Data science benchmarking has a problem that software engineering does not. The title covers at least three different jobs, and the pay gap between them is now wider than the gap between junior and senior within any one of them.
Before you look at a single number, decide which of these you are actually hiring:
- The analyst. SQL, dashboards, stakeholder reporting. Real value, lowest band.
- The scientist. Modelling, experimentation, causal inference. The classic definition.
- The ML engineer. Production systems, pipelines, deployment. Currently the most expensive of the three by a clear margin.
A benchmark that averages all three tells you nothing about any of them.
Germany, 2026
| Role and level | Typical base range |
|---|---|
| Data scientist, junior | €45,000 to €60,000 |
| Data scientist, 3 to 5 years | €70,000 to €90,000 |
| Data scientist, overall band | €50,000 to €92,000 |
| ML engineer, entry | around €70,000 |
| ML engineer, mid | €58,000 to €82,500 |
| ML engineer, 8+ years | up to €113,740 |
The published averages for machine learning engineers range from €70,000 on Glassdoor to €100,264 on SalaryExpert. That is a 43% spread, and it is not measurement error. It reflects a genuine split between employers who have production machine learning and employers who have a data team with ambitions.
United States, 2026
| Measure | Figure |
|---|---|
| Median total compensation | $176,000 to $180,000 |
| 75th percentile | $245,000 |
| 90th percentile | $330,000 |
By employer, levels.fyi medians vary far more than the headline suggests: Block $307,000, Indeed $306,000, Meta $297,000, Google $290,000, Microsoft $255,000, Glassdoor $190,000, IBM $164,000.
The spread from IBM to Block is nearly double, for the same job title, in the same country, in the same year. If you are competing for the same people as the top of that list, the market median is not your benchmark. The top quartile is.
The ML premium is the story
The single most useful thing in this guide: production machine learning experience is now worth more than seniority.
An engineer with four years who has shipped and maintained a model serving live traffic will frequently out-earn a data scientist with eight years of analysis and experimentation work. Ten years ago that would have been backwards. The reason is supply. Plenty of people can build a model in a notebook. The population who have owned one in production, with monitoring, retraining and a rollback plan, is small and did not grow as fast as demand.
If your specification says "data scientist" but your first year of work is making the pipelines reliable and getting a model deployed, you are hiring an ML engineer and should benchmark accordingly. This is the most common and most expensive briefing error we see in this discipline.
How to benchmark this properly
Write the first year, not the title. List the five things this person will actually deliver in twelve months. If three of them are engineering, benchmark as engineering.
Separate the analyst band out. If the role is genuinely reporting and stakeholder analysis, you are competing in a much deeper and cheaper market, and paying ML rates for it wastes budget you need elsewhere.
Check whether your data is ready. A large proportion of failed data science hires are not hiring failures. The person arrives, finds no reliable pipeline, spends a year doing data engineering they did not sign up for, and leaves. If that is the actual first year, hire the data engineer first. It is cheaper and it works.
Assume a counteroffer. In this segment specifically, salary inflation has run ahead of the wider engineering market and incumbent employers respond aggressively. Prepare a non-financial answer before the offer goes out.
What we are seeing in live searches
The clearest pattern of the year is that clients who fix the definition before the search runs fill the role, and clients who do not, do not. We turn down data science briefs where the specification and the first-year workload contradict each other, because taking them is how a search burns three months and ends where it started.
The second pattern: German employers are increasingly losing candidates to fully remote US contracts rather than to local competitors. If you are hiring in Berlin or Munich and benchmarking only against German employers, you are benchmarking against the wrong market.
We recruit data and machine learning teams across the UK, DACH and the US. Our data science and analytics recruitment page sets out how we scope these briefs, and if the definition question above sounds familiar, tell us what you are trying to hire and we will tell you honestly which role it is.
Sources: levels.fyi, Glassdoor, SalaryExpert. Figures are gross annual and were current at the date of publication. Treat them as a starting point for a conversation, not as a determination.