Data scientist salary

§Four ways the Levels.fyi by-company number misleads in 2026

If you searched "data scientist salary at [Company X]" and read the median total comp, you got a number. That number is misleading in four specific ways that compound over a 6-year horizon. Each costs six figures of expected comp if you miss it.

One: title arbitrage. At Meta IC5 the same human earns $343K as Analytics DS, $440K as Applied Scientist, $475K as Research Scientist with PhD, $418K as ML Engineer. A 50 percent spread inside one company, one level, one location. Most candidates shop by company. The dominant variable is the title on the offer letter.

Two: the AI-lab tier is missing from a company search. OpenAI L5 Member of Technical Staff sits at $1.1M nominal; Anthropic L5 at $850K. Searching "data scientist at OpenAI" returns nothing useful because the internal title is MTS, not Data Scientist. If your mental anchor for senior DS comp is $400K, the actual market ceiling is 3x higher.

Three: the AI-skills premium is bimodal. PwC's headline 56 percent AI wage premium hides a 25x spread. Production ML work captures 80 to 120 percent of the premium; dashboard and experimentation work captures 5 to 15 percent. If your tickets are mostly A/B test write-ups, you do not have the AI premium even if you use ChatGPT daily.

Four: Analytics DS is structurally trapped at FAANG in 2026. The Lead → Staff → Senior Staff Analytics DS path that reached comp parity with Applied Scientist at L6 pre-2023 is mostly closed post-LLM. Joining as Analytics DS without knowing this lands you at a year-6 ceiling 30 to 50 percent below an Applied Scientist colleague hired the same year.

The rest of this page is the breakdown. The calculator above runs the (company × level × title × education × location) intersection so all four patterns show up as visible spreads in the result.

§Title arbitrage: the same human earns $343K to $475K at Meta IC5

The Meta IC5 spread is not a special case. It holds at every company in the panel that maintains separate title tracks.

A Senior Applied Scientist at Meta and a Senior Applied Scientist at Google differ by 4 percent on Levels.fyi medians. A Senior Analytics DS at Meta and a Senior Research Scientist at the same Meta differ by 50 percent. Cross-company variation is noise; cross-title variation is signal.

The track choice happens at the offer letter, not after. Promotion across tracks (Analytics → Applied → Research) requires switching ladders, often switching companies, and increasingly requires a PhD. Internal lateral from Analytics DS to Applied Scientist exists at Meta and Google but typically requires shipping a production ML project first, which is gated by the team you already sit on. The window is narrow.

The negotiation move most candidates never make: ask if the title can be re-leveled. "Can this offer be re-tracked from Senior Analytics DS to Senior Applied Scientist, given my background in [production ML project]?" The base/stock package usually moves with the title rather than separately. The cross-title bar chart in the calculator above shows the spread at your specific intersection, which is the data you bring into that conversation.

§AI labs are not paying for AGI, they are paying for the labor-market arbitrage they created

OpenAI did not invent the $1M MTS package because ML research is suddenly worth $1M. It invented the package because in late 2023 it discovered that the marginal Anthropic researcher was leaving for $700K, and that the marginal Google DeepMind researcher was being held in place by a counter-offer of $900K. By Q2 2024 every AI lab understood that the next hire on the wrong side of a comp gap would walk to the lab across the street.

The result is a textbook bidding war on a finite talent pool. The 2026 OpenAI L5 MTS at $1.1M nominal is the equilibrium price of a researcher who could, with one email, get a competing $850K offer from Anthropic. The premium against Google L6 RS at $660K is the price of preventing a Google-DeepMind counter-offer from closing.

Three implications follow for someone holding an AI-lab offer.

The nominal number is not the cash number. OpenAI pays in PPUs, not equity. PPUs cap returns at a fixed share of distributed profit; even if OpenAI's valuation doubles, your PPU payout does not. Anthropic pays in RSUs repriced at each funding round, but those only convert via secondary tender or IPO.

Both labs are now close to public-market exit. OpenAI closed a $122B Series F at $852B post-money in March 2026 and is targeting a Q4 2026 IPO at up to $1T. Anthropic raised $65B in May 2026 at a $965B valuation, eclipsing OpenAI for the first time and targeting a late-2026 IPO. No S-1 filed from either as of May 2026, but the trajectory is set.

The calculator's stock-realization slider defaults to 70 percent for OpenAI and 75 percent for Anthropic, reflecting the Q4 2026 and late-2026 IPO timing of each. At 70 percent, OpenAI L5 nominal $1.1M converts to $770K expected cash. That is above Google L6 Research Scientist at $660K but below the headline $1.1M, and the band depends entirely on whether the IPO prices at or above the last private mark.

The current comp level depends on IPO execution. Both labs spend a large share of revenue on compensation. The current curve holds while private rounds keep clearing at higher valuations; if either lab IPO prices flat or down relative to its last private mark, the implicit equity value resets and the bidding war loses its anchor. Given the H2 2026 IPO timing of both, the next 6-12 months are the moment when "nominal $1.1M offer" either converts to public-market cash at a similar level or resets materially lower.

The career risk-adjustment is real. Researchers who join OpenAI at $1.1M nominal are accepting that their 4-year cash outcome is bounded by the probability of a liquidity event at or above the entry valuation. Researchers who join Google at $660K are accepting that their 4-year cash outcome is bounded by the GOOG share price, a much lower-variance bet. The expected value can be similar; the variance is not. The compare mode in the calculator lets you toggle PPU realization assumptions on the OpenAI offer side-by-side with a Google offer to see where the breakeven lives.

§The PwC "56% AI premium" hides a 25x spread that decides whose career it actually moves

The widely-cited PwC AI Jobs Barometer 2025 number is that workers with AI skills earn a 56 percent wage premium over colleagues in the same role without those skills. The 56 percent is real, sourced from 800M job ads across 9 countries. It is also a mean that conceals a bimodal distribution.

PwC publishes the headline number but does not decompose it by role type. The Levels.fyi DS data tells you why the mean is misleading. At the same FAANG company and level, Applied Scientist and ML Engineer roles (production ML systems work) clear the 56 percent premium against the same-level Analytics DS comp band. Analytics DS roles (dashboards, A/B tests, ad-hoc SQL) do not. The 56 percent is a weighted average across two labor markets sharing a job title, not a premium that everyone in the title captures.

The practical question is which side of the split you sit on right now. The signal is not your job title, it is what fills your week.

You are on the AI-premium side if your tickets are mostly model training pipelines, ranking or recommendation systems, MLOps and infrastructure, production inference debugging, or shipping new model versions to users. The title might say "Data Scientist", "Applied Scientist", or "ML Engineer". Doesn't matter; the work pulls you toward Applied Scientist comp bands.

You are not on the AI-premium side if your tickets are mostly dashboards, cohort analyses, A/B test write-ups, ad-hoc SQL for PMs, or experimentation read-outs. The work pulls you toward Senior Analytics DS and a comp plateau, even if you use LLMs daily to write the SQL faster.

Check the last 10 tickets in your queue. If 7 or more are dashboard / experimentation / ad-hoc analysis, you are on the wrong side of the 56 percent premium, and the pivot needs to happen this quarter or next.

The decision framework for someone 0 to 4 years into a DS career: the cost of being on the wrong side of this distribution compounds. Years 1 and 2 in Analytics DS at FAANG pay similarly to years 1 and 2 in Applied Scientist (mid-$200K range either way). Year 5 differs by $200K+. The switch window is years 1 to 3, ideally inside one company via internal transfer; after year 3, switching ladders typically requires switching companies and a step back in level.

§Analytics DS is the title trap of 2026

The structural problem with Analytics DS at FAANG in 2026 is not the comp at any single level, it is the ceiling. Median Analytics DS at Meta tops out at IC6 around $420K. Median Applied Scientist at Meta IC6 is $581K and Research Scientist with PhD at IC6 is closer to $630K. Past IC6, the Analytics DS title rarely promotes; the next level is typically Director of Analytics, a manager track, which pays from $500K to $750K but is gated by org headcount and headcount-limited promotion windows. Applied Scientist at IC7+ clears $800K to $1M+ as an IC.

Pre-2023, Analytics DS was a respected dual-track career: deep SQL plus statistical experimentation, with promotion through Lead → Staff → Senior Staff Analytics DS that reached comp parity with Applied Scientist at L6.

The post-LLM market broke this. LLMs absorbed the "ad-hoc analytical request" workload that Analytics DS used to specialize in. A Sales VP who needed a cohort analysis used to ping the Analytics DS team and wait two weeks; in 2026 they ask Claude or GPT and get a usable answer in ten minutes. The marginal value of an Analytics DS hour fell, and FAANG comp committees adjusted: widening the gap to Applied Scientist, capping Analytics DS senior bands.

This shows up as the 0.70 to 0.85 title multiplier in the calculator. Amazon (0.70) cut the hardest; Microsoft (0.85) the least. The trend is one-directional: the 2026 multiplier is wider than the 2024 multiplier was, and 2027 will likely be wider again.

The actionable consequence: if you are joining FAANG in an Analytics DS role with intent to stay long-term, the comp ceiling at year 6 will be 30 to 50 percent below an Applied Scientist colleague hired the same year. The switch window is years 1 to 3, through one of three doors:

• Internal lateral to Applied Scientist. Requires shipping a production ML project. Easiest at Meta.
• External switch to a smaller company that bundles titles. Databricks and Stripe collapse Analytics / Applied / ML Engineer into one ladder.
• PhD pivot toward Research Scientist. 5-year detour. See the next section for whether the math works.

Staying in Analytics DS past year 5 increasingly looks like accepting the ceiling, not optimizing within it.

§PhD ROI breaks even only at Research Scientist with 10+ year tenure

The common claim is that a PhD in computer science or statistics pays back in DS compensation. The math does not support this at most title tracks.

A US CS PhD takes 5 to 7 years. During that window, the candidate forgoes industry compensation. The opportunity cost is not the stipend ($35K to $50K/year), it is the foregone industry job: a strong CS graduate could enter FAANG at $200K to $250K mid-level by year 2 of what would have been a PhD program. Over a 5-year PhD horizon, the foregone earnings run $1.0M to $1.4M before tax.

The PhD premium at Applied Scientist is $35K to $50K/year over an MS at the same company × level, mostly in the stock line. Even at the high end, the payback period against $1.2M foregone is 24 to 35 years. Over a typical 20-year post-PhD career, the PhD-vs-MS calculation at Applied Scientist is roughly break-even at best, more often negative.

The math changes at Research Scientist for one reason: PhD is not a premium, it is access. Major FAANG and AI labs do not hire into Research Scientist without a PhD, except for acquired-team transfers and internal promotions from Applied Scientist with a publication track record.

The PhD-vs-no-PhD comparison at Research Scientist is therefore not "$1.2M foregone vs $35K/year premium" but "$1.2M foregone vs job that does not exist for non-PhDs". If the candidate values being on a Research Scientist track at all, the foregone earnings calculation is irrelevant because the alternative path does not lead there.

The break-even framing that does work: PhD pays back if the candidate stays on a Research Scientist track at FAANG or an AI lab for 8+ years post-graduation at IC5+ levels. Below 8 years, the foregone earnings dominate. Above 8 years, the cumulative comp at $500K to $900K (Research Scientist Staff and Principal bands at FAANG, AI lab L5+ MTS) clears the gap. At AI labs the break-even is shorter (3 to 5 years post-PhD) because the comp is higher; at Microsoft Research the break-even is never reached on comp alone because the bands are capped.

The practical decision: choosing between PhD admission and a strong industry MS job, the dollar math favors industry unless you specifically intend to stay on the Research Scientist track long-term at a top lab.

The decision should be made on research interest, intellectual fit, and career flexibility, not on the assumption that PhD pays back automatically. The calculator quantifies the per-year delta at the company × level × title × education state you select, so you can run the math on your own intended path instead of the headline "PhD pays more" claim.

§How to use the calculator against your actual offer

The calculator collapses a 480-cell decision space (8 companies × 5 levels × 4 titles × 3 educations) into one screen. The intended workflow against a real offer:

Step 1: locate the intersection. Match company, level (using the translation table below), title, education, and location. The hero shows median TC at that intersection. Public companies show nominal; private companies (OpenAI, Anthropic, Databricks) show risk-adjusted using the stock-realization slider.

Step 2: check the title arbitrage. The first chart in the result compares all 4 title tracks at your company × level × education. This is the largest visual signal in the result and usually surfaces a $100K to $300K spread. If your offer is below the highest bar in that chart, you have a negotiation lever: ask why the title is X instead of Y.

Step 3: benchmark against the 8-company panel. The cross-company bar (further down) shows your selection vs the other 7 companies at the same level + title + education state. If your offer is within 10 percent of the median, it is a normal offer. If it is more than 20 percent below, the offer is below market and worth a negotiation push. If it is more than 20 percent above, congratulations on the negotiation.

Step 4: stress-test the refresher rate. The TC-by-year line projects 6 years at your chosen refresher percentage. The "comp cliff" between year 4 and year 5 is the most common offer-letter blind spot: a year-1 TC of $500K can decay to $380K by year 5 if the refresher rate is 15 percent. Levels.fyi medians do not surface this because they aggregate year-1 packages.

Step 5: for AI-lab offers, run two PPU scenarios. The defaults assume the IPO lands at or above the last private mark (OpenAI 70 percent, Anthropic 75 percent given Q4/late-2026 IPO targeting). The conservative read (50 percent) models an IPO at a meaningful discount to the last round; the optimistic read (90 percent) models an IPO above the last round. The compare mode lets you A/B these scenarios side-by-side.

§The translation table below

The unified ladder maps to company-specific level and title designators as follows.

Two details matter. At OpenAI and Anthropic, all four title selectors collapse into Member of Technical Staff because the labs do not split the technical ladder into Analytics / Applied / Research sub-tracks. The calculator surfaces this as an editorial mismatch when "Analytics DS" is selected at an AI lab. Microsoft L60-L67 numbering is internal; the unified Senior label maps to L63-64 because L5 in Microsoft's own taxonomy is an intern level.