AI Engineer Salary Guide 2026

Generic salary sites flatten AI engineering into one number. That hides the useful part: 2026 AI pay depends less on the words "AI engineer" and more on the bottleneck you can own.

Selected postings show the spread, but they don't define a market-wide range. Google's Software Engineer, PhD, Early Career, AI/Machine Learning, 2026 Start posting lists a US base range of $147K-$211K plus bonus and equity; minimum qualifications lead with a PhD or equivalent practical experience, so this is not a generic early-career AI eng band.^{[1]Reference 1Software Engineer, PhD, Early Career, AI/Machine Learning, 2026 Starthttps://www.google.com/about/careers/applications/jobs/results/122258040807137990-software-engineer/} OpenAI's Research Engineer page listed $250K-$445K plus equity when checked August 2, 2026, while Anthropic's RL research and safeguards Greenhouse cards list Annual Salary of $500K-$850K and $350K-$500K when checked the same day (equity is not printed on those cards).^{[2]Reference 2Research Engineerhttps://openai.com/careers/research-engineer-san-francisco/[3]Reference 3Research Engineer, Machine Learning (Reinforcement Learning)https://job-boards.greenhouse.io/anthropic/jobs/4613568008[4]Reference 4ML/Research Engineer, Safeguardshttps://job-boards.greenhouse.io/anthropic/jobs/4949336008} These are selected examples, including high-end frontier roles, not estimates for a typical AI engineer.

If you're mapping the role, read the AI engineer role guide first. Salary bands make more sense once you separate applied AI product work, production ML, research engineering, and ML infrastructure.

Benchmark the right market

Three rules keep compensation data honest:

The practical question isn't "What does an AI engineer make?" It's:

What role family am I being evaluated against, and what hard outcome am I expected to own?

Selected compensation anchors

The available sources don't support a reproducible population-wide salary range by level. Use the rows below as source-specific anchors, not percentiles, averages, or promises.

Mid-market gap: Available evidence doesn't include a reproducible mid-market cash band for Series B startups, non-Bay Area remote roles, or non-lab SWE+LLM product engineering. The hard numbers below are ceiling-heavy on purpose (frontier postings + one big-tech early PhD band + one 2024 H-1B MLE median). Don't read the table as a typical-pay distribution.

Snapshot discipline: No fixed Meta Levels.fyi median appears here. Read the live page's displayed median total compensation field before an offer call, because new submissions can change it.

The Meta row above comes from H1BData.info, not directly from the US Department of Labor's OFLC disclosure files. H1BData.info is a third-party aggregator that republishes DOL Labor Condition Application disclosure data; its wage records don't include the stock and bonus reported on total-compensation aggregators. Levels.fyi is a separate user-submitted aggregator and shouldn't be presented as a DOL or employer record.^{[5]Reference 5Meta Platforms MACHINE LEARNING ENGINEER H-1B Salary 2024https://h1bdata.info/index.php?em=META+PLATFORMS+INC&job=MACHINE+LEARNING+ENGINEER&city=&year=2024[7]Reference 7Meta Machine Learning Engineer Salaryhttps://www.levels.fyi/companies/meta/salaries/software-engineer/title/machine-learning-engineer}

Role family matters more than title

Two people can both be called "AI engineer" and sit in different pay markets.

One person builds product workflows on hosted APIs: prompt routing, retrieval wiring, eval dashboards, and user-facing features. That work matters, but it often benchmarks near product software engineering unless the role owns reliability, cost, or quality outcomes.

Another person owns a scarce bottleneck: training throughput, inference latency, model routing, eval methodology, safety classifiers, or agent reliability. That work maps closer to ML systems, research engineering, or infrastructure engineering, and those families carry the highest public bands.

Don't negotiate from the title. Negotiate from the bottleneck you own.

Company tier anchors

The selected frontier-lab postings are high-end public examples, not evidence that every lab role or AI role pays within those bands. OpenAI's Research Engineer role listed $250K-$445K plus equity; Anthropic's RL research and safeguards roles listed Annual Salary $500K-$850K and $350K-$500K when checked August 2, 2026 (equity not on those Greenhouse cards).^{[2]Reference 2Research Engineerhttps://openai.com/careers/research-engineer-san-francisco/[3]Reference 3Research Engineer, Machine Learning (Reinforcement Learning)https://job-boards.greenhouse.io/anthropic/jobs/4613568008[4]Reference 4ML/Research Engineer, Safeguardshttps://job-boards.greenhouse.io/anthropic/jobs/4949336008}

For a big-tech total-compensation check, Levels.fyi exposes live median fields for Google and Meta machine learning engineers and OpenAI software engineers. Read those fields at negotiation time rather than carrying an old number forward. Use the H1BData.info wage rows only as a base-wage cross-check, with the aggregator and filing year stated explicitly.^{[6]Reference 6Google Machine Learning Engineer Salaryhttps://www.levels.fyi/companies/google/salaries/software-engineer/title/machine-learning-engineer[7]Reference 7Meta Machine Learning Engineer Salaryhttps://www.levels.fyi/companies/meta/salaries/software-engineer/title/machine-learning-engineer[8]Reference 8OpenAI Software Engineer Salaryhttps://www.levels.fyi/companies/openai/salaries/software-engineer[5]Reference 5Meta Platforms MACHINE LEARNING ENGINEER H-1B Salary 2024https://h1bdata.info/index.php?em=META+PLATFORMS+INC&job=MACHINE+LEARNING+ENGINEER&city=&year=2024}

Location still matters

Remote work changed compensation, but it didn't make geography disappear.

The cited Google posting offers multiple US working locations and says individual pay depends on skills, experience, and education.^{[1]Reference 1Software Engineer, PhD, Early Career, AI/Machine Learning, 2026 Starthttps://www.google.com/about/careers/applications/jobs/results/122258040807137990-software-engineer/} Anthropic says staff are expected to be in an office at least 25% of the time on the cited roles.^{[3]Reference 3Research Engineer, Machine Learning (Reinforcement Learning)https://job-boards.greenhouse.io/anthropic/jobs/4613568008 [4]Reference 4ML/Research Engineer, Safeguardshttps://job-boards.greenhouse.io/anthropic/jobs/4949336008} The cited OpenAI Research Engineer role is San Francisco-based; hybrid and office policies vary by role and posting.^{[2]Reference 2Research Engineerhttps://openai.com/careers/research-engineer-san-francisco/}

Before using a salary number, ask whether the band is tied to a specific office, state, country, or remote policy. A Bay Area frontier-lab anchor isn't the same thing as a national remote band.

Read the whole offer, not base salary alone

Offer math gets noisy because companies package value differently. Normalize each offer into a four-year view: base and location policy, sign-on timing and clawbacks, equity type and vesting, bonus target, refresher expectations, and liquidity.

Worked example: compare two offers

Both offers below are mid-level applied AI roles. Year one looks close, but the four-year view changes the decision: Offer A is public big tech with $175K base, $110K/year RSUs, 15% target bonus, and an illustrative $40K refresher in years three and four. Offer B is a private AI lab with $190K base, $95K/year options, 10% target bonus, and no announced tender window. Public stock doesn't always win; paper value, vesting, refreshers, and liquidity have to be compared together.

Base isn't total compensation: Public postings show salary bands. Senior offers can differ through equity, sign-on, bonus target, refreshers, liquidity, and location policy.

For private-company equity, ask for the last preferred price, common strike price, latest 409A, tender-window policy, runway after compute commitments, and refresher policy before you treat the grant as cash.

What raises your band

The strongest compensation stories connect engineering work to scarce outcomes. "I integrated an LLM API" is weak. "I reduced p95 inference latency by 28% by changing batching, KV cache policy, and model routing" is strong because it ties work to cost, quality, reliability, or experiment velocity.

Use one compact story shape: bottleneck, constraint, decision, result, business impact. Example: internal FAQ answers took 4.2 seconds at p95 because every request used the highest-cost model with a full document dump; a router, answer cache, and shorter context dropped p95 to 1.1 seconds and cut token cost per conversation by 62%.

What weakens your band

Three signals make a role or candidate look cheaper even when the title sounds impressive: no production ownership, no metrics, and no eval discipline. Fix those by building a complete system with ingestion, chunking, retrieval, reranking, evals, cost tracking, and fallback behavior. Pick one user-value metric, such as grounded answer rate on 100 held-out tasks, and run it after every model or prompt change.

Negotiation checklist

Bring current numbers, but don't let a salary page do all the work. Your negotiating power comes from peer-set fit plus evidence.

• Benchmark yourself against the closest role family: applied AI, ML engineering, ML systems, agent systems, or research engineering.
• Quantify technical impact in business units: latency, GPU cost, token spend, eval lift, incident reduction, adoption, or revenue protected.
• Ask reverse-interview questions about current pain: inference cost, training throughput, eval reliability, data freshness, safety review, or agent failure recovery.
• Negotiate the whole package: base, sign-on, bonus, equity, vesting, refreshers, liquidity, and location adjustment.
• Re-check public data before the offer call. Job postings and salary snapshots move.

Where the numbers come from

The exact employer bands above come from selected postings checked August 2, 2026. The H1B row is a 2024 filing-year snapshot captured August 2, 2026. Levels.fyi supplies live user-submitted total-compensation fields, while H1BData.info supplies a third-party view of DOL LCA base-wage disclosure records. These sources measure different things and change on different schedules, so don't combine them into a single market range without a documented sampling method.