AI Research Scientist Salary Guide 2026: What the Labs Actually File

Last updated: November 4, 2026

By Robert Ardell, Co-Founder and Strategic Advisor, KORE1

AI research scientist salaries in 2026 start at a $140,300 federal median, reach about $183,000 to $194,000 at Meta and Google, and hit $265,000 to $380,000 at Anthropic and OpenAI, before equity. The lab figures are medians from 2025 H-1B wage filings with the Department of Labor, which list base salary only. Stock, signing money, and retention bonuses sit on top, and at the frontier labs they are most of the package.

Most salary pages give you one number for this job. It’s the wrong number for almost everyone reading.

Bias first. KORE1 runs an AI research scientist staffing desk, so I have a reason to tell you this hire is hard to price, and it is. Gregg Flecke already covered the hiring side in our guide to hiring an AI research scientist, including the three jobs that hide under one title. I’m not going to re-argue that here. This page is about the money, and specifically where the money numbers come from, because the sources disagree by more than $200,000 at the median and almost nobody tells you why. So here’s why.

What Does an AI Research Scientist Salary Actually Measure?

An AI research scientist salary is the base pay for someone hired to develop new machine learning methods rather than ship known ones, usually with a doctorate and a publication record. Every public figure measures a different slice of that population, which is why the figures don’t agree.

Here’s what four common sources say, side by side.

The Bureau of Labor Statistics number is the floor everyone quotes. It’s honest. It’s also a category of about 38,600 jobs that includes university-adjacent labs, defense research, and plenty of people who have never trained a neural network, so it tells you what the field pays in general and very little about what a strong LLM researcher will accept. Projected growth is 22% from 2025 to 2035, with about 2,900 openings a year. Good growth. Wrong benchmark.

Payscale sits even lower, at $129,987 for its AI research scientist title. That’s below the federal median for the broader occupation, which tells you it isn’t describing the people the big labs hire, whoever it is describing. Glassdoor’s machine learning research scientist page reports $192,960, but that’s total pay, bonus included, on only 44 submissions. Strip the bonus and its base range runs about $131,000 to $197,000. Small sample. Still closer to the market we recruit in.

For base pay, I’d bet on the bottom row over the other three. Nobody filled it out for fun.

What the Big Labs Tell the Department of Labor

When a U.S. employer sponsors someone on an H-1B visa, it files a Labor Condition Application stating the job title, the work city, and the wage it will pay. The Department of Labor publishes those filings. Nobody self-reports and nobody rounds up to impress a friend. It’s what the company wrote on a federal form.

Three caveats, and they matter. The figure is base salary only, with no stock or bonus. A filing is a stated wage for a petition, not proof somebody took the job. And the sample is limited to sponsored hires, which leaves out every researcher who didn’t need a visa.

Here’s 2025.

Look at the title column before you look at the money. Anthropic and OpenAI didn’t file a single application under “research scientist” in 2025. Not one. Both put researchers and engineers under Member of Technical Staff, so if you search salary data for the job title, the two labs setting the top of this market simply aren’t in your results. That alone keeps the top of the market out of any average built on the job title. Search the title, miss the market. At OpenAI, the 24 filings whose titles also named research carried a median base of $380,000.

Meta is the opposite story. Its middle half sits inside a $35,000 band, $179,159 to $214,032, across 441 filings, which is what tight leveling looks like on paper. Tight. You don’t negotiate your way out of a Meta band. You negotiate your way into a higher level, and then the equity grant does the heavy lifting.

Google’s top filing, $400,000, is one person. Treat it that way.

Does the City Change the Number?

Less than you’d think at the big employers, more than you’d think everywhere else.

Meta’s 2025 research scientist filings had a median of $182,000 in Menlo Park, $182,293 in Bellevue, and $192,761 in New York. That’s nearly flat across three very different cost-of-living markets. Same level, same pay. A researcher moving from Menlo Park to Bellevue at Meta keeps roughly the same base, which means a candidate in Seattle comparing your offer to a Meta number is comparing it to a Bay Area number whether either of you realizes it or not. OpenAI’s 20 Bellevue filings sat at $255,000 against $325,000 for its 84 in San Francisco, so not every lab works the same way, but the pattern at Meta is hard to miss.

Across the whole occupation, geography moves the median a lot. These are BLS figures for May 2025, all computer and information research scientists, not just AI.

Two things jump out. Washington has more of these jobs than any metro on the list, 3,420, and pays below the Bay Area and Seattle. San Diego has slightly more of these jobs than San Francisco, 1,680 to 1,670, and a median about $50,000 lower. The BLS doesn’t say who the employers are in each metro, so I won’t pretend to know the reason, but if you’re hiring in Southern California, the San Diego figure is the one a local candidate has probably seen.

Austin barely registers. Three hundred jobs.

Base Is the Smaller Part at the Top

Everything above is salary. At the labs that set the market, salary is the boring line on the offer letter. Read the grant.

Glassdoor’s machine learning research scientist data puts additional cash pay, mostly bonus, at roughly $24,000 to $45,000 a year on top of base. Equity is where it stops being a normal job. Meta offered some researchers packages worth as much as $300 million over four years in 2025, and OpenAI responded with faster vesting and retention bonuses as high as $1.5 million, according to Wall Street Journal reporting summarized by DeepLearning.AI’s The Batch. Maybe a few dozen researchers saw offers like that. They also move what the next hundred people ask for. Every time.

For planning, our cluster uses three bands, and they hold up against the filings with one exception.

• Frontier research scientists at the big labs, $200,000 to $300,000 base, with total packages from $600,000 to well past $1.5 million.
• Applied research scientists who adapt published work to a company’s product, $180,000 to $260,000 base and $220,000 to $400,000 total. This is the hire most companies actually make.
• Strong ML engineers carrying a research title, $160,000 to $230,000 base. Not a knock on them. Just a different job, priced like one in our machine learning engineer salary guide.

The exception is OpenAI. Its filed median of $325,000 already sits past the top of the frontier base band, and the research-titled filings at $380,000 sit well past it. That end of the market moved fast in 2025. If you’re benchmarking against OpenAI specifically, use its filings, not a band.

What Moves One Offer Above Another

Credentials matter less than people expect and the subfield matters more. Subfield is the price tag.

The BLS lists a master’s degree as the typical entry-level education for the whole occupation, and plenty of applied researchers we’ve placed don’t hold a doctorate. For frontier work, a PhD is close to standard. What actually separates two PhDs at offer time is first-author papers at venues like NeurIPS, ICML, and ICLR, and whether those papers are in the area the company is spending on right now. In 2026 that means LLM pretraining and post-training, reinforcement learning, inference efficiency, and interpretability. A brilliant vision researcher and a merely good RL researcher can get offers that land in the reverse order of their talent. Annoying. True.

Supply is getting tighter, too. The Stanford HAI 2026 AI Index found that the number of AI scholars moving to the United States has dropped 89% since 2017, and 80% in the last year alone. Fewer arrivals. Same demand. You can guess what that does to price.

One more lever, and it’s the one hiring managers forget. Compute. A researcher choosing between two offers will often ask how many GPUs they can actually use and whether they can publish. I watched a candidate take roughly $40,000 less in base last spring because the smaller company guaranteed a dedicated training cluster and conference travel, and the bigger one couldn’t say who would own the results. No regrets a year later. Money was close enough. The work wasn’t.

Setting a Band When You Aren’t a Frontier Lab

You aren’t going to match OpenAI. Neither is anyone else outside about five companies, and the candidates know it, so stop treating the top row of the filings table as your competitor. It isn’t. A Series B company that tries to pull a Google DeepMind researcher on base alone usually spends six weeks on the search, loses at the counteroffer anyway, and reopens the req with a smaller number and a slightly worse name among the small group of people who talk to each other about these jobs.

Start with which seat you’re filling. If the work is applied, the realistic base band is $180,000 to $260,000, and the Meta and Google filings, clustered between about $179,000 and $217,000 in their middle halves, are the fair comparison. In Southern California outside the Bay Area, expect the lower half of that range. Then decide what you can offer that the big labs can’t, usually a named problem, publication rights, and a shorter path to leading something. Run the exact role through our salary benchmark assistant before you commit to a number. Then commit.

Most of these seats are a direct hire, because research people want equity and a lab that will still exist in three years. The exception is a bounded question with a deadline, where a contract research scientist for a quarter or two can answer it before you build a team around it. And if the role starts looking more like model serving than model invention, the LLM engineer salary guide is probably the band you want.

KORE1 has been placing technical people since 2005, across 30-plus U.S. metros, and our average time-to-hire on IT roles is 17 days with 92% of placements still in the seat a year later. Research searches run longer than that on purpose, usually four to eight weeks, because the screen is reading someone’s papers, not checking a keyword. The comp conversation happens on day one. Not week six.

What People Ask About Research Scientist Pay

Does OpenAI really pay researchers more than $300,000 in base?

Its 2025 filings say yes, with a $325,000 median base across 109 Member of Technical Staff applications and $380,000 across the 24 whose titles named research.

That’s before stock, which at OpenAI is usually the larger number. Those filings cover sponsored hires only, so treat them as a strong signal rather than a census.

How much is the PhD worth in dollars?

Nobody publishes a clean PhD premium for this job, so be wary of anyone who quotes you one.

The filing data doesn’t record degrees, and the aggregators mix applied and frontier roles. What we see in offers is that the doctorate gets you into the frontier conversation and the publication record sets the number once you’re in it. A strong master’s with shipped applied work can out-earn a PhD with thin publications, and it happens more than people admit.

What does a first industry research job pay?

$135,000 to $180,000 in base is the realistic entry range at the large tech employers, based on the bottom quarter of 2025 research scientist filings at Meta and NVIDIA.

Meta’s lowest filing was $135,516 and its 25th percentile $179,159. NVIDIA’s 25th percentile was $160,000. At a startup, base can land lower and the equity higher, and you’ll want to know what that equity is worth before you count it. Ask.

If we can’t win on equity, should we overpay on base?

Pay the top of the fair band, not above it, and spend the rest of your effort on the parts of the offer big labs can’t match.

An extra $30,000 rarely flips a researcher who has a frontier-lab offer in hand. A guaranteed compute budget, a real research mandate, and the right to publish sometimes do. I’ve seen it go both ways, but the second way more often. Cash rarely breaks the tie.

Why do Payscale and Glassdoor disagree by $60,000?

Partly because they measure different things, Payscale an average salary for its AI research scientist title and Glassdoor total pay, bonus included, for 44 machine learning research scientists.

The rest is sampling. Neither site publishes who its respondents work for, so I can’t tell you which employers are pulling each number up or down, and I’d distrust anyone who claims to. Compare base to base. Then decide which seat you’re hiring for, because neither average is your market until you have.

If you’re about to write a number into an offer letter for this role, talk to one of our AI recruiters first. We’ll tell you which column of the filings table your candidate is actually comparing you to, and whether the seat you’ve scoped is research at all.

Filing figures are from 2025 H-1B Labor Condition Application disclosures, as indexed by h1bdata.info and pulled November 4, 2026. BLS figures are from the May 2025 Occupational Employment and Wage Statistics release for occupation 15-1221.