Career Hub

Data Scientist Hub: Product DS, ML DS, and Analytics DS Are Three Different Jobs in 2026

By Blake Crosley · Last verified

The three-track split is the decision this page helps with

Data scientist is a job title that covers three different jobs in 2026. Product DS at Meta, Airbnb, Netflix Studio, and Stripe is SQL, A/B testing, causal inference, and product judgment; the partner is PM and the artifact is a launch decision. ML DS / MLE at Anthropic, OpenAI, Google DeepMind, and Databricks Mosaic is PyTorch, eval design, distributed training, and inference deployment; the partner is research and the artifact is a shipped model. Analytics DS at most non-FAANG tech companies is SQL plus dashboarding plus stakeholder analysis; the partner is the business function and the artifact is the recurring readout. These three jobs share a vocabulary and a degree path. They do not share a ladder, a comp band, or an interview shape. The single largest career mistake at IC3 is interviewing for two of the three at once. The portfolios read as unfocused because the portfolios for the three roles look different. This hub helps you pick the track first, then level up inside it.

Key claims this hub defends

  • The three-track split is real and load-bearing. Product DS, ML DS / MLE, and Analytics DS have different leveling rubrics at the same company, different interview rounds, and different reference patterns. Meta lists Data Scientist and Research Scientist as separate ladders; Anthropic lists Member of Technical Staff with research and engineering subtracks; levels.fyi 2026 self-reports separate Data Scientist from Machine Learning Engineer at most reporting companies.1
  • AI labs pay a tier above FAANG at senior+, and the gap widened in 2025. Anthropic Senior MTS and OpenAI Senior MTS levels.fyi self-reports cluster $700k to $1.4M+; OpenAI peak-vesting cycles have produced reported senior-MTS total comp above $2M in some years. Meta E5 / E6 Data Scientist medians sit $380k to $580k. The caveat: AI-lab equity is private-company equity that clears at a liquidity event, not quarterly. The numbers are directionally right and precision-soft.12
  • The 2026 senior product-DS interview is causal-inference fluent. Senior product DS rounds at Meta IC4+, Airbnb IC4+, Stripe Senior, and Netflix Studio now probe propensity scoring, instrumental variables, difference-in-differences on staggered rollouts, and synthetic control on geo-experiments. The canonical practitioners' textbook is Hernan and Robins, Causal Inference: What If; the working academic body is Susan Athey's NBER causal-ML papers. A/B testing alone clears the junior bar and not the senior bar.7
  • The 2026 frontier-MLE interview compressed toward model-from-scratch in 60 minutes. Anthropic, OpenAI, and Google DeepMind ML rounds now commonly ask the candidate to implement scaled-dot-product attention or a small transformer in PyTorch in a 60-minute coding round, then discuss the eval-set design that would catch a regression. The Vaswani Attention Is All You Need progression through the LLaMA and Qwen architecture papers should be conversational background, not memorization.3
  • The 2020-era deep skills are still load-bearing; what is new is causal inference at scale. Statistics, A/B testing, ML fundamentals, and SQL still gate every level. The 2026 addition is the senior+ expectation that you can extract causal estimates when randomization is impossible (Susan Athey's heterogeneous-treatment-effect papers on arxiv are the reference, and the EconML library at github.com/py-why/EconML is the working toolkit). The deep skill that mattered in 2020 and matters in 2026 is statistics. The deep skill that became table-stakes in 2026 at senior+ is causal inference.7
  • What is not load-bearing in DS hiring at FAANG and AI labs in 2026. Kaggle medals were a signal in 2018 and are not in 2026; the Borisov et al. tabular-deep-learning survey put a quiet end to the Kaggle-as-credential argument. Coursera and DataCamp certificates do not move interview teams. Generic data-analyst experience without a modeling or experimentation layer reads as Analytics DS, not Product DS or ML DS. R-only practitioners are at a real disadvantage at FAANG production ML, where the production language is Python with PyTorch and the data plane is SQL on Spark or BigQuery.5

Product DS, ML DS / MLE, and Analytics DS are three different jobs

The wrong answer to the "which track" question is "all three; I am flexible." Flexibility reads as no portfolio. The right answer is to name the track, name the artifact you have shipped that fits the track, and name the next artifact you will ship to close the remaining gap. The three tracks:

  • Product DS (Meta IC, Airbnb IC, Netflix Studio, Stripe). The job is SQL + A/B testing + causal inference + product judgment. The interview shape is a SQL screen, a stats round, a product-sense round, and a behavioral. The senior round in 2026 probes causal inference. The partner is PM. The artifact is a launch decision that the company would not have made without the analysis.
  • ML DS / MLE (Anthropic MTS, OpenAI MTS, Google DeepMind, Databricks Mosaic, Meta GenAI, Hugging Face). The job is PyTorch + eval design + distributed training + inference deployment. The interview shape is a coding round (often "implement attention from scratch"), an ML system design round, an eval-design round, and a research-engineering depth round. The partner is research. The artifact is a shipped model with a documented eval.
  • Analytics DS (most non-FAANG tech companies, finance, healthcare). The job is SQL + dashboarding + stakeholder analysis. The interview shape is SQL screen + business case + behavioral. The partner is the business function. The artifact is the recurring readout that the function depends on. The leveling rubric is shallower than Product DS or ML DS / MLE; the IC ladder typically tops out at staff equivalent rather than principal.

Two career mistakes the network sees often. First, applying to Product DS roles at Meta and ML DS / MLE roles at Anthropic in the same week with the same resume. The resume reads as either / or to both teams; both pass. Second, taking an Analytics DS job at a FAANG with the plan to "transfer to Product DS." The internal transfer to Product DS is an external-equivalent re-interview at Meta, Airbnb, and Google; the Analytics DS tenure does not count. The honest read: pick the track, then optimize. Trying to keep the option open is the option-value trap.

The 2026 interview bar shifted; the gap is causal inference and model-from-scratch

Two shifts happened between 2022 and 2026. The first is the senior product-DS bar moving from A/B-test fluency to causal-inference fluency. The second is the frontier-MLE bar moving from "you have used the OpenAI API" to "you can implement scaled-dot-product attention from scratch in 60 minutes and design an eval-set that would catch a regression."

The causal-inference shift at senior product DS is the larger change in absolute interview hours. Meta IC4+ product DS rounds in 2026 commonly include a round structured around an observational analysis where randomization is impossible (a pricing change rolled out by market, a policy change applied to a self-selected subgroup, a feature shipped in geos that opted in). The candidate is expected to articulate the identification strategy: propensity scoring with a documented matching procedure, instrumental variables with a defended instrument, difference-in-differences with a parallel-trends sanity check, synthetic control on panel-data settings. The canonical practitioners' textbook is Hernan and Robins, Causal Inference: What If, free PDF; the working academic body is Susan Athey's NBER causal-ML papers and the EconML library. The 2020 toolkit (CUPED for variance reduction, Bonferroni and Benjamini-Hochberg for multiple comparisons, sequential testing for early stopping) is still required and is what Microsoft ExP published at exp-platform.com; the 2026 toolkit adds the observational-causal layer.

The frontier-MLE shift is the larger change in interview prep cost. The "implement attention from scratch in PyTorch" round at Anthropic, OpenAI, and Google DeepMind in 2026 is not a curiosity; it is the modal coding round at senior+. The candidate is expected to write the scaled-dot-product attention forward pass, reason about the masking pattern for a causal LM vs an encoder, and explain the numerical-stability tricks (softmax minus max, the sqrt(d_k) scale factor in the original Attention Is All You Need). The eval-design round is the second compression: write the acceptance criteria for the next training run, defend why the eval-set is contamination-resistant, articulate which metrics catch the failure mode you most fear. The "you have used the API" junior frame is no longer the senior bar.

  • Causal Inference for DS / MLE: DiD, synthetic control, instrumental variables, heterogeneous treatment effects. The senior-product-DS differentiator at Meta, Airbnb, Stripe, and Netflix Studio.
  • Statistics and Experimentation for DS: CUPED, SRM, sequential testing, multiple-comparisons. The Microsoft-ExP-anchored experimentation toolkit; required at every level.
  • LLM and Foundation Models for MLE: PEFT fine-tuning, RAG architectures, eval design, vLLM serving. The frontier-MLE differentiator at Anthropic, OpenAI, and Google DeepMind.
  • ML Fundamentals for DS / MLE: bias-variance, gradient boosting, neural-net optimization, cross-validation. Non-negotiable at every level; the first-round screen-out at AI labs.

AI labs now pay a tier above FAANG at senior+, with caveats the page names

The 2026 senior-DS compensation landscape splits into three bands. Junior FAANG-tier Product DS or MLE: $200k to $290k total comp in the US, mostly cash with a 4-year RSU vesting tail. Senior FAANG- tier (Meta E5 / E6, Google L5 / L6 Data Scientist or MLE): $380k to $580k clustering at IC5, with stock vesting. AI-lab Senior MTS (Anthropic, OpenAI Senior Member of Technical Staff): $700k to $1.4M+ commonly, on heavy private-company equity, with OpenAI peak-vesting cycles having produced reported senior-MTS total comp above $2M in some years per levels.fyi 2026 self-reports. Scale AI senior MLE clusters $500k to $900k; Hugging Face senior MLE is materially lower than the frontier-lab tier but still above median FAANG.

Three caveats this page does not hide. First, the comp data for DS on levels.fyi is roughly 5x to 10x sparser than for SWE, and the AI-lab self-report counts are in the low tens at senior MTS and single digits at principal MTS. The intervals are wide and the precision implied by a six-figure point estimate is false. Second, AI-lab equity is private-company equity that clears at a liquidity event (tender offer, secondary sale, IPO), not quarterly RSU vesting. A $1.4M Anthropic Senior MTS offer and a $580k Meta E6 offer are not directly comparable; one is a concentrated bet on the private-company valuation holding through your vest, the other is liquid quarterly. Third, the AI labs do not publish their compensation bands. The Anthropic careers page and OpenAI careers pages list roles without bands; what is reported on levels.fyi is what individual MTS chose to share. The numbers are directionally right and precision-soft.

The decision the page helps with: whether to optimize for FAANG or AI-lab in your next job-search cycle. If you are early-career (IC3 / E4 / L4 equivalent) and your portfolio is heavier on product DS, FAANG is the better risk-adjusted path; the senior ladder is published, the comp clears quarterly, and the IC5 / E5 number is large enough that the AI-lab tier difference is not decisive after tax and after equity-discount. If you are mid-to- senior with ML-research-engineering portfolio (a fine-tuned model on Hugging Face Hub with a documented eval, a NeurIPS / ICML workshop paper, a deployed eval framework), the AI-lab tier is where the marginal return on portfolio is highest, and the comp gap is large enough to absorb the private-equity-risk discount.

  • MTS at Anthropic: the careers-page-anchored guide to Member of Technical Staff at Anthropic; research and engineering subtracks, the comp tier, the interview shape.
  • MTS at OpenAI: the OpenAI MTS guide; comp tier, equity structure, the peak-vesting reality.
  • Data Scientist / MLE at Meta: the Meta DS and Production ML guide; E5 / E6 leveling, the product-DS round shape, the GenAI track.
  • Data Scientist / MLE at Stripe: the Stripe DS guide; fraud-detection ML, experimentation at scale, the senior bar.
  • Data Scientist at Airbnb: the Airbnb DS guide; the experimentation platform, the causal-inference-heavy senior round.

Landing the first DS / MLE role: portfolio over credentials

Junior Product DS at FAANG-tier (Meta E3, Airbnb IC2, Netflix Studio, Stripe) typically requires 0 to 3 years of experience or a portfolio with the equivalent depth. The portfolio piece that gets the interview: a shipped analysis with named lift on a defined metric, written up at the level a PM would read. Kaggle medals were the signal in 2018 and are not in 2026. The portfolio piece that does not move the round: certificates from Coursera, DataCamp, or DataQuest.

Junior MLE at AI-lab tier (Anthropic MTS-entry, OpenAI MTS-entry, Google DeepMind research engineer, Hugging Face MLE) typically requires either a relevant graduate degree (PhD strongly preferred for research-track, MS acceptable for engineering-track) or a portfolio with research-engineering depth. The three artifacts that move the needle: a fine-tuned open model on Hugging Face Hub with a documented eval, a NeurIPS / ICML / ACL workshop paper (co-authored is fine), a deployed eval framework on GitHub with non-trivial downstream use. Total comp at junior FAANG-tier runs $200k to $290k; AI-lab entry sits $250k to $500k+ depending on equity structure per levels.fyi 2026 data.1

Making senior is the central plateau and the longest tenure

Mid (3 to 5 yrs) and senior (5 to 8 yrs) is the central plateau for most DS / MLE careers. Senior is the level where the company expects you to own the analytical or modeling work end-to-end. For Product DS at IC4+: own the experimentation strategy for a product surface, run the causal-inference analyses when randomization is impossible, partner with PM on the launch decision and own the after-the-fact accountability. For MLE at IC4+: own ML system design end-to-end (data pipeline, training loop, eval pipeline, deployment surface, monitoring), LLM and foundation-model fluency at production scale (PEFT fine-tuning via the Hugging Face peft library, eval methodology, inference deployment via vLLM), partnership with research and engineering, and mentorship of juniors. Senior FAANG-tier DS / MLE 2026 total comp clusters $380k to $580k at L5 / E5 / IC5 on levels.fyi.1

Staff and principal are the scope step, not the seniority step

The senior IC track in DS / ML engineering runs deep. Staff (8 to 12 yrs) goes to Senior Staff (10 to 15 yrs) to Principal (12 to 20+ yrs) to Distinguished MTS at AI labs, which carries some of the most senior IC titles in the industry. The transition from senior to staff is not "more senior;" it is "wider scope." Staff scope expands beyond a domain to multi-team or org-level work: a foundation-model strategy across multiple recommendation surfaces, an eval platform that other teams adopt, a distributed-training infrastructure used by the whole ML org, a causal-inference framework that becomes the company-wide standard for observational analyses.

Total comp at staff+ commonly clears $620k at FAANG-tier; at principal it commonly exceeds $1M; AI-lab principal MTS peak- vesting cycles can exceed $5M in reported total comp, with OpenAI principal MTS at the top of the IC market in 2026. The caveats from the AI-lab section apply with extra weight at this tier: the principal self-report count on levels.fyi is in the single digits, the equity is private, the precision is soft.1

Each company guide covers what is verifiably published, not what is inferred

The company pages cover what each company has actually published about hiring: how levels map to titles, what is known about the interview process from first-party careers pages, compensation data from levels.fyi with the sparsity caveats named, and the org / culture artifacts the company has chosen to share publicly. Where companies do not publish their internal interview rubrics, the pages name the gap rather than fabricate authority. AI labs (Anthropic, OpenAI, Scale AI, Hugging Face) are the tier-above- FAANG comp companies in 2026; FAANG and FAANG-adjacent (Meta, Google, Netflix, Airbnb, Stripe) and ML-platform companies (Databricks) round out the published hiring landscape.

The 2020-era deep skills are still load-bearing; what is new is causal inference at scale

The DS / ML deep-skill bar in 2026 is the 2020 bar plus one addition. Statistics and A/B-test methodology (the Microsoft ExP toolkit at exp-platform.com; CUPED, SRM detection, sequential testing, multiple comparisons) is gating at every level. ML fundamentals (bias-variance, gradient boosting, neural-net optimization, cross-validation including the scikit-learn TimeSeriesSplit pattern for temporal data) gates every MLE round. LLM and foundation-model fluency (PEFT fine-tuning, RAG, eval design, inference deployment) is the mid+ MLE differentiator.

The 2026 addition is causal inference at scale. Senior product DS interviews at Meta, Airbnb, Netflix Studio, and Stripe now probe observational-causal methodology in a way they did not in 2020. The toolkit: propensity scoring on observational data, instrumental variables when randomization is impossible, difference-in-differences on staggered rollouts, synthetic control on panel-data settings. The canonical practitioners' textbook is Hernan and Robins; the working academic body is Susan Athey's NBER causal-ML papers; the working library is EconML at github.com/py-why/EconML. The wrong answer at the senior product-DS round is "we would A/B test it;" the right answer names the identification strategy for the case where A/B testing is not an option.

Frequently asked questions

How do I pick between Product DS and ML DS / MLE if I'm a strong programmer?
The deciding question is what your portfolio already shows. If you have shipped A/B tests with named lift, written analyses PMs cited in launch decisions, and your strongest references are PMs, that is Product DS. If you have shipped a fine-tuned model with documented eval, written training code that runs at scale, and your strongest references are ML engineers, that is ML DS / MLE. Strong programmer alone is not a tiebreaker; the ML DS / MLE track at AI labs and FAANG production ML wants strong programmers, but the Product DS track wants programmers who also have product judgment. The interview-prep cost is real: Product DS prep is SQL plus causal inference plus product sense; MLE prep is PyTorch plus eval design plus ML system design. Trying to prep both in one job-search cycle is the failure mode the network sees most often.
Is the AI-lab tier real or is FAANG comp catching up?
It is real at senior+ as of 2026 and the gap has widened, not narrowed, since 2024. Anthropic and OpenAI senior MTS levels.fyi self-reports cluster materially above the Meta E5 / E6 and Google L5 / L6 medians at the same scope, with OpenAI peak-vesting cycles producing reported total comp above $2M for senior MTS in some years. The caveat the page does not hide: AI-lab equity is private-company equity. It clears at a liquidity event (tender offer, secondary, IPO). FAANG RSU value clears every vest. A $1.4M Anthropic Senior MTS offer and a $580k Meta E6 offer are not directly comparable; one is a concentrated bet on the private-company valuation holding through your vest, the other is liquid quarterly. Treat the AI-lab number as a range whose lower bound is below the FAANG number if the tender prices fall.
What's the smallest credential change that opens AI-lab interviews?
One published artifact the interview team can read in fifteen minutes. The three artifacts that move the needle: a fine-tuned open model on Hugging Face Hub with a documented eval set and a README explaining what failed and why; a NeurIPS / ICML / ACL workshop paper, co-authored is fine; a deployed eval framework on GitHub with non-trivial use (stars are weak signal, downstream forks and citations are strong signal). The credentials that do not open AI-lab interviews: Kaggle medals (was the signal in 2018, is not in 2026), Coursera certificates, generic data-analyst experience without a modeling layer, R-only practitioners at the production-ML tier. The honest framing: AI labs hire on demonstrated research-engineering taste, and the artifact is the proof.
Do FAANG data-scientist interviews still over-index on SQL and stats?
Junior yes, senior no. Junior Product DS at Meta, Airbnb, Netflix Studio in 2026 still runs a heavy SQL screen plus a stats round; the bar has risen (joins on 5+ tables, window functions, the candidate has to argue why their query is correct, not just produce output), but the shape has not. Senior product DS has shifted. The 2026 senior round at Meta IC4+, Airbnb IC4+, and Stripe Senior probes causal-inference fluency: propensity scoring on observational data, instrumental variables when randomization is impossible, difference-in-differences on staggered rollouts, synthetic control on geo-experiments. SQL plus A/B-test fluency clears the junior bar; the senior bar wants A/B-test fluency plus the toolkit for the cases where you cannot run an A/B test.
How sparse is the comp data for DS compared to SWE?
Materially sparser, and the page does not pretend otherwise. levels.fyi has roughly 5x to 10x more SWE self-reports than data-scientist self-reports at most companies; the median comp at L5 / E5 Data Scientist is computed off a few hundred data points at FAANG, low-tens at Anthropic and OpenAI, and AI-lab principal MTS is computed off self-reports in the single digits. Treat all AI-lab numbers as wide intervals. The published Anthropic and OpenAI careers pages do not list comp bands; what is reported is what individual MTS have chosen to share. The honest read: the numbers are directionally right (AI labs pay above FAANG; FAANG senior+ pays in the $400k to $900k range with stock vesting) but the precision implied by a six-figure point estimate is false.
Should a strong Analytics DS try to move into ML DS at the same company?
Possible but uphill. The internal transfer path that works: the Analytics DS finds a project where the analytical question is also a modeling question (a churn analysis that becomes a churn predictor; a pricing analysis that becomes a pricing model), ships it with an MLE partner, and uses that as the portfolio piece for the transfer. The path that does not work: requesting a transfer because Analytics DS comp lags ML DS comp. The leveling rubrics at Meta, Google, and Airbnb treat the two ladders as parallel, not promotable; you re-interview for the new track, and the bar is the same as for an external hire. The honest read: if you are doing Analytics DS at a company with strong ML, the credential that opens the transfer interview is a shipped modeling project, not a tenure argument.

Sources

  1. levels.fyi; Data Scientist and ML Engineer compensation (2026 self-reports). Self-reported total compensation by level across FAANG-tier and AI-lab companies; sparsity caveat applies (DS data is roughly 5x to 10x sparser than SWE; AI-lab senior MTS counts are in the low tens, principal in the single digits).
  2. Anthropic Careers; Member of Technical Staff postings. Research and engineering subtracks; entry MTS through Distinguished MTS leveling. The page does not list comp bands; the AI-lab tier numbers come from individual MTS self-reports on levels.fyi.
  3. Vaswani et al.; Attention Is All You Need (NeurIPS 2017). The foundational transformer paper. The 2026 frontier-MLE round commonly asks the candidate to implement scaled-dot-product attention from scratch in 60 minutes.
  4. Chip Huyen; Designing Machine Learning Systems and ML Interviews Book. The canonical mid-to-senior production-ML reference; the ML-interview-book covers the interview-round structure at FAANG and AI labs.
  5. Hello Interview; Understanding FAANG Job Levels. Cross-company DS / MLE leveling reference; the Meta E5 / E6, Google L5 / L6, Anthropic MTS, OpenAI MTS leveling map.
  6. OpenAI Evals; open-source eval framework. The canonical eval-design reference at AI labs; paired with EleutherAI's open-source lm-evaluation library, the two frameworks the eval-design round expects the candidate to know.
  7. Hernan and Robins; Causal Inference: What If (free PDF). The canonical practitioners' textbook for causal-inference methodology; the 2026 senior product-DS round is the page where this material binds.
  8. EleutherAI; open-source LLM evaluation framework on GitHub. The canonical community-maintained LLM eval framework; the open-source counterpart to OpenAI Evals at the AI-lab eval-design round.
  9. Deng, Xu, Kohavi, Walker; CUPED (Microsoft ExP, 2013). The canonical variance-reduction paper; the Microsoft ExP body of work at exp-platform.com is the experimentation-toolkit reference for every level.

Resources for data scientists and ML engineers

  • Software Engineer: the broader software-engineering hub. ML engineers live in the seam; this is the discipline you partner with on production-ML systems.
  • Product Manager: your product partner. Brief them well, ground their roadmap in evidence, push back when an experiment is unprestructured.
  • Product Designer: your design partner. UX research overlap, experiment-design partnership, design-of-experiments craft.
  • iOS Engineer: if your team ships on-device ML (Core ML, Foundation Models in iOS 26), the iOS engineering depth is here.
  • Engineering Manager: the management track if you're considering the move from senior IC ML work.
  • Design Manager: your design-leadership partner. Design decisions get evaluated against experiment data; you're the discipline they partner with.
  • Frontend Engineer: the engineers who ship the dashboards, A/B test surfaces, and product-analytics interfaces that DS / ML teams consume daily.
  • Backend Engineer: the data partner. Backend engineers ship the data-platform infra (Kafka, Airflow, dbt, warehouses) that DS teams build on.
  • UX Researcher: the qualitative counterpart. UXR + DS partnership is the modern evidence stack: signal, story, and significance combined.
  • Data Engineer: the infrastructure partner. DS uses the data infra DEs build (Spark, dbt, warehouses, streaming).
  • Android Engineer: mobile analytics ingestion partner. Android event streams flow through DS pipelines.
  • DevOps / SRE Engineer: the production-platform partner. ML systems run on the infra SREs maintain; experiment platforms, metrics pipelines, and model-serving rely on SRE.
  • Security Engineer: ML-security and AI-safety partner. Adversarial robustness, prompt-injection defense, NIST AI RMF compliance, and PII-in-training-data governance live at the DS / Security seam.

Cross-cutting career-strategy guides

Topic-style guides that apply across every role track, from referral to onboarding. Pair the role-specific content above with these guides for the parts of the job-search arc that are not role-specific: