What AI Engineers & Researchers Actually Earn in 2026 — and How to Read an Offer

If you've read that an AI researcher got a "$100 million signing bonus," or that OpenAI engineers "make $1.5 million," you've been served the hype, not the reality. The truth is more useful: AI-lab compensation is genuinely the highest in the history of software — but the number that matters is almost never the one in the headline, and most candidates evaluate offers using the wrong figure entirely.

We pulled this together from three kinds of evidence: U.S. Department of Labor H-1B wage filings (which are legally binding and list base salary only), verified employee-reported total compensation, and on-the-record reporting from Bloomberg, TechCrunch, CNBC, and the WSJ. Where a number is contested or thinly sourced, we say so. By the end you'll know what these jobs actually pay by level, why two offers with the same headline can differ by hundreds of thousands of real dollars, and the exact questions to ask before you sign.

First, the trap: "base salary" and "total comp" are different planets

This is the single most important thing to understand, and it's where almost everyone goes wrong. There are two completely different numbers floating around for the same job:

The base salaries quoted in government wage filings are real and reliable — but at a frontier AI lab, base is only 30–60% of the package. The rest is equity. At senior levels, equity is 40–70% of total compensation. So when one source says "OpenAI pays $310K" (a base-salary filing) and another says "OpenAI pays $870K" (total comp), both can be correct — they're measuring different things. Internalize this before you compare any two offers.

The floor: what base salaries actually are

These come from U.S. Department of Labor H-1B/LCA filings — the most reliable public data that exists, because companies are legally bound to the figures they file. Remember: base only, no equity, no bonus.

Two things jump out. First, base salaries at OpenAI, Anthropic, and xAI cluster remarkably tightly around $275K–$315K median — the labs are competing on equity, not base. Second, the spread within a single title is enormous (Anthropic's filed MTS base runs from $156K to $690K) because one title can span everything from a new grad to a manager.

The real number: total compensation by level

Now the figure that actually lands in your account over four years. These are verified, employee-reported total-comp packages (base + annualized equity + bonus). Treat them as ranges, not promises — equity values move with each funding round.

The takeaway for a candidate: a senior role at a frontier U.S. lab realistically clears $550K–$870K+ in total comp, and the differences between offers live almost entirely in the equity line — which, as we'll see, is the part you can least take at face value.

One big geographic caveat: London frontier-lab comp runs roughly 30–40% below the San Francisco Bay Area in dollar terms. A DeepMind L5 in London is a very different number than the same level in Mountain View. If you're comparing across geographies, convert carefully.

Researchers vs. engineers: the premium hides in the equity

A persistent question: do "researchers" out-earn "engineers"? The honest answer is nuanced. At Google DeepMind, base salary is essentially identical across tracks at a given level — research scientists get roughly 5–15% larger equity grants, not higher base. At Anthropic, research roles run about 15–30% higher total comp per level, and the gap widens at senior tiers.

At OpenAI the title-level data shows research scientists earning nearly 2× the median of software engineers — but that's largely a leveling effect (researchers cluster at higher levels), not a per-level premium. The defensible conclusion: at the same level, the researcher premium is modest and lives in equity, not base. The eye-watering researcher packages you read about aren't a different pay band — they're a tiny number of frontier specialists (pretraining, reinforcement learning, reasoning) being bid on individually.

Outside the labs, the pattern rhymes: machine-learning engineers typically earn a meaningful premium over generalist software engineers — on the order of 25–40% — though the size varies sharply by employer.

The equity is the whole game — and it's not simple

If 40–70% of your package is equity, you need to understand what kind of equity, because the instruments behave completely differently.

OpenAI: Profit Participation Units (PPUs) — and a major 2025 change

OpenAI historically paid equity as Profit Participation Units: a contractual claim on future profits, not ownership shares, vesting 25% per year over four years. The catch was a return cap (originally around 10×), with value above it flowing to OpenAI's nonprofit — and because the cap was a multiple of the original issue price, it effectively shrank as the valuation rose (toward ~4× for later grants).

That changed materially: in its October 2025 restructuring into a public-benefit corporation, OpenAI removed the capped-profit model, and employees now reportedly hold more conventional equity participating proportionally in upside. If you're reading an older guide that describes a hard 10× PPU cap, it's out of date.

Anthropic, and private startups: double-trigger RSUs and options

Anthropic grants trend toward double-trigger RSUs: they require both time-based vesting (typically 4 years) and a liquidity event (IPO, acquisition, or an approved tender) before they turn into anything sellable. Earlier-stage startups grant stock options (ISOs/NSOs) with a strike price set to the company's 409A valuation — which is deliberately far below the headline "preferred" price investors pay.

Public labs (Google, Meta, Microsoft): liquid RSUs

This is the structural advantage that rarely makes headlines: public-company RSUs are essentially cash at vest. They settle into shares you can sell on the open market, they never go "underwater," and there's no liquidity event to wait for. A $400K public RSU and a $400K private grant are not worth the same thing.

How you actually get paid: liquidity is the hidden variable

Private equity only becomes money when you can sell it. In 2025–2026 that happened mainly through company-run tender offers, and the details matter:

• OpenAI ran a roughly $6.6B tender in late 2025, raising the per-employee cash-out cap from $10M to $30M; about 75 people hit the $30M maximum.
• Anthropic ran a roughly $6B tender in April 2026 priced at a $350B valuation — and notably, employees largely refused to sell, betting on higher prices. Secondary-market indications implied valuations approaching $1T. (Reported by Bloomberg.)
• In May 2026, both OpenAI and Anthropic cracked down on unauthorized secondary transfers (SPVs, forward contracts, tokenized rights) — a reminder that you generally cannot sell private shares whenever you want.
• Both filed confidentially for IPOs in mid-2026, and SpaceX (which had absorbed xAI) completed a record IPO in June 2026 — so the liquidity picture is shifting fast.

The lesson for an offer: tender liquidity is probable-but-not-guaranteed, on a timeline you don't control. Treat private equity as illiquid until proven otherwise.

The eye-popping numbers, reality-checked

The talent-war headlines are real news, but they're routinely misread. Here's what holds up:

Sam Altman said publicly (June 2025) that Meta offered his staff "$100 million signing bonuses." But Meta's own CTO, Andrew Bosworth, pushed back — the figure applied to "a few people, very senior leadership," and was structured as RSUs vesting over years tied to performance, not an upfront cash bonus. One named recipient, researcher Lucas Beyer, flatly called the $100M-sign-on framing "fake news." So: nine-figure total packages for a handful of elite people, yes; $100M cash on signing, no.

The WSJ reported Meta offered Thinking Machines co-founder Andrew Tulloch a package "as much as $1.5B over six years." Meta called that description "inaccurate and ridiculous," and the figure was contingent on stock performance. Tulloch did eventually join Meta — reportedly for less. Treat $1.5B as a contested headline, not a fact.

The genuinely confirmed mega-numbers are structural, not salaries. Bloomberg reported Meta hired Apple's Ruoming Pang on a package over $200M. And the truly enormous figures are "reverse acqui-hires" paid to founders and teams via licensing deals: Google–Windsurf (~$2.4B), Google–Character.AI (~$2.7B), and Microsoft–Inflection (~$650M). These are how the billion-dollar numbers get paid — and they're not job offers you can apply for.

What's driving all of it is scarcity: a few hundred people can credibly train a frontier model, and labs that raised at $32B (SSI, with no product) or $350B–$1T (Anthropic, OpenAI) will pay almost anything to keep them. OpenAI reportedly responded to the poaching with retention grants and CRO Mark Chen telling staff he'd "fight to keep every one of you."

How to read — and negotiate — your own offer

This is the part that actually changes your life, and it's where the research compounds into something practical. Six rules:

1. Negotiate your level first — it dwarfs everything else

The jump from one level to the next is worth far more than haggling within a band. At OpenAI, L4 total comp runs roughly $560K–$750K while L5 runs ~$900K–$1.3M. Fighting a down-level, or pushing to be slotted one level higher, is the single highest-value move you can make. (Levels aren't standardized — a Google L5 is not a Meta E5 — so research the specific ladder.)

2. Convert every offer to risk-adjusted cash before comparing

Never compare headline-to-headline. Take a public RSU at roughly face value (it's liquid at vest). Discount private equity 20–30% for a well-funded lab with an active tender market — more (30–50%) if there's no tender history or the valuation is stale. Only then are two offers comparable.

3. Treat PPUs / private grants as capped, illiquid, and unverifiable until you ask

At offer time you often aren't told how many units exist or what percentage you own — so you can't independently value the grant. Ask for the unit count and the implied total. Apply both an illiquidity discount and (for any capped instrument) a cap discount.

4. Run the equity diligence checklist

Before you sign, get answers to: option type (ISO/NSO/RSU); strike price; the most recent 409A valuation and its date; what % of fully-diluted shares your grant represents; the vesting schedule and cliff; the post-termination exercise window; tender-offer history and cadence; and what happens to unvested (and vested-but-illiquid) equity if you leave.

5. Probe the liquidation overhang

Ask the uncomfortable question: "What exit value has to be cleared before common stock is worth anything?" Investors hold preferred shares with liquidation preferences that get paid first. If the capital raised exceeds a plausible exit value, your common equity can be worth roughly zero today — regardless of the headline valuation.

6. Model the tax and the triggers, not just the grant

The traps that quietly cost people the most: single- vs. double-trigger RSUs (double-trigger is worthless if the liquidity event never comes); the 90-day post-termination window to exercise options (miss it and ISOs convert to NSOs); and the AMT bill you can owe on the paper gain from exercising — tax on money you can't yet access. Budget cash for the exercise and the tax before you'd ever need it.

The bottom line for job seekers

AI-lab compensation in 2026 is extraordinary and real: a senior engineer or researcher at a top U.S. lab is looking at $550K–$870K+ in total comp, and the genuine frontier specialists command far more. But the headline numbers conflate base with total comp, treat illiquid private equity as if it were cash, and amplify a handful of negotiated outliers into a fake "market rate."

Your edge as a candidate is to do what most don't: separate base from total comp, risk-adjust the equity, negotiate the level, and run the diligence checklist before you sign. The offer with the bigger headline is frequently the smaller real number.