Research Engineer salary in San Francisco & total compensation (2026)

Priced on research taste, not coding speed

At the frontier labs the Research Engineer implements, scales, and refines the systems research runs on — safety tooling, evaluation infrastructure, and training pipelines, including distributed RL infrastructure across GPU clusters. It sits just below the Research Scientist on comp (OpenAI research-scientist total runs $771k–$1.28M+ at L4–L6) but carries a higher floor, because labs recruit it on engineering excellence with demonstrated research instinct. Anthropic's technical-staff bands reach up to $1.38M base.

Research Engineer pay in San Francisco

San Francisco is the price-setting anchor for AI comp: Levels.fyi's Wrapped 2025 put median SWE total comp at $278k in the Bay Area versus $193k in New York — a ~44% gap that has not closed. The metro holds about 35% of US AI engineers, because frontier-lab HQs and NVIDIA all anchor here. The trade-off is the steepest cost of living in the country (~20% above the US average), so COL-adjust before comparing a Bay number to anywhere else.

Research Engineer pay by seniority

Level is the single biggest driver of a Research Engineer offer — a one-level move can change total comp by tens of thousands a year. The bands below are computed from posted pay-transparency ranges in our index:

• Mid (~$300k) — well-scoped work with guidance, ramping to independent delivery.
• Senior (~$300k) — owns features and systems end to end; where most experienced hires land.
• Staff (~$300k) — drives architecture and tradeoffs across teams.
• Principal (~$516k) — sets technical direction; comp here is individualised and equity-dominated.

The four things the interview actually tests

Anthropic's Research Engineer loop screens for research intuition, research taste, calibrated epistemic honesty, and ML-native coding fluency — telling promising directions from dead ends, knowing which problems matter, communicating uncertainty honestly, and implementing model code fluently. Across the labs the bar pivoted from "published papers" toward applied research taste plus building skill in 2024–26, with acceptance under 1%. An RE who can independently own a training-pipeline acquisition is worth far more than one who only executes on existing infra.

AI-native vs traditional pay for a Research Engineer

The same title is priced differently depending on whether the employer is AI-native. Comparing posted bands for Research Engineer roles in our index at AI-native companies against everyone else:

That 53% premium is the single best reason to position yourself as AI-native rather than "a software engineer who also uses AI." The framing alone changes which band a recruiter benchmarks you against.

The equity instrument behind the headline

At this tier the equity type matters as much as the size. OpenAI pays in Profit Participation Units — a synthetic, capped-profit instrument, not stock, valued at vest and at tender (its Oct 2025 tender ran at ~$500B). Anthropic pays RSUs with a tender roughly every 6–12 months and an IPO reportedly targeted as early as October 2026. Notably, the front-loaded vesting sweeping late-stage FAANG is not a lab phenomenon — OpenAI and Anthropic still use a standard 4-year vest with a 1-year cliff, so the "AI exception" is grant size, not schedule.

What a Research Engineer actually does

The Research Engineer builds the systems research runs on: safety tooling, evaluation infrastructure, and training pipelines, and — increasingly — the distributed reinforcement-learning infrastructure across GPU clusters, from clean training abstractions to experiment management. DeepMind frames the role as a software engineer with deep ML understanding — engineering, maths, and ML knowledge in one person. The distinguishing trait versus a pure ML engineer is research taste: an RE who can independently own a training-pipeline acquisition is worth far more than one who only executes on existing infra.

Research Engineer vs Research Scientist vs ML Engineer

Against the Research Scientist, the RE usually sits slightly lower on median comp but carries a higher floor, because it is recruited on engineering excellence plus research instinct rather than publication output — OpenAI research-scientist total runs $771k–$1.28M+. Against the ML Engineer, the RE demands demonstrable research taste (problem selection, a real commitment to alignment/safety), which is the selectivity signal that lifts the band.

Career ladder into research engineering

Two entry paths dominate: PhDs or research assistants in ML/AI, and senior ML engineers who have shown research instinct through publications or serious open-source work. From there the ladder runs Senior RE → Staff RE / Research Scientist → Principal. Across the labs the bar shifted from "published papers" toward applied research taste plus building skill in 2024–26, which widened the door for strong engineers who can demonstrate judgement, not just credentials.

Which companies hire Research Engineers in San Francisco

The most active AI-native hirers for this role in our index right now — a company scaling a function usually means clearer levelling and more room to negotiate. Each links through to its open roles and comp range:

Negotiating a Research Engineer offer

The published lever ranking is consistent across 500+ negotiated AI-lab offers: level is the highest-impact lever — an L4→L5 jump at a frontier lab moves total comp by $150k–$300k a year — and a written competing offer is "the single most reliable" way to move a number. Base salary is the hardest component to shift: both OpenAI and Anthropic run strict bands, so anchor on level and equity, not base. At the RE tier, a competing offer from the other frontier lab is the highest-leverage input — the OpenAI/Anthropic talent war is documented, and each treats the other's offer as the benchmark to beat.

How we calculated these numbers

Every salary figure on this page is computed from pay-transparency bands posted in live job listings in our index — 84 Research Engineer postings in San Francisco currently disclose a band, out of 372 live matching roles. Where a listing posts a range we take the midpoint, and we refresh weekly. These are calibrated ranges, not offers.

Sources

• Anthropic Research Engineer interview 2026 (Sundeep Teki)
• Business Insider — Anthropic salaries revealed (2026)
• Levels.fyi — OpenAI PPU compensation explained
• Levels.fyi — front-loaded vesting