What a software engineer at OpenAI or Anthropic earns in 2026 is not a salary. It is a stake in the two most valuable private companies on earth.

The median software engineer at OpenAI reports total compensation near $870,000 a year, and the number climbs past $1.37 million at the senior individual-contributor level - Levels.fyi. At Anthropic, a Staff Software Engineer reports roughly $1.25 million, of which about $843,000 is equity rather than cash - Levels.fyi. These are not one-off offers to celebrity researchers. They have become the going rate for engineers who, three years ago, would have taken a $250,000 package at a big-tech company and called it a career win.

But the headline number hides more than it reveals. Most of that compensation is illiquid private equity, not money in a bank account. The base salary, the part you can actually spend, sits far lower: government visa filings show OpenAI paying a median base of $310,000 to its members of technical staff, and Anthropic around $300,000 - h1bdata.info. The gap between $300,000 in cash and $1 million on paper is the entire story of frontier-lab pay, and understanding it is the difference between an offer that changes your life and one that merely looks like it will.

This guide breaks down exactly what the two labs pay in 2026, how the base-versus-equity split works, why a tender offer is the only moment the paper becomes real, and how the 2026 talent war (nine-figure offers from Meta, a rejected $1.5 billion package, a leaked "someone has broken into our home" memo) reset the ceiling for everyone. It is written for the engineer wondering whether one of these jobs is worth chasing, and for the recruiter or founder trying to compete for the same people without a trillion-dollar balance sheet.

Contents

1. The headline: what a frontier-lab engineer earns in 2026
2. How to read the numbers: base, equity, and the paper-versus-cash problem
3. Inside OpenAI's pay machine: PPUs, the $500B tender, and $1.5M in stock
4. Inside Anthropic's pay machine: RSUs, the $350B tender, and the retention edge
5. The talent war that reset the ceiling: Meta, $100M, and a $1.5B "no"
6. How OpenAI and Anthropic pay stacks up against everyone else
7. Why the numbers are this high: scarcity, winner-take-all, and the AI premium
8. Can you actually get one of these jobs? The bar, the interview, the visa
9. What it means if you are hiring, or being hired
10. The future: liquidity events, cooling signals, and agentic hiring

1. The headline: what a frontier-lab engineer earns in 2026

Total compensation at OpenAI and Anthropic scales from roughly $250,000 at entry to well over $1.2 million at the top of the engineering ladder, and researchers sit a tier above that again. That single fact reorders how ambitious engineers think about their careers, because it means a strong mid-career move to a frontier lab can more than triple a Big Tech package. The rest of this guide is mostly about the asterisks attached to those numbers, but the numbers themselves are the reason anyone reads on.

At OpenAI, self-reported data puts an entry-level (L2) software engineer around $253,000, an L4 around $653,000, an L5 around $1.16 million, and L6 packages averaging close to $1.19 million with the single highest reported figure near $1.585 million - Levels.fyi. Anthropic's ladder runs on a parallel track: an entry Software Engineer near $367,000, Senior near $591,000, Lead near $780,000, and Staff near $1.25 million - Levels.fyi. The two ladders are not labelled identically, which matters when you compare them, but the shape is the same: pay roughly doubles with each big step up in seniority, and equity does almost all of the doubling.

The chart below places the two ladders side by side using approximate career stages. Read it with one caveat in mind: OpenAI's L5 is an unusually senior individual-contributor rung, which is why OpenAI spikes hard at the "Senior" stage while Anthropic keeps climbing to a higher "Staff" peak. The point is not that one lab beats the other at every rung, but that both operate in a band no traditional employer can match.

What the chart cannot show is the tier that sits above ordinary engineering entirely. Research scientists and research engineers at OpenAI report median total comp closer to $1.25 million, one rung higher than the equivalent software-engineering level, because the people who can push a frontier model forward are scarcer than the people who can ship the product around it - Levels.fyi. At the very top, superstar researchers are a different market again, priced in tens of millions rather than hundreds of thousands, a dynamic we return to in chapter five. For a normal, excellent engineer, though, the ladders above are the real reference points, and they are already extraordinary.

It is worth pausing on how recent all of this is. As late as 2023, a strong senior engineer choosing between big-tech offers was optimizing inside a $250,000 to $450,000 band, and the frontier labs were small enough that most engineers had never seriously considered them. The mid-rungs that now anchor negotiations barely existed at these numbers: OpenAI's L3, one step above entry, already reports around $333,000, more than many principal engineers earn elsewhere - Levels.fyi. The compression of roughly a decade of pay inflation into about two years is why so many engineers feel disoriented reading these figures, and why benchmarking against your own past offers is now actively misleading.

That research premium is worth a note for anyone early in their career. Self-reported data puts the base salary of a senior OpenAI research scientist near $396,000, before the equity that lifts the total past $1.4 million, and the pattern repeats at Anthropic - Levels.fyi. The premium exists because the skill is rarer: many excellent engineers can build the product around a model, but very few can meaningfully improve the model itself. If you are choosing where to specialize, that gap is a signal worth heeding, because the research and research-engineering tracks are where the scarcity, and therefore the pay, concentrates most, and where the ceiling stays highest over a career.

The practical takeaway for anyone benchmarking an offer is to anchor on the level, not the company. A "Senior" title means something very different across these two ladders, and the dollar difference between adjacent rungs (often $400,000 or more) dwarfs the difference between the two labs at the same rung. If you are negotiating, the highest-leverage question is not "OpenAI or Anthropic" but "which rung of which ladder does this offer actually put me on," because that single classification decides most of the money.

2. How to read the numbers: base, equity, and the paper-versus-cash problem

Every frontier-lab number you have read so far is base salary plus illiquid equity, with almost no cash bonus, and the two halves behave nothing alike. Base salary is guaranteed, taxed as ordinary income, and lands in your account every two weeks. Equity is a bet on a private company's future valuation that you cannot sell on the open market, may not be able to sell at all for years, and could be worth far more or far less than the number on your offer letter. Confusing the two is the most expensive mistake a candidate can make.

The most reliable window into the cash half is government data. When a lab sponsors a work visa, it must file the offered base salary with the U.S. Department of Labor, and those filings are public. They show OpenAI members of technical staff at a median base near $310,000, with 97% of filings above $200,000, and Anthropic technical staff clustering around a $300,000 median base, ranging up past $425,000 for senior roles - h1bdata.info. Business Insider's analysis of Anthropic's 2026 filings found individual technical-staff base salaries stretching as high as $1.38 million for the most senior roles, and research-operations and technical-sales roles reaching $500,000, all before a dollar of equity - Business Insider. Because these are legally filed figures, they are the closest thing to an audited floor on frontier-lab pay.

The sheer volume of those filings is itself a data point. OpenAI's operating entity filed roughly 129 H-1B petitions in a single recent fiscal year, of which about 125 were approved, and it onboarded dozens of overseas hires in a single quarter - MyVisaJobs. Anthropic's filings are similarly heavy and concentrated in California. For a candidate, the useful implication is that base salary is both high and unusually transparent at these labs, because the immigration system forces disclosure the equity component never receives. If you want to know the floor, read the filings; if you want to know the ceiling, you have to understand the equity, and the equity is where the risk lives.

Combining the two data sources gives a realistic estimate for any specific offer. Take the visa-filed base as a firm floor, add the equity portion from self-reported totals at the same level, then apply the illiquidity discount from the previous section, and you arrive at a defensible expected value rather than a headline. In practice that means an offer advertised at, say, $900,000 total might be understood as roughly $350,000 in guaranteed base plus $550,000 in equity that is worth its face value only if the valuation holds and a sale actually materializes. Doing this arithmetic before you accept is the single most valuable habit a candidate can build, because it converts an intimidating number into a set of assumptions you can interrogate one at a time: how firm is the base, how likely is the valuation to hold, and how soon can the equity be sold. An offer you can decompose is an offer you can compare; a headline you cannot is just marketing.

The equity half is where the eye-popping totals come from, and where the risk lives. At Anthropic, roughly 70% of a package at senior levels is equity, granted as restricted stock that vests over four years with a one-year cliff and then monthly - Levels.fyi. The diagram below traces how a single package splits into a part you can spend and a part you can only hope to spend later.

The crucial node in that diagram is the second-to-last one. Until a tender offer (a company-organized sale where investors buy employee shares) or an IPO happens, vested equity is paper. You cannot pay a mortgage with it, and its value swings with every funding round. This is why the single most important due-diligence question a candidate can ask is not "how much equity" but "when and how can I sell it," a question we answer for each lab in the next two chapters.

The tax and timing mechanics compound the paper-versus-cash gap. Modern grants at both labs use double-trigger vesting, meaning shares are not fully yours until both a time-based schedule and a liquidity event occur, and even vested private shares can sit behind lockups and transfer restrictions. That structure protects the company but shifts risk onto the employee, who owes ordinary-income tax on cash base every year while the equity, the majority of the package, stays frozen and unsellable. A candidate who mentally spends the full seven-figure total is budgeting against money that may not become real for years, if ever. The discipline that separates a good decision from a dazzled one is refusing to count equity as cash until there is a concrete, credible path to selling it.

A final reason to distrust any single headline figure is dispersion. A reported "median" flattens a wide spread of real outcomes that depend on hire date, negotiation, level, and how the equity has appreciated. The distribution below, from Levels.fyi's annual report, shows how far senior software-engineer pay spreads even inside the biggest and most transparent tech employers, and frontier labs are wider still because so much of the package is a moving valuation.

The dispersion in that image is the reason two engineers with identical titles at the same lab can be $500,000 apart. When you read "the median is $870,000," picture that spread behind it, and treat the median as the middle of a wide cloud rather than the price of the job. With that framing in place, the individual pay machines at each lab start to make sense.

3. Inside OpenAI's pay machine: PPUs, the $500B tender, and $1.5M in stock

OpenAI ran the most unusual equity program in tech, paid the highest average stock compensation of any startup in history, and then rebuilt the whole thing in October 2025. Understanding that arc explains both why OpenAI packages ballooned and why the company had to rewire its own comp structure under competitive pressure. It starts with a piece of financial engineering called the PPU.

For most of its history, OpenAI did not grant stock, options, or RSUs. It granted Profit Participation Units, a synthetic instrument that entitled the holder to a slice of the company's future profits rather than ownership of shares - Levels.fyi. PPUs vested over four years at 25% per year, carried a historical two-year lockup, and, in older grants, were capped so a unit could appreciate only about tenfold before the upside stopped. A typical senior offer looked like a $300,000 base plus a large annual PPU grant, sometimes $500,000, sometimes $2 million of notional value, layered year over year. The structure existed because OpenAI began life inside a capped-profit nonprofit, and it produced a generation of paper millionaires who owned no actual stock.

That model reached a breaking point in 2025. According to figures OpenAI presented to investors, the company's stock-based compensation averaged about $1.5 million per employee, described as the highest of any tech startup in history, roughly 34 times the typical pre-IPO norm, and consuming an astonishing 46% of revenue - Fortune. Paying nearly half of every dollar of revenue in equity is not sustainable indefinitely, and it collided with a second problem: employees holding paper they could not sell while rivals waved cash.

Paying nearly half of revenue in equity is not a quirk; it is a structural strain that shapes behavior. Stock compensation at that scale dilutes existing investors, complicates the path to profitability, and creates enormous pressure to give employees an exit before resentment sets in. It is a large part of why OpenAI restructured at all: the capped-profit PPU model had become both a recruiting liability, because rivals could offer cleaner and more familiar equity, and a financial one, because the paper obligations kept growing faster than the mechanisms to satisfy them. The 2025 rebuild was as much a release valve as a reorganization, designed to make the equity legible enough that employees would trust it and stay.

The scale is easier to grasp in comparison. Where a mature software company like Alphabet directs roughly 15% of revenue to stock compensation, and Meta under 6%, OpenAI's reported 46% is in a different universe, several times the roughly 6% average across large pre-IPO tech firms - Fortune. The only companies in the same conversation are other frontier labs. That intensity is a direct measure of how badly OpenAI wanted to hold its people during the raids of 2025: it was willing to hand out an unprecedented share of its own value to keep the team intact, betting the talent was worth more than the dilution. For an employee, it also explains why the tender mattered so much, because a company paying that much in paper has to eventually make the paper real or watch morale curdle.

OpenAI's answer came in two moves. First, in its October 2025 restructuring into a public benefit corporation, it converted PPUs into conventional equity and removed the old profit cap, so holders could finally share in uncapped appreciation and new hires could receive familiar double-trigger RSUs - Levels.fyi. Second, it gave staff a way to cash out. In early October 2025 OpenAI closed a $6.6 billion employee share sale at a $500 billion valuation, letting more than 600 current and former employees each sell up to $30 million, with roughly 75 people hitting that cap - CNBC. For an employee, that tender was the moment years of paper turned into a wire transfer.

The mechanics of that sale reveal how deliberately OpenAI engineered liquidity. The company raised the per-person cap from an earlier $10 million to $30 million in response to investor demand, and while roughly 75 people sold the maximum, the wider group averaged closer to $11 million each - CNBC. It was not the first such event, either: a roughly $1.5 billion SoftBank-led tender in late 2024 had already let staff cash out at earlier valuations of $65 billion and then $91.5 billion. Each tender is effectively a scheduled bonus round whose size depends entirely on the latest valuation, which is why OpenAI employees track funding rumors the way public-company staff track an earnings date.

The reason those grants keep getting more valuable is the valuation curve underneath them, and it is nearly vertical. The line chart below tracks how OpenAI's private valuation moved against Anthropic's over roughly a year, and it is the real engine behind every equity figure in this guide: the same grant issued twelve months apart can be worth wildly different amounts.

By March 2026 OpenAI had raised a fresh $122 billion round at an $852 billion valuation, with Amazon, Nvidia, and SoftBank among the buyers, roughly quintupling the paper value of a 2024-era grant in under eighteen months - CNBC. That is why OpenAI could hand out retention grants reported around $1.5 million to roughly a thousand employees, and up to $5 million for its most sought-after researchers, without technically raising anyone's "salary" - Fortune. The lesson for a candidate is that at OpenAI the valuation trajectory, not the offer letter, determines what the job is worth, and the tender calendar determines when you get to keep it.

4. Inside Anthropic's pay machine: RSUs, the $350B tender, and the retention edge

Anthropic pays nearly as much as OpenAI, structures its equity more conventionally, and keeps its people better than any rival, and it does all three while publicly refusing to fight bidding wars. That combination makes it the most interesting counterexample in the market: proof that a lab can hold elite talent without matching the highest number on the table. Its pay machine is worth studying precisely because it looks so different from OpenAI's.

On raw compensation, the two are close. Anthropic's software engineers report totals from $367,000 at entry to $1.25 million at Staff, with senior and lead engineers between those poles, and roughly 70% of the package delivered as restricted stock on a standard four-year vest - Levels.fyi. Unlike OpenAI's former PPUs, this is ordinary private-company equity, which makes it easier to understand and, arguably, easier to trust. Geography stretches the range further: as Anthropic built out a London office with room for hundreds of staff, it advertised research-engineering roles with base salaries up to £630,000 and a senior trust-and-safety engineering role at £265,000 to £370,000 - City AM. The interview below, recorded in mid-2025, gives the Anthropic side of the story in CEO Dario Amodei's own words, including his view of the spending war around him.

Anthropic's liquidity story mirrors OpenAI's but with a revealing twist. In early 2026 the company organized an employee tender offer at a $350 billion valuation, aiming to let staff sell $5 billion to $6 billion of stock - Bloomberg. When the sale completed that spring, it came up short of the money investors had lined up, because many employees declined to sell, betting their shares would be worth more in a future IPO - Bloomberg. That is a remarkable signal. Given a rare chance to convert paper to cash, a large share of Anthropic's staff chose to keep the paper, which tells you more about internal conviction than any survey could. The valuation supporting that conviction had climbed from $61.5 billion in March 2025 to $183 billion that September and past $350 billion by early 2026 - Forbes.

Underneath those valuations sits an economic model that makes Anthropic's discipline affordable. The company runs strikingly lean, having grown from around 240 employees in early 2023 to a few thousand by the end of 2025 while its annualized revenue run-rate reportedly approached $30 billion, an unusually high revenue-per-employee that gives it room to pay elite packages without Meta-scale headcount - Fortune. A smaller, more selective team means each hire is more valuable and easier to retain, which turns the fairness stance from idealism into strategy: when you employ fewer, better-aligned people, you rarely face the internal inequity that a one-off nine-figure counteroffer would detonate across the org.

The equity mechanics reward reading the fine print. Anthropic has been shifting newer grants toward restricted stock units rather than options, and toward double-trigger vesting that fully converts only on a liquidity event, a structure cleaner and more predictable than OpenAI's former profit-participation units but still fundamentally illiquid until a tender or IPO - Levels.fyi. For a candidate weighing an Anthropic offer, the questions that matter are the refresh policy (how much new equity you receive each year), the strike price if any options remain, and the company's tender cadence. Those three details decide whether a headline package compounds into real wealth or merely looks impressive on paper, and they vary enough between individual offers that two engineers hired at the same level can end up in very different financial positions.

The retention numbers turn that anecdote into data. Per SignalFire's 2025 State of Talent Report, Anthropic posted the highest two-year retention of any frontier lab at 80%, ahead of Google DeepMind at 78%, OpenAI at 67%, and Meta at 64% - SignalFire. Engineers were reportedly about eight times more likely to move from OpenAI to Anthropic than the reverse, and Anthropic converted 88% of technical offers and 95% of go-to-market offers it extended - SignalFire. People do not just join Anthropic, they stay, which structurally lowers the price it must pay to keep a team together.

That edge is not costless, and Anthropic did quietly reinforce it with money as well as mission, granting large retention bonuses to technical staff during the height of the 2025 raid. But the striking part is the ratio: with about 88% of technical offers accepted and eight-to-one net inflows from OpenAI, Anthropic spends far less per retained engineer than a lab hemorrhaging people into bidding wars - SignalFire. Even Amodei has privately voiced unease that the industry's mission could be "losing to money," a worry that is easier to hold when your own attrition is the lowest in the field. Retention, in other words, is Anthropic's cheapest form of recruiting.

Most striking is that Anthropic achieves this while explicitly declining to match nine-figure poaching offers. "We are not willing to compromise our compensation principles, our principles of fairness, to respond individually to these offers," Amodei said, adding that many staff "wouldn't even talk to Mark Zuckerberg" - Fortune. The practical lesson, for founders especially, is that mission, equity conviction, and internal fairness can substitute for raw dollars at the margin, but only if the base package is already elite. Anthropic is not cheap; it is disciplined. That discipline is exactly what the 2026 talent war tested to its limit.

5. The talent war that reset the ceiling: Meta, $100M, and a $1.5B "no"

In 2025, Meta tried to buy an AI lab's worth of talent one researcher at a time, and in doing so it reset the market's sense of what a single person can be paid. Whatever you think of the individual figures, the war permanently changed the reference points that OpenAI, Anthropic, and everyone else now negotiate against. It is impossible to understand 2026 pay without it.

The opening shot became public when Sam Altman said on a podcast that Meta had tried to poach OpenAI staff with signing bonuses "as high as $100 million," plus larger annual compensation, and claimed "so far none of our best people have decided to take them up on that" - CNBC. The CNBC segment below captured the moment the figure entered the mainstream, and it is the single best primary-source clip for the war's headline number.

The $100 million figure was immediately contested, and the dispute is instructive. Meta's own CTO, Andrew Bosworth, told staff in a leaked meeting that "Sam is just being dishonest," clarifying that packages of that scale were offered to only a small number of very senior leadership roles and were blended totals, not simple sign-on cash - TechCrunch. At least one poached researcher publicly called the sign-on framing "fake news." The honest reading is that nine-figure numbers were real but rare, reserved for a handful of elite targets, and structured as multi-year packages rather than a check on day one.

Untangling those disputed figures is itself a useful exercise, because it teaches how to read talent-war headlines. The reliable pattern is that the biggest numbers describe multi-year, mostly-equity packages for a small set of elite targets, not broad signing bonuses, and that individual cases resolve messily. Andrew Tulloch, the researcher who reportedly turned down a $1.5 billion Meta offer, ultimately did leave Thinking Machines Lab for Meta months later on undisclosed terms - Business Standard. The lesson for anyone benchmarking against these stories is to discount the headline, check whether an offer was accepted or merely made, and separate the handful of true nine-figure packages from the far more common multimillion-dollar ones. The ceiling is real, but it is narrow, and most engineers, even excellent ones, are competing in the tier below it.

What is not disputed is that some packages genuinely reached that tier. Meta secured 24-year-old researcher Matt Deitke with a reported $250 million over four years, including as much as $100 million in the first year, after he initially declined a smaller offer and Zuckerberg personally doubled it - Yahoo Finance. It poached Apple's foundation-models lead Ruoming Pang with a package exceeding $200 million - Bloomberg. And it reportedly offered Thinking Machines Lab co-founder Andrew Tulloch around $1.5 billion over six years, an offer he turned down and Meta later called inaccurate - Business Standard. Meta also spent $14.3 billion for a 49% stake in Scale AI to bring founder Alexandr Wang aboard as its first Chief AI Officer, an acqui-hire in all but name - CNBC.

The raid was wide as well as deep. Meta reportedly approached around 50 researchers before freezing AI hiring in August 2025, and the names who left OpenAI alone included the entire Zurich team of Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai, plus researchers like Jason Wei and Shengjia Zhao, who became chief scientist of Meta Superintelligence Labs - TechCrunch. Meta spent more than $1 billion on top of that to bring in investors Nat Friedman and Daniel Gross by buying out their venture fund. The scale signalled a strategic bet that talent density, not compute alone, was the binding constraint on catching up, and that a few dozen exceptional people could be worth more than the entire budget spent to acquire them.

The defenders responded in character. OpenAI's chief research officer, Mark Chen, sent a raw internal memo saying "I feel a visceral feeling right now, as if someone has broken into our home," and promised the company was "recalibrating comp" while refusing to abandon fairness - Entrepreneur. OpenAI even ordered a company-wide week-long shutdown to let exhausted staff recover from the recruiting siege - Futurism. Not everyone played: Microsoft AI chief Mustafa Suleyman, watching from the sidelines, said flatly of Meta's offers, "I don't think anyone's matching those things" - CNBC.

Not every move was a clean win, which is part of the lesson. OpenAI's own roughly $3 billion deal to buy the coding startup Windsurf collapsed over Microsoft intellectual-property concerns, whereupon Google swept in with a $2.4 billion reverse-acqui-hire for Windsurf's chief executive and key researchers, leaving most of the staff behind - Computerworld. These structured talent deals, buying a company largely to acquire a handful of people, became a favored workaround when direct offers failed, and they blurred the line between hiring and mergers in ways antitrust regulators are still digesting. The takeaway is that at the very top, compensation and corporate strategy have fused: a single hire can be worth structuring a billion-dollar transaction around.

The retention scoreboard is the clearest way to see who won the war, and it is not the biggest spender. The chart below shows two-year retention across the four major labs, and it captures the central irony of 2025: the lab that refused to match nine-figure offers retained its people best.

The interpretation matters for anyone competing for this talent. Meta's raid proved that money can buy almost any individual, but not loyalty at scale: by August 2025 it had frozen AI hiring after the spree, and the lab with the lowest offers kept the most people. Compensation sets the floor a candidate will consider, but mission, colleagues, and the credibility of the equity decide who actually stays. That is the single most important thing a smaller employer can learn from the war, and we come back to it in chapter nine.

6. How OpenAI and Anthropic pay stacks up against everyone else

OpenAI and Anthropic sit at the very top of the pay market, but the more useful comparison is not who pays the biggest number, it is whose equity you can actually sell. Ranking labs by headline total comp is easy and slightly misleading; ranking them by risk-adjusted, liquidity-adjusted value is harder and far more useful. Both views matter, so start with the raw ranking.

The Levels.fyi chart below ranks the top employers of AI engineers by total compensation, and it is the cleanest single picture of where the frontier labs sit relative to the broader market. OpenAI and Anthropic cluster near the top, but the spread underneath them is instructive: a great many well-known companies pay AI engineers a fraction of the frontier-lab rate.

The most important rival to understand is Google DeepMind, because it exposes the liquidity trade-off directly. DeepMind maps to Google's ladder, where an L5 research scientist reports around $413,000 and an L6 around $590,000, meaningfully below OpenAI and Anthropic at the top - Levels.fyi. But DeepMind's equity is publicly traded Alphabet stock that is liquid the moment it vests. There is no waiting for a tender, no valuation you cannot exit. A lower headline number that converts to cash on a schedule can beat a higher one trapped in a private company, and for risk-averse candidates it often does.

The right way to price that difference is with an explicit discount. A rational candidate should mentally haircut private, illiquid equity against public shares to reflect the years of delay, the valuation risk, and the chance that a liquidity event slips or disappoints. Applied honestly, a $1 million Anthropic package heavy in private stock and a $700,000 DeepMind package in liquid Alphabet shares can converge in real, spendable value, especially for someone who needs cash on a known timeline. This is not an argument against the private labs, whose upside is genuine and, in the last two years, spectacular. It is an argument for comparing risk-adjusted value rather than headline totals, which is the single most common failure in these decisions.

The private challengers, by contrast, compete on upside. xAI pays software engineers a median around $640,000, with staff packages above $1 million, on par with the frontier labs' mid-tier but paid in illiquid private stock - Levels.fyi. Europe's leading lab, Mistral, pays well by regional standards but is capped by a smaller balance sheet: it raised at roughly $14 billion in September 2025 and was reportedly in talks near $23 billion in 2026, an order of magnitude below the U.S. leaders, which limits how large a paper-equity component it can offer - Bloomberg. The pattern is consistent: valuation size sets the ceiling on equity-driven pay, which is why the two most valuable private labs also top the compensation tables.

When you join is itself part of the compensation. An engineer who took OpenAI equity before its valuation ran from roughly $300 billion to $852 billion in about a year captured that entire step-change; one who joins afterward inherits the same grant at a far higher entry price, worth much less in future appreciation - CNBC. The same logic governs the fast-moving challengers, from xAI's tie-up with SpaceX to Mistral's strategically backed European rounds: each rewards early conviction and penalizes late arrivals, which is why sophisticated candidates weigh a lab's trajectory, not just its headline valuation, when they compare offers.

The synthesis for a candidate weighing offers is to stop comparing single numbers and start comparing four things at once: base salary, equity size, equity liquidity, and the credibility of the valuation behind it. A DeepMind offer optimizes liquidity, an OpenAI or Anthropic offer optimizes size and upside, and an xAI or Mistral offer sits somewhere in between on different axes. The "best-paying" lab is genuinely different for a 27-year-old willing to gamble than for a 40-year-old with a mortgage, and pretending otherwise is how people talk themselves into the wrong job. This same scarcity logic, viewed from the employer's side, is what makes the whole market so expensive.

7. Why the numbers are this high: scarcity, winner-take-all, and the AI premium

Frontier-lab pay is extreme because a tiny number of people can move a trillion-dollar prize, and the labs would rather overpay than ship their next model a quarter late. This is not a bubble of irrational exuberance so much as a rational response to genuine scarcity layered on winner-take-all economics. Once you see the supply-and-demand math, the numbers stop looking crazy and start looking almost inevitable.

The supply side is brutally thin. Industry estimates put the number of people capable of leading frontier AI research at roughly 500 to 1,000 worldwide, of whom each lab needs dozens, against a demand curve set by companies convinced the payoff is measured in the trillions - HackerRank. When the prize is that large and the qualified pool that small, a $200 million package for a proven researcher is not extravagance; it is a rounding error against the value of shipping a better model months sooner. This is the same logic that prices franchise athletes and star fund managers, applied to a field where the "franchise" might reorganize the economy.

The winner-take-all structure is what turns scarcity into extreme prices. In a field where the gap between the best and second-best model can decide market leadership, the marginal researcher who shortens the path to that lead is worth a slice of the entire prize, not a market-clearing wage. Analysts describe superstar frontier researchers earning north of $30 million a year on exactly this logic, priced against outcomes rather than hours - HackerRank. It is the same reason a studio overpays for a director who can reliably deliver a billion-dollar franchise: when the payoff is convex and concentrated in a few hands, the people who can bend the outcome capture an outsized share of it, and competitors who refuse to pay simply lose the person to someone who will.

The premium is not confined to the superstar tier, either. The chart image below quantifies how much more AI and machine-learning engineers earn than generalist engineers at the same level across major tech companies, and the gap widens with seniority.

That premium shows up in the broader labor market too, not just at the labs. PwC's 2026 Global AI Jobs Barometer, which analyzed more than a billion job ads, found that roles requiring AI skills now carry a 62% wage premium, up from 57% a year earlier, and that AI-skill postings grew 69% against 9% for the market overall - PwC. In other words, the same forces inflating frontier-lab packages are raising pay for anyone who can credibly claim AI skills, just at a smaller magnitude.

The premium also widens sharply with seniority, which explains the shape of the whole market. PwC's data shows the AI wage premium climbing from single digits at entry level to well over 70% for senior roles, and varying enormously by sector, from modest gains in the public sphere to more than a doubling in some consumer markets - PwC. Stack that seniority curve on top of the frontier-lab scarcity premium and the extreme top-end numbers stop looking arbitrary: a senior researcher at a top lab is being paid the compounded product of two separate premiums, one for AI skills generally and one for the tiny pool who can work at the frontier specifically. Multiply two large multipliers and you arrive naturally at seven and eight figures.

This is also the strongest argument against the simple "it is all a bubble" reading. Broad AI pay may well cool if valuations correct or the supply of competent engineers catches up, and the median-salary plateau in the next chapter hints that it already is. But the scarcity at the very top is structural, not speculative: there is no fast way to manufacture a researcher who can lead a frontier training run, and no valuation swing changes that overnight. So even a market correction would likely compress the broad premium while leaving the superstar tier intact or rising. Anyone betting a career on this market should hold the two ideas separately: the ordinary AI job is a good, high-paying job that may normalize, while the frontier tier is a scarce franchise that is likely to stay scarce for years.

The result is a labor market that has split cleanly in two. A capable machine-learning engineer at a mainstream company earns a U.S. median around $170,000 to $179,000, a strong salary by any normal standard - Glassdoor. A frontier-lab engineer with the same job title earns $600,000 to over $1 million. Same skills on paper, three-to-five times the pay, because one group works on the model that might define the decade and the other works on the product around it. Understanding which side of that split a given role sits on matters far more than the title printed on it, which is exactly the question the next chapter tackles from the candidate's point of view.

8. Can you actually get one of these jobs? The bar, the interview, the visa

The bar at OpenAI and Anthropic is extreme, but it is not defined by pedigree, and the most common reason strong candidates get rejected is not a failed coding round. That surprises people who assume these jobs go only to PhDs from a handful of universities. The reality is more open and, in some ways, harder, because the labs screen for things a resume cannot fake.

Anthropic is explicit that it does not require a PhD: roughly half its technical staff hold one, meaning half do not, and about half had no prior machine-learning experience before joining - DataExec. Its own careers guidance tells applicants to put independent research, an insightful blog post, or substantial open-source contributions at the very top of their resume, because the company weighs "direct evidence of ability" over credentials. The signal that gets you in the door is demonstrated capability: a repo people use, a paper people cite, a system you actually built, not a logo on your degree.

This is a genuine shift from how elite tech hiring worked a decade ago, and it favors builders over test-takers. Because both labs treat public artifacts as primary evidence, the highest-leverage thing an aspiring hire can do is ship something real and visible: a widely used open-source library, a faithful reproduction of a notable paper, a technical writeup that experts actually reference. Credentials still help at the margin, but they have become a tiebreaker rather than a gate, and a self-taught engineer with a serious portfolio can out-compete a PhD with none. That is unusually good news for career-changers, who can build the exact evidence these labs weigh without going back to school, on their own time and at their own pace.

The interviews reward building over trivia. OpenAI's software-engineering loop has largely moved away from classic algorithm puzzles toward recreating a small working system in about an hour, followed by a scalability-focused system-design round, across roughly four to six onsite sessions and an end-to-end process that can run eight to twelve weeks - Exponent. Anthropic runs two live coding rounds at medium-to-hard difficulty plus system design, but with a telling twist: its values and mission-alignment round is the single most common reason for post-onsite rejection, ahead of technical performance - Final Round AI. At a lab whose entire pitch is careful, safety-minded development, culture fit is not a formality; it is a gate.

The concrete shape of the process is worth knowing before you start. Anthropic's coding rounds run roughly 45 to 60 minutes each at medium-to-hard difficulty, typically in Python, and lean on system design and reasoning as much as raw algorithms - Final Round AI. OpenAI's build-a-real-system round rewards engineers who can produce working software under time pressure rather than recite textbook solutions, and the full loop plus scheduling routinely stretches across two to three months. Budget for that timeline, prepare for depth over trickery, and treat the values conversation as seriously as the technical ones, because at these labs it is frequently the round that actually decides the outcome.

Both labs also run deliberate on-ramps for non-traditional candidates, which is where career-changers should look. The clearest examples are structured programs rather than lateral hires:

• OpenAI Residency is a six-month paid program aimed at researchers and engineers from adjacent fields like math, physics, and neuroscience, plus strong self-taught engineers - OpenAI.
• Anthropic Fellows funds people to work on AI-safety research even without a PhD, prior ML experience, or published papers.
• Open-source and independent work functions as a live audition, since both labs treat public artifacts as primary evidence.

These programs matter because they convert the intimidating "you need to already be famous" narrative into a concrete path. The person who spends six months shipping a serious open-source project or a well-argued research writeup is, in these labs' own framing, building exactly the portfolio that gets interviews. That is a more actionable plan than chasing another degree, and it is available to people already in the workforce.

The last hurdle for many candidates is immigration, and it doubles as a salary revelation. Both labs sponsor visas heavily, which is why their base-pay data is public in the first place, and top researchers increasingly use the O-1A "extraordinary ability" visa, which has no annual cap and no lottery, as an alternative to the H-1B and a bridge to a green card - Casium. For an international engineer, the practical sequence is to build the citable, adopted, public body of work that both the hiring bar and the O-1A criteria reward, because the same evidence serves both goals. Get that right and the visa becomes a paperwork exercise rather than a wall.

One last practical filter is location, because neither lab is remote-friendly. Both are office-first: OpenAI runs a structured hybrid that expects employees to live near an office and show up regularly, while Anthropic advertises only a small share of roles as remote and generally expects even those staff on site around a quarter of the time - JobsByCulture. Pay is also location-adjusted, which is why the same title can carry very different numbers in San Francisco, London, or elsewhere. For a candidate, this means the offer is inseparable from a relocation decision, and the headline compensation should be read against the cost and disruption of moving to an expensive hub. A $900,000 package that requires uprooting your life to the Bay Area is a different proposition than the same number where you already live, and the labs are well aware of that difference when they make the offer.

9. What it means if you are hiring, or being hired

If you are trying to hire this talent without a trillion-dollar balance sheet, the 2026 data delivers one liberating message: money sets the floor, but it does not decide who stays. The lab that refused to match nine-figure offers kept the most people, which means the game is winnable on axes other than cash. But it also means the sourcing and retention playbook has changed, and clinging to the old one is how good teams lose candidates they never even reached.

Start with the demand reality, because it is stark. AI skills are now the single hardest capability to hire for globally, with 72% of employers reporting difficulty filling roles in ManpowerGroup's worldwide survey, and demand for AI skills appearing in a rising share of postings year over year - MSH. The best AI engineers are almost never active applicants; they are employed, courted weekly, and invisible to a job board. Waiting for inbound applications in this market is a strategy for hiring the people nobody else wanted.

The scarcity is intensifying, not easing. Demand for AI skills now shows up in about 2.5% of U.S. job postings, up more than 50% in a single year and nearly 300% over the decade, while the supply of people who can credibly fill those roles grows slowly - MSH. When demand climbs that fast against a fixed pool, the employers who win are not the ones with the most open requisitions but the ones who can identify and reach the right passive candidate before three other companies do. Speed and reach, not budget, become the scarce resource, and they are the two things a smaller team can actually build.

That has pushed sourcing away from keyword searches on a single network toward AI systems that read context across the whole open web. Gartner projects that 82% of HR leaders plan to deploy agentic AI for recruiting by mid-2026, and modern platforms now run natural-language searches across 800 million to more than a billion profiles spanning LinkedIn, GitHub, and beyond - Mokka. This is the category where a tool like HeroHunt.ai fits as one option among several: its AI Recruiter takes a plain-language role brief, autonomously searches over 1 billion profiles, screens candidates on context rather than keyword matches, and runs personalized outreach on autopilot, which is exactly the shape of sourcing that a scarce, passive, GitHub-native talent pool demands. For a small team competing against Meta's budget, leverage on the sourcing side is one of the few places the playing field can be levelled, and it costs nothing to start - HeroHunt.ai.

The sourcing shift is not about replacing recruiters but about widening the top of the funnel to include people who will never apply. An AI sourcing pass can surface a systems engineer whose only public footprint is a well-starred GitHub repository, or a researcher who published once and then disappeared into an unrelated job, candidates a keyword search on a single network would never return. Reaching them first, with a message that shows you actually understand their work, is worth more in this market than a marginally higher band, because the binding constraint is discovery, not willingness to pay. The team that finds the person nobody else found rarely has to win a bidding war for them.

Retention is the other half, and here the labs teach a clear lesson. The tactics that actually held teams together in 2025 were rarely a single giant check:

• Rolling equity refreshes that keep two-to-three years of unvested stock on the table at all times, so leaving always means walking away from money.
• Credible liquidity, via tender offers that convert paper to cash on a known cadence, which is what lets equity conviction substitute for a bigger base.
• Mission and colleagues, the intangibles Anthropic leaned on when it declined to match offers and still retained 80% of staff over two years - SignalFire.

For a startup, the trick is to borrow the labs' retention playbook at a smaller scale. You can offer annual equity refreshes so there is always unvested upside on the table, commit to a transparent secondary or tender program so employees believe the paper can become cash, and tie the equity story to a mission specific enough that people can picture why staying matters. None of that requires a nine-figure budget; it requires clarity and follow-through. The most common self-inflicted wound is opacity: a candidate who cannot understand or trust an equity grant will discount it to near zero, which is how smaller companies lose people they could have kept for a fraction of a lab's cost.

The practical synthesis for a founder or talent leader is to compete where you can win. You will not out-pay Meta, but you can out-source it by reaching passive candidates first, out-structure it by making your equity legible and liquid, and out-mean it by offering work people believe in. Deep dives on that playbook live in HeroHunt's guides to winning the AI talent war and recruiting AI engineers in 2026, and the compensation benchmarking in its AI talent compensation guide pairs naturally with the numbers here. The teams that win the next hire are the ones that treat pay as the price of entry, not the whole strategy.

10. The future: liquidity events, cooling signals, and agentic hiring

The next phase of frontier-lab pay hinges on three things: whether these companies go public, whether the broader AI wage premium cools, and whether AI itself changes who gets hired at all. Nobody can predict the exact numbers, but the forces are visible now, and they point toward a market that stays extraordinary while becoming more legible and, at the margins, more disciplined. Each force cuts a different way.

The biggest wildcard is liquidity. Both labs are marching toward the moment their paper becomes tradeable, whether through continued tenders or eventual IPOs, and that transition will reprice everything. When Anthropic employees declined to sell into a $350 billion tender because they expected more from an IPO, they were making a concrete bet that the illiquid-equity phase is temporary - Bloomberg. A public listing would convert today's paper millionaires into actual ones and, in doing so, remove the single biggest asterisk on every offer letter in this guide. It would also make frontier-lab comp directly comparable to public Big Tech for the first time, which could either validate the premiums or compress them.

The timing of that event is itself a compensation variable. An engineer joining now is implicitly betting on when, and at what price, their private equity becomes tradeable, and that bet can dominate the headline offer. Anthropic staff who declined the 2026 tender were wagering that patience pays; OpenAI staff who sold into the $500 billion tender were locking in certainty. Neither is obviously right, and the correct choice depends on personal risk tolerance far more than on which lab's brand is stronger. As IPOs approach, expect this timing question to become the central negotiation, quietly displacing the raw grant size that dominates offer conversations today.

There are early signs the raw AI premium is stabilizing rather than climbing forever. The image below tracks median AI-engineer pay over time, and it shows a peak followed by a plateau, a hint that the market for ordinary AI talent is maturing even as the superstar tier keeps escalating.

A cooling middle does not mean a cooling top. The two markets have decoupled: median AI-engineer pay can plateau while a proven researcher's price keeps rising, because the scarcity that drives the superstar tier is not eased by a larger supply of competent generalists. Watch for this divergence to widen, with the vast majority of AI engineers settling into a high-but-normal band and a few hundred people commanding numbers that look more like sports contracts than salaries.

For anyone tracking this market, a few concrete signals will tell you which way it is breaking. Watch the frequency and pricing of employee tenders, because a lab that stops offering liquidity is either close to an IPO or short on demand for its shares. Watch for actual IPO filings, which would convert paper to a public price and reset every benchmark in this guide. And watch whether Meta, having frozen hiring after its 2025 spree, re-enters the market, because a second raid would push the top tier higher again. The direction of those three indicators, more than any single salary datapoint, will tell you whether 2027 pay looks like a continuation of the boom or the beginning of its normalization.

The third force is the strangest: AI is starting to reshape hiring itself. The same labs paying these sums are also shrinking the number of people it takes to build a product, and OpenAI's own leadership signalled a shift in early 2026, with Sam Altman telling staff he wanted to slow hiring and "do more with fewer people" even as headcount plans reached toward 8,000 - Engadget. If agentic tools let a leaner team do more, the labs may bid even harder for the few people who can direct that leverage, concentrating pay further at the top. On the recruiting side, the rise of autonomous sourcing and screening means the contest for scarce engineers will increasingly be fought by software on both sides of the table.

There is a reflexive quality to this worth naming. The labs building agents that compress engineering teams are simultaneously the biggest bidders for the humans who build those agents, so their own product may keep shrinking the number of roles even as it raises the price of the few that remain. The same automation is arriving in recruiting itself: autonomous systems that source, screen, and draft outreach are moving from novelty to default, which means the competition for scarce engineers will be decided partly by who deploys the better software rather than who employs the larger recruiting team. The market that pays the most for AI talent is, fittingly, being reshaped by AI itself, and the people and companies who understand that early will spend the next few years a step ahead of the ones still optimizing yesterday's playbook.

For the individual engineer, the strategic reading is that the window to enter this market on favorable terms is open but not permanent. The base salaries are already historically high, the equity upside is real but tied to liquidity events that will eventually arrive, and the bar rewards demonstrated capability over credentials, which is the one thing a motivated person can build starting today. For employers, the message is the mirror image: the talent will only get scarcer and more expensive, so the durable advantage goes to those who can find, reach, and keep people better, not merely pay them more.

Conclusion: how to actually use these numbers

The right way to read frontier-lab pay is as a level-indexed range of base salary plus illiquid equity, discounted for the risk and delay of turning that equity into cash, not as the single seven-figure number in a headline. Do that, and the market becomes navigable rather than dazzling. For a candidate, the decision framework is compact: identify which rung of which ladder an offer really puts you on, separate the guaranteed base from the equity bet, and interrogate exactly when and how you can sell that equity, because a DeepMind offer that vests into liquid Alphabet stock and an Anthropic offer that vests into a private valuation are different instruments even at the same dollar total.

For employers, the framework is the inverse. You almost certainly cannot win a bidding war against a company willing to pay one researcher $250 million, and the 2026 data says you do not have to: the lab that refused to match kept the most people. Compete on the axes where scale is not decisive, by reaching passive candidates before your competitors do, by making your equity legible and liquid, and by offering work people find worth staying for. Match your comp to the level, not the hype, and spend your scarce advantage on sourcing and retention rather than on a single spectacular offer that resets everyone's expectations.

The two labs at the center of this story got where they are by pricing talent as the scarcest input to the most valuable technology of the decade. Whether you are trying to join them or trying to hire against them, the winning move is the same: understand precisely what the number means before you chase it. The people who do that, on both sides of the table, will spend the next few years making clear-eyed decisions in a market engineered to dazzle everyone else.

Written by Yuma Heymans (@yumahey), who built HeroHunt.ai, the world's first AI Recruiter, now used by 15,000+ recruiters to source and engage engineers from over a billion profiles on autopilot. He has spent years on the hiring side of exactly the talent market this guide dissects, and writes about how scarcity, equity, and AI-driven sourcing are quietly rewriting the rules of who gets hired and for how much.

This guide reflects the AI compensation landscape as of August 2026. Private valuations, tender terms, and pay bands change quickly, so verify current figures before making a decision based on them.