paris. Source NYC. And do not spend another week trying to poach Anthropic.

FAQ

Is Meta's 64% retention actually worse than the number suggests?

Probably yes. SignalFire's 2026 State of Tech Talent Report explicitly excluded Meta AI from the frontier-lab retention analysis because Meta AI is not a distinct LinkedIn entity, only a subteam label. That means churn inside MSL, FAIR, and Reality Labs' AI groups may not be fully reflected in the 64% figure, which was based on 2025 data collected before the October 2025 FAIR cuts, LeCun's November 2025 departure, and the January 2026 Reality Labs layoffs. Assume the real 2026 number is lower.

Why can't I just poach Anthropic engineers directly?

Because the math is against you. Anthropic retains 80% of two-year hires, has 88% offer acceptance for tech roles and 95% for GTM, and Dario Amodei has publicly refused to match Meta's offers on principle. Engineers there are named as the #1 dream company more often than any other lab. You can occasionally win one, but as a systematic sourcing strategy it produces very low yield per hour. Meta's outbound response rates on the FAIR and Reality Labs cohorts are structurally higher because those cohorts are already emotionally outside the building.

How do I find FAIR alumni if "FAIR" is not a real LinkedIn org?

You cannot do it with a clean Boolean, which is exactly why so few recruiters have run this search well. You need to combine paper authorship (FAIR publications 2021 to 2024), tenure filters (joined Meta before 2023), LeCun's co-authorship graph, and geo filters (Paris, NYC, Montreal, Menlo Park). Refolk handles this natively because you can describe the cohort in plain English instead of trying to reverse-engineer it into a search string that LinkedIn does not support.

What about the Microsoft Frontier Co. reskilling pool?

Amy Coleman's July 6, 2026 memo made the reskill mandate explicit, and Microsoft is investing $2.5B to redeploy 6,000 employees as forward-deployed engineers. Not everyone wants that role. Senior ICs who prefer platform, infra, or research work over customer-facing implementation work are quietly evaluating options right now. It is a product-engineering persona, not a research persona, so treat it as an adjacent pool with its own outreach playbook, not as an extension of the Meta cohort.