Andrew Tulloch is leaving Meta for Anthropic, and the timing is what makes the move sting. Business Insider reported on September 10, 2026 that Tulloch will exit Meta’s Superintelligence Labs after less than a year on the job to join Anthropic’s inference and performance team. An Anthropic spokesperson told the outlet he will start “next week,” putting his official arrival at the company in mid-September 2026. The AI research and engineering outlet shattered.io and industry tracker A47 News confirmed the same departure within hours of each other.
On paper, this is a single research hire changing employers. In practice, it is the latest data point in a talent war that has reshaped how frontier AI labs compete, spend, and retain the small pool of engineers capable of training and serving models at the scale of GPT, Gemini, and Claude. Tulloch’s exit follows Meta’s costliest recruiting campaign in company history, one built specifically to prevent moves like this one.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Business Insider and Other Outlets Are Reporting
The core facts, as confirmed across multiple outlets, are narrow but significant. Andrew Tulloch is departing Meta’s Superintelligence Labs (MSL) to join Anthropic, where he will work specifically on training and inference, according to Business Insider, shattered.io, and A47 News. Business Insider’s report states plainly that he’ll join “the startup’s inference and performance team, where he’ll aim to optimize the training of Anthropic’s AI models and how they respond to users.” That framing matters: it is not a generic research placement, it is a hire aimed squarely at the operational core of how Anthropic’s Claude models get built and run in production.
Neither Meta nor Anthropic has issued a lengthy public statement explaining the move. The only on-record comment identified in reporting so far is the Anthropic spokesperson’s confirmation to Business Insider of Tulloch’s start date. Meta has not been quoted directly on the departure in the coverage reviewed. That silence is itself notable given how vocally Meta has talked about its AI hiring in the past 14 months.
Shattered.io frames the move in blunter terms, describing it as a signal “that money alone may not be enough to keep top researchers loyal to Meta’s Superintelligence Labs.” That framing is worth sitting with, because Meta’s approach to MSL was built almost entirely on the premise that money would be enough.
Who Is Andrew Tulloch?
Tulloch is not a rank-and-file engineer. Before Meta, he co-founded Thinking Machines Lab alongside Mira Murati, OpenAI’s former chief technology officer, according to reporting cited by DeepLearning.ai’s The Batch newsletter. That pedigree put him firmly inside the small circle of researchers who have shaped frontier model development at more than one lab.
The same DeepLearning.ai piece, citing reporting from The Wall Street Journal, notes that Tulloch initially turned down a Meta compensation package that included bonuses reportedly worth $1.5 billion before ultimately joining the company during its mid-2025 hiring spree. The Wall Street Journal returned to that same $1.5 billion figure in September 2026, re-reporting it as his Anthropic move became public and underscoring just how large the package was that he ultimately walked away from. He joined Meta’s Superintelligence Labs less than a year before this week’s Anthropic move, meaning his entire tenure at Meta lasted under 12 months from hire to departure.
Shattered.io’s account adds that the exit comes “after a reported $1B+ pay package,” a figure the outlet frames as unconfirmed by either company. No source reviewed for this story publishes an official, company-confirmed compensation number for Tulloch at either Meta or Anthropic. Every dollar figure connected to his moves should be read as reported, not verified.
Inside Meta Superintelligence Labs: The Team Tulloch Is Leaving
Meta Superintelligence Labs was formed on June 30, 2025, consolidating the company’s AI research, foundation model, and AI product teams under a single division built around what Meta has called “personal superintelligence.” The formation came with an unusually public recruiting push. Mark Zuckerberg personally named roughly 11 new hires poached from rival labs, including Trapit Bansal, Shuchao Bi, Huiwen Chang, Ji Lin, Hongyu Ren, and Jiahui Yu from OpenAI; Jack Rae and Pei Sun from Google DeepMind; Joel Pobar from Anthropic; and Johan Schalkwyk from Sesame AI.
The lab’s growth was not linear. In October 2025, Meta cut roughly 600 roles from MSL, leaving the division at just under 3,000 employees. Despite the reduction, MSL kept shipping: on August 5, 2026, the lab released Muse Code, a coding agent in public beta powered by its Muse Spark 1.2 model. That product cadence is part of why Tulloch’s exit lands as more than routine turnover, since it suggests MSL is losing senior research talent even while trying to maintain a competitive release schedule.
Meta’s own account of that hiring push, reported separately by Business Insider in mid-2025, described the addition of Shengjia Zhao, a co-creator of ChatGPT and former OpenAI lead scientist, as chief scientist of MSL, framing the move as an escalation of the talent war with OpenAI directly. Tulloch is not the only high-profile name to leave Meta’s AI organization in 2026. In February 2026, OpenAI re-hired two researchers, Ruoming Pang and Pengchuan Zhang, from Meta in the same week, according to reporting on the broader AI talent war. Counting Tulloch, that makes at least three named departures from Meta’s frontier AI ranks in 2026 alone, on top of what one Axios-cited account describes only in aggregate as “several” recruits who “quickly decamp” after joining MSL.
What Tulloch Will Actually Do at Anthropic
The reporting is specific on this point: Tulloch is joining Anthropic’s inference and performance team in September 2026, after less than a year at Meta, according to Business Insider. Business Insider describes his mandate as optimizing “the training of Anthropic’s AI models and how they respond to users.” Shattered.io adds that the team’s scope covers both training efficiency and tuning how models behave once they are answering real production traffic, work the outlet describes as sitting “at the center of Anthropic’s enterprise pitch.”
That distinction between training and inference is not cosmetic. Training determines how capable a model is when it ships. Inference determines how fast, how cheap, and how reliably that model runs once millions of users and enterprise customers are hitting it every day. As Anthropic pushes further into enterprise deployments, cutting inference latency and cost per token becomes as commercially important as raw benchmark performance. A researcher who can move the needle on both sides of that equation is a strategic hire, not a headcount add.
Anthropic’s recent product cadence backs up why this role matters now. The company has been expanding its enterprise safeguards and tooling, a push covered in Anthropic’s rollout of new Claude enterprise safeguards to more than 100 firms, and its models continue to be benchmarked directly against OpenAI’s lineup, as seen in the head-to-head look at Claude Opus 4.8 against GPT-5.6 Sol on coding tasks. That rivalry sharpened further on September 22, 2026, when Anthropic shipped Claude Opus 5.5, which TechCrunch reported now runs at $4 per million input tokens and $20 per million output tokens, and which 9to5Mac reported outscored GPT-5.6 Sol by 11 points on CursorBench. Faster, cheaper inference is the kind of unglamorous engineering work that keeps those enterprise commitments viable at scale, and it is precisely the lever Opus 5.5’s pricing drop shows Anthropic pulling.
The Compensation Numbers Behind the Talent War
Meta’s willingness to pay eye-watering sums to build MSL is well documented. According to reporting from Calcalistech summarizing Wired’s coverage, Meta offered compensation deals worth up to $300 million over four years to pull top researchers away from OpenAI, Anthropic, and Google DeepMind. Separate research from AgentMarketCap on 2026 AI compensation puts Anthropic’s median research scientist package at roughly $746,000, with top packages reaching about $1.05 million, figures that sit far below Meta’s most aggressive offers but are still well above typical tech industry pay.
Those numbers help explain why Tulloch’s move reads as counterintuitive on its face. If Meta’s packages genuinely reach into the hundreds of millions for its most sought-after hires, a move to a company reportedly paying under $1.1 million at the top of its research scientist band looks like it runs against pure financial logic, unless Anthropic is compensating him well outside that median band, or unless factors other than raw pay drove the decision. Neither company has disclosed Tulloch’s actual new compensation, so any specific number attached to his Anthropic deal beyond shattered.io’s unconfirmed “$1B+” reference should be treated as speculation.
| Lab | Reported Compensation Range (2026) | Source |
|---|---|---|
| Meta Superintelligence Labs | Up to $300 million over four years for top targeted hires | Calcalistech, citing Wired |
| Anthropic | ~$746,000 median research scientist package; up to ~$1.05 million at the top | AgentMarketCap 2026 compensation analysis |
| Meta (Tulloch’s original offer) | Reported bonuses worth $1.5 billion, initially declined | DeepLearning.ai’s The Batch, citing The Wall Street Journal |
| Industry-wide top packages (various labs) | Reaching roughly $5 million in select cases | AgentMarketCap 2026 compensation analysis |
A Timeline of Meta Superintelligence Labs’ Turbulent First Year
Meta’s superintelligence push has moved fast in both directions, adding headcount aggressively and then losing some of it just as quickly. The table below lays out the key milestones reported since MSL’s formation through Tulloch’s departure this week.
| Date | Event |
|---|---|
| June 30, 2025 | Meta Superintelligence Labs formally created, consolidating AI research and product teams |
| July 2025 | Meta launches a hiring spree offering hundreds of millions of dollars to researchers from OpenAI, Google DeepMind, and Anthropic; Zuckerberg names roughly 11 recruits, including Andrew Tulloch |
| July 2025 | Meta hires Shengjia Zhao, a co-creator of ChatGPT and former OpenAI lead scientist, as chief scientist of MSL |
| October 2025 | Meta cuts roughly 600 roles from MSL, leaving the division at just under 3,000 employees |
| February 2026 | OpenAI re-hires Ruoming Pang and Pengchuan Zhang, both previously at Meta, in the same week |
| August 5, 2026 | MSL releases Muse Code, a coding agent in public beta powered by Muse Spark 1.2 |
| September 10, 2026 | Business Insider reports Andrew Tulloch is leaving Meta for Anthropic’s inference and performance team after less than a year at MSL |
Why Money Alone Isn’t Holding MSL Together
The pattern emerging across 2025 and 2026 is that Meta can win a hire but struggle to keep one. The company’s strategy assumed that sufficiently large compensation offers would neutralize any cultural or mission-based pull from rival labs. Tulloch’s exit, alongside the Pang and Zhang departures earlier in the year, suggests that assumption has limits. Shattered.io’s read on the move, that it signals money alone may not secure loyalty at MSL, lines up with a broader pattern researchers and industry observers have pointed to: fast-moving reorganizations, the October 2025 layoffs, and public scrutiny of MSL’s output can all erode retention even when the paycheck is enormous.
Anthropic, by contrast, has built a reputation around a narrower, mission-driven research culture and has continued attracting senior talent even while paying less than Meta’s most extreme offers. That contrast is part of why some of Meta’s competitors in the talent war, including Google-backed ventures, have found success recruiting Meta alumni directly, a dynamic also visible in Crux AI’s recent hire of Meta researcher Alan Duong, backed jointly by Google and Blackstone.
Anthropic’s Growing Appetite for Infrastructure Talent
Anthropic’s decision to prioritize an inference and performance hire fits a broader company posture. As Claude models scale to more enterprise customers, the cost of running inference becomes a direct line item that affects margins and pricing, and Anthropic’s own pricing history shows how fast that line item is moving: CNBC reported that Claude Opus 5 launched on July 24, 2026 at $5 per million input tokens and $25 per million output tokens, a rate Anthropic then undercut just two months later with Opus 5.5’s $4/$20 per-million-token pricing. Hiring a researcher with Tulloch’s background at Thinking Machines Lab and Meta signals that Anthropic is treating serving efficiency as a competitive differentiator, not just a backend cost center.
That focus also lands at a moment when Anthropic faces continued scrutiny over its growth trajectory and governance. Commentary around the company’s path toward a public listing, discussed in coverage of CNBC’s Jim Cramer hyping an Anthropic IPO shortly before one of its own researchers cited a 10% AI risk estimate, shows the company operating under a spotlight where both its technical execution and its internal safety culture are being watched closely. A high-profile infrastructure hire like Tulloch adds technical credibility to that narrative, even as questions about compensation transparency remain unanswered.
How This Fits the Broader AI Funding and Valuation Picture
Talent moves of this size do not happen in a vacuum. They track alongside massive capital flows into frontier AI labs. Rival lab Mistral AI, for instance, saw its valuation climb past €21 billion following a €3 billion investment tied to Samsung, detailed in Tech Insider’s coverage of Mistral’s Samsung-backed funding round. That kind of capital injection is precisely what allows labs across the industry, not just Meta, to offer packages large enough to poach senior researchers from each other. None of the sources reviewed for this story publish a confirmed 2026 valuation figure specifically for Anthropic, so any number circulating for the company should be treated cautiously until officially disclosed.
What is clear is that the capital available to frontier labs keeps climbing, and that capital is increasingly being spent on people rather than just compute. Every dollar committed to a nine-figure researcher package is a dollar not spent on training runs or data center capacity, a tradeoff that is starting to show up in how labs prioritize their budgets heading into 2027.
Historical Context: The AI Talent War Since 2024
The current wave of poaching did not start with MSL. Since 2024, OpenAI, Google, Anthropic, and Meta have traded senior researchers in a cycle that has only accelerated. Google’s own talent war coverage, cited in Business Insider’s reporting, has tracked a string of high-profile moves across Anthropic, OpenAI, and Google DeepMind well before Meta’s MSL formation in mid-2025. What changed with MSL was scale: Meta’s willingness to offer sums as large as $300 million over four years, according to Calcalistech’s reporting, effectively reset the ceiling for what a single researcher hire could cost.
That reset created a feedback loop. Once Meta proved willing to pay nine figures for a single hire, rival labs had to either match those numbers for their most critical staff or accept the risk of losing them, a dynamic AgentMarketCap’s analysis frames as pushing some individual packages as high as roughly $5 million even outside Meta’s most extreme offers. Tulloch’s original decision to decline Meta’s reported $1.5 billion package, only to accept a later offer and join anyway, illustrates just how unstable even top-of-market compensation had become as a retention tool by mid-2025.
Competitive Landscape: Meta, Anthropic, OpenAI, and Google DeepMind
Each of the four major labs is fighting this war differently. Meta has leaned hardest on raw compensation, backed by its scale and ad-revenue balance sheet, but MSL’s October 2025 layoffs and its string of 2026 departures suggest that money is being spent faster than it is buying stability. OpenAI has shown it can win back its own former researchers, as seen with the February 2026 re-hires of Pang and Zhang. Google DeepMind continues to lose and gain talent as part of the same churn, with Joel Pobar’s move from Anthropic to Meta in 2025 as one example running in the opposite direction of Tulloch’s move now.
Anthropic’s position looks different: it is not outspending Meta, based on the compensation figures reviewed, yet it keeps landing senior hires who could likely command Meta-scale packages elsewhere. That suggests Anthropic’s pitch rests on factors beyond pay, whether that is research autonomy, the specific technical problems on offer, or the company’s public positioning on AI safety. Whatever the mix, it is proving durable enough to pull a co-founder of Thinking Machines Lab away from one of the best-funded AI divisions in the industry.
Market and Industry Reaction
No source reviewed for this story reports a measurable stock price reaction tied specifically to Tulloch’s departure, and none of the coverage cites an analyst note addressing the move directly. That is not unusual for a single researcher exit, even a high-profile one; these moves tend to matter more for internal morale and recruiting perception than for public market pricing in the short term. Where the reaction is visible is in trade press and industry commentary, with shattered.io’s framing of the move as a “signal” being the closest thing to an analytical verdict currently in circulation.
What Comes Next: Five Predictions
- Meta will likely respond with additional public hiring announcements in the coming weeks to offset the narrative of losing talent, following the same pattern it used after the Pang and Zhang exits in February 2026.
- Anthropic’s inference and performance team will probably become a more visible unit in the company’s public communications, especially if serving costs and latency become larger talking points in its enterprise sales pitch.
- Expect further scrutiny of MSL’s retention numbers heading into 2027, particularly if additional named researchers leave following Tulloch, Pang, and Zhang.
- Compensation figures across the industry are likely to keep climbing rather than stabilizing, since AgentMarketCap’s data already shows some individual packages approaching $5 million even outside Meta’s headline offers.
- Reporting on Anthropic’s funding and valuation trajectory should sharpen in the coming months as the company continues to add senior technical staff, a pattern that historically precedes fresh funding disclosures at other frontier labs, as seen with Mistral AI’s own valuation jump earlier in 2026.
Why This Matters Beyond One Researcher
It is tempting to read a single departure as noise. But the pattern across 2025 and 2026, Meta’s aggressive hiring, its October layoffs, the Pang and Zhang exits, and now Tulloch’s move, points to something structural: even the largest compensation packages in tech history are not sufficient on their own to guarantee retention in frontier AI research. That has implications for how every lab in this race, not just Meta, thinks about what it actually takes to keep its best people once they are in the building.
For Anthropic, the immediate implication is more concrete: a researcher with hands-on experience training and serving frontier models at two other top labs is now working on the exact bottleneck, inference cost and reliability, that determines whether its enterprise ambitions scale profitably. Whether that translates into measurable product improvements will likely show up first in Claude’s latency and cost-per-token metrics rather than in any public statement from the company.
Frequently Asked Questions
Who is Andrew Tulloch?
Andrew Tulloch is an AI researcher who co-founded Thinking Machines Lab with Mira Murati, OpenAI’s former chief technology officer, before joining Meta’s Superintelligence Labs in 2025 and now moving to Anthropic in September 2026, according to Business Insider and DeepLearning.ai’s The Batch.
Why is Andrew Tulloch leaving Meta for Anthropic?
Neither company has published an official explanation. Shattered.io frames the move as evidence that compensation alone has not been enough to retain top researchers at Meta’s Superintelligence Labs, though no on-record statement from either company confirms a specific motive.
What will Andrew Tulloch do at Anthropic?
According to Business Insider, he is joining Anthropic’s inference and performance team, working on optimizing model training and how Anthropic’s AI models respond to users in production.
How much was Andrew Tulloch reportedly paid at Meta?
Reporting cited by DeepLearning.ai’s The Batch, sourced to The Wall Street Journal, states Tulloch initially turned down a Meta package including bonuses reportedly worth $1.5 billion before later joining the company. Shattered.io separately references an unconfirmed “$1B+” pay package tied to his time at Meta. Neither figure has been officially confirmed by Meta.
What is Meta Superintelligence Labs?
Meta Superintelligence Labs (MSL) is the division Meta formed on June 30, 2025, consolidating its AI research, foundation model, and AI product teams to pursue what the company calls personal superintelligence. It cut roughly 600 roles in October 2025 and has since lost several senior researchers, including Ruoming Pang, Pengchuan Zhang, and now Andrew Tulloch.
Has Anthropic confirmed Andrew Tulloch’s compensation?
No. An Anthropic spokesperson confirmed to Business Insider only that Tulloch will start “next week” and join the inference and performance team. No compensation figure for his new role has been officially disclosed.
Is Andrew Tulloch the only researcher to leave Meta’s AI division in 2026?
No. In February 2026, OpenAI re-hired Ruoming Pang and Pengchuan Zhang, both previously at Meta, in the same week. Other reports reference additional unnamed departures from MSL without specifying names.
Does this affect Meta’s AI product roadmap?
There is no confirmed reporting linking Tulloch’s departure to a specific delay or change in Meta’s AI product plans, including its Muse Code and Muse Spark model releases. Any impact on Meta’s roadmap remains speculative at this stage.
