The most-cited AI lab salary comparison in 2026 says Anthropic pays $665K median total comp vs OpenAI at $555K - and names Anthropic the winner. On our analysis of 300+ verified Levels.fyi data points, that headline is backwards at the levels where most engineers reading this are actually competing. An OpenAI L5 earns $829K-$871K in total compensation per year (Levels.fyi 2026). An Anthropic Lead Software Engineer - roughly equivalent seniority - earns $759K. OpenAI wins by $70K-$110K per year at that level, and the gap grows further up the ladder. The equity mechanics that deliver - or withhold - that difference are the real story, and most comp articles never get there.
Before getting into the numbers, the comparison is harder than it looks because all three companies measure compensation differently. Anthropic does not publish internal level names publicly. OpenAI uses L3-L7. Meta uses E3-E9. Levels.fyi maps self-reported submissions to inferred bands, with sample sizes at Anthropic numbering in the dozens and at Meta numbering in the thousands. A 'senior engineer' at all three companies might be doing comparable work but sitting at different points in each company's hierarchy, and each firm weights base vs. equity very differently.
Base salary at each lab, by level
Start with the base salary, because it is the only part of TC you can spend the day your paycheck deposits. Anthropic Senior SWEs earn approximately $318K in base; Lead SWEs earn approximately $329K (Levels.fyi 2026). These are high by any standard - roughly 2.4 times what a senior engineer earns at a regional employer. But in the San Francisco AI talent market, they sit around the median for this peer group. OpenAI L4 engineers earn $278K-$340K base; L5 engineers earn $315K-$347K base (Levels.fyi 2026). Meta ML Engineers at E5 earn approximately $225K base with a 15% bonus on top, bringing guaranteed cash to roughly $259K.
Strip out equity and the comparison narrows dramatically. On base salary plus bonus alone - the part of TC you can actually spend - Anthropic Lead SWEs and OpenAI L5s earn roughly the same: $330K-$350K. Meta E5s trail by about $75K in guaranteed cash. The entire advantage of Anthropic and OpenAI over Meta sits in equity. Whether that equity is worth what the offer letter claims depends on questions that most candidates do not ask during recruiting.
| Feature | Anthropic | OpenAI |
|---|---|---|
| Mid-level TC (L4/Senior equiv.) | ~$563K (Senior SWE) | ~$646K (L4) |
| Senior TC (L5/Lead equiv.) | ~$759K (Lead SWE) | ~$871K (L5) |
| Staff TC (L6 equiv.) | ~$850K+ | ~$1.23M+ (L6) |
| Research Scientist median | ~$746K | ~$771K (L4 equiv.) |
| Equity cliff | 1-year cliff, forfeit all before Month 12 | No cliff (Day 1 vesting since Dec 2025) |
| Equity type | RSUs (private company, illiquid) | RSUs (converted from PPUs Jan 2026, still illiquid) |
| Liquidity | Periodic tender offers, company discretion | Periodic tender offers, company discretion |
One detail the comparison table cannot capture: Anthropic's higher blended median ($665K vs OpenAI's $555K) is a sample artifact. Anthropic has a smaller, more senior-skewed workforce. OpenAI has more L2-L3 engineers pulling the average down. When you compare by specific level, OpenAI wins at every band from L4 up. The blended-median comparison that circulates in most AI salary articles is misleading for any individual engineer trying to evaluate an actual offer.
The equity structure is where the real comparison happens
OpenAI's equity was not always a standard RSU. From 2019 through 2025, OpenAI issued Profit Participation Units (PPUs) - an instrument tied to future profits rather than company equity, with a 10x cap on each grant's value (Pin.com 2026). OpenAI converted PPUs to RSUs in January 2026 as part of its restructuring into a for-profit company. The conversion clarified the structure, but engineers who joined between 2022 and mid-2025 held PPUs during years when OpenAI was losing approximately $5 billion annually against $13 billion in revenue. Those grants vested, but their realized value depended entirely on secondary tender offers at company-set prices - not a public market.
Anthropic's structure is simpler but has a cliff that OpenAI eliminated. Anthropic grants RSUs on a 4-year vest with a 1-year cliff: 25% vests at Month 12, then monthly thereafter (jobsbyculture.com 2026). OpenAI removed its cliff entirely in December 2025, moving to Day 1 vesting - a direct response to xAI doing the same in summer 2025 and to retention pressure across the frontier-lab talent market (Fortune 2025). The difference matters more than it appears in a salary comparison: an engineer who leaves Anthropic at Month 11 - by choice or due to a layoff - forfeits all equity. At OpenAI since December 2025, an engineer who leaves at Month 11 keeps 11 months of vested equity.
The illiquidity problem affects both companies in structure but differs in degree. In February 2026, Anthropic ran a tender offer at a $350 billion valuation where investors hoped to buy $5-6 billion in employee shares. Most employees held (Bloomberg 2026). Their reasoning: they expected a future IPO to price higher than the tender's implied value. That is a rational bet if Anthropic continues toward its reported $30 billion ARR trajectory. But it is also an employee choosing to leave hundreds of thousands of dollars on the table against an uncertain future payout. Anthropic's current valuation sits at approximately $380 billion (Series G, 2026); secondary market prices have implied a $1 trillion valuation. The gap between those two numbers tells you exactly how speculative the equity component of any Anthropic offer is.
OpenAI pays $70K-$110K more per year at equivalent senior levels (Levels.fyi 2026), has no equity cliff since December 2025, and its post-conversion RSUs are structurally cleaner than the PPU instrument they replaced. The exception is research scientists: Anthropic's environment has an 80% two-year retention rate - the highest among frontier labs - and its research scientist median TC of $746K is within striking distance of OpenAI's $771K (HeroHunt.ai 2026). If you are doing frontier AI research, the choice is close and Anthropic's culture is a genuine differentiating factor. If you are doing production ML engineering, OpenAI wins on the numbers and on the equity structure. Meta belongs in this conversation only at E6 and above, where its $786K TC in fully liquid public stock competes on risk-adjusted terms with either frontier lab.
What most comp transparency articles miss
The standard Levels.fyi comparison reports base, bonus, and equity grant at time of hire. That analysis misses at least four structural factors that meaningfully change realized compensation over four years. The first: mark-to-market equity risk. Levels.fyi values grants at the most recent tender or secondary market price. Anthropic raised at $380 billion in 2026; OpenAI raised at approximately $300 billion in recent rounds. Those valuations are the denominators used to calculate how many RSUs your dollar grant translates into. The share count is fixed at grant; if either company IPOs below that price - a real possibility given that 70% of AI startups saw valuation corrections in 2025 (Pin.com 2026) - your offer letter TC will have overstated what you actually receive.
The second gap is role commingling. Levels.fyi aggregates software engineers, research engineers, and research scientists into the same dataset without clean separation. At Anthropic, a Research Scientist L6 earns roughly 1.5 to 2 times what a Software Engineer at the same nominal level earns (Blind 2026). When an article cites 'Anthropic median $665K,' that blends roles in a way that inflates the number for pure engineers and understates it for researchers. The role-type spread matters more at Anthropic than at OpenAI or Meta because Anthropic's ratio of research to engineering headcount is higher than the industry norm.
The third gap is mission discount. Anthropic retains 80% of two-year hires while paying cash-equivalent amounts that, by the numbers, should face more attrition (HeroHunt.ai 2026). A meaningful share of Anthropic's workforce is pricing in the non-financial value of working on safety-aligned AI. Anthropic recruits on this deliberately. The practical effect: Anthropic offers lower effective cash value than its headline TC suggests, because part of the workforce is voluntarily accepting what wealth advisors call a mission discount. If you are not in that segment, Anthropic's TC is worth less than the offer letter states. The fourth gap is tax exposure: Anthropic RSUs are taxed as ordinary income at vest, and the withholding rate often leaves employees owing six-figure tax bills in April - a reduction in realized comp that no transparency site captures (Augustus Wealth 2026).
“Equity now comprises 60 to 70 percent of total compensation at frontier labs, up from 35 to 45 percent two years ago. That means the gap between headline total comp and what you can actually spend has never been wider.”
Where Meta AI fits in this picture
Meta is not a direct talent competitor to Anthropic and OpenAI in most conversations, but it belongs in this comparison because it answers a question neither frontier lab can match: what does the pay look like in fully liquid form? A Meta E5 ML Engineer earns $472K in total comp: $225K base, $222K in RSUs, and $26K in bonus (Levels.fyi 2026). Those RSUs are Meta public stock, liquid the day they vest. An E6 earns $786K in the same structure. There is no cliff, no tender window, no secondary market required. You sell vested shares at the market price the day they vest. In May 2026, Meta also announced 8,000 layoffs while simultaneously redeploying 7,000 employees into AI-focused roles - a signal that Meta is consolidating around AI engineering rather than retreating from it (NBC News 2026).
The real comparison for a senior engineer is not $871K OpenAI L5 vs $786K Meta E6. It is approximately $340K in guaranteed cash (OpenAI base plus bonus) vs $269K (Meta base plus bonus), plus $530K in OpenAI RSUs at current valuation vs $469K in Meta public stock. The OpenAI RSUs require trusting an IPO timeline, a post-IPO lockup, and a company valuation that holds. Meta RSUs convert to cash quarterly. Financial advisors working specifically with AI lab employees commonly apply a 15-25% liquidity discount to private-company RSUs depending on estimated time to IPO (Augustus Wealth 2026). At a 20% discount, the OpenAI L5's $530K in RSUs becomes $424K in expected-value terms - putting the total risk-adjusted offer at roughly $764K, about $22K above Meta E6. That is a much closer race than the headline numbers suggest.
- OpenAI: no equity cliff since December 2025, Day 1 vesting, L5 TC highest of the three at $871K, RSU structure cleaner post-PPU conversion
- Anthropic: best research culture of the three (80% two-year retention), research scientist median $746K competitive with OpenAI, equity upside highest if IPO prices above $380 billion
- Meta: fully liquid public RSUs vesting quarterly from day one after cliff, no private-company valuation risk, E6 TC at $786K competes on risk-adjusted basis with OpenAI L5
- OpenAI: PPU-to-RSU conversion created years of structural uncertainty; prior PPU cap of 10x grant value limited senior-level upside; equity grant levels reportedly pulled back from 2022-2024 peaks
- Anthropic: 1-year cliff forfeits all equity before Month 12; tender offers are infrequent and company-controlled; base salary runs $20K-$30K below OpenAI at equivalent levels
- Meta: E5 TC at $472K materially trails both frontier labs; annual refresher grants cut by 10%+ in 2025 (FAANG Fire 2025); no exposure to frontier-lab equity upside if either Anthropic or OpenAI generates a high-multiple exit
How to think about this for your next career move
The number to run is the four-year break-even, not the annual TC. Take the difference between each offer's expected four-year payout. Apply a liquidity discount to the private equity: most financial advisors working with AI lab employees use 15-25% for Anthropic and OpenAI RSUs depending on their estimated time to IPO (Augustus Wealth 2026). Factor in the cliff: if there is any realistic scenario where you leave Anthropic in Year 1 - health, family, role mismatch, reorg - the expected value of that equity at Month 11 is zero, not the grant amount. Then factor in tax exposure: RSUs vest as ordinary income, and the withholding gap can cost you $50K-$150K in year-end tax bills on top of what you already paid via withholding.
The fastest path into these compensation tiers is demonstrated production ML experience rather than credentials alone. For engineers currently building toward senior AI roles, the /certifications/google-ml-engineer certification and the /certifications/aws-ml-specialty are the two credentials most cited by frontier lab hiring managers as signals of production-readiness. Neither gets you to L5 alone; both help you speak the right technical vocabulary in the interview process. We cover the full AI/ML engineering career path at /careers/ai-ml-engineer and the role itself in depth at /learn/what-does-an-ai-ml-engineer-do-2026. For adjacent roles that frontier labs also hire into, the /learn/what-does-a-devops-engineer-do-2026 guide covers the infrastructure-adjacent skills that appear frequently in frontier lab job descriptions.
For engineers not yet at the frontier-lab tier, the broader US AI/ML engineer market pays a median $162K in base salary (Glassdoor 2026) or $187K-$270K in total comp depending on company tier - significantly less than the frontier-lab numbers but growing fast. AI Engineer was LinkedIn's number one fastest-growing US job title in 2026, with postings up 143% year-over-year and a global supply-demand gap of 3.2 qualified candidates per open role (LinkedIn Workforce Report 2026). The AI premium is real at senior levels - 14% above comparable non-AI senior engineers and 18% above non-AI staff engineers (Levels.fyi 2026) - but only 6% at entry level, and entry-level AI hiring at top tech firms fell 25% from 2023 to 2024. The money is at the top, and getting there requires specific deployment experience that Coursera's machine learning specializations (around $200 for 6 months of access) and LinkedIn Learning's LLM deployment tracks are the most cited starting points for engineers building that skillset (LinkedIn Workforce Report 2026).
Does OpenAI or Anthropic pay more for AI engineers in 2026?+
At equivalent senior levels, OpenAI pays more: an L5 earns $829K-$871K TC vs $759K for an Anthropic Lead SWE (Levels.fyi 2026). Anthropic's higher blended median ($665K vs $555K for OpenAI) is a sample artifact - Anthropic has fewer junior roles pulling the average down. For research scientists specifically, the gap narrows to roughly $25K and Anthropic's research culture becomes the stronger differentiator.
What was an OpenAI PPU and how does it differ from an RSU?+
Profit Participation Units were OpenAI's proprietary equity instrument from 2019 to 2025. They were tied to future profits rather than company equity value, with a 10x cap on individual grant value. OpenAI converted all PPUs to standard RSUs in January 2026 as part of its restructuring into a for-profit company. Current offers at OpenAI are RSU-based. Engineers who joined before 2025 may still hold PPUs alongside their post-conversion RSUs.
What is the 1-year equity cliff at Anthropic and why does it matter?+
Anthropic grants RSUs on a 4-year vest with a 1-year cliff: nothing vests in the first 12 months, then 25% vests at Month 12, followed by monthly vesting thereafter (jobsbyculture.com 2026). If you leave before Month 12 - voluntarily or via layoff - you receive zero equity regardless of how close to the cliff date you were. OpenAI eliminated its cliff entirely in December 2025, moving to Day 1 vesting. For any engineer with a realistic chance of leaving in Year 1, the expected equity value at Anthropic should be discounted substantially.
Is Meta competitive with Anthropic and OpenAI for AI engineer compensation?+
At E5, Meta pays $472K vs $871K at OpenAI L5 - a large gap. But Meta's RSUs are fully liquid public stock, vesting quarterly. After applying a 15-25% liquidity discount to OpenAI's private RSUs (Augustus Wealth 2026), the effective gap narrows from $399K to roughly $240K-$320K over four years. Meta E6 at $786K in liquid stock competes more directly with OpenAI L5 on risk-adjusted terms. For engineers who value compensation certainty over upside, Meta is a rational choice over either frontier lab.
How reliable are Levels.fyi salary numbers for these companies?+
Meta has the most reliable data (public company, large sample size of thousands of submissions). Anthropic and OpenAI data is thin - fewer than 100 verified Levels.fyi submissions each - which creates upward bias from self-selection. Levels.fyi values private-company equity at the most recent tender or secondary market price, not an IPO price. Treat Levels.fyi frontier lab figures as directional with a 15-25% haircut on equity for illiquidity risk and small-sample uncertainty.
What skills do I need to be competitive for these AI engineering roles?+
Based on 2025-2026 job postings (LinkedIn Workforce Report 2026), frontier labs hire for PyTorch proficiency, demonstrated LLM fine-tuning experience, inference optimization knowledge (vLLM, quantization, batching), and RAG pipeline implementation. Research roles additionally require publications or equivalent independent research output. The /certifications/google-ml-engineer and /certifications/aws-ml-specialty certifications are the most cited entry-point credentials for engineers transitioning into ML-focused roles at the enterprise tier; frontier labs prioritize direct work history over credentials at senior levels.
Is the AI engineer job market strong enough to justify targeting these labs in 2026?+
Yes, with important caveats. AI Engineer was LinkedIn's number one fastest-growing US job title in 2026, with postings up 143% year-over-year and a 3.2:1 demand-to-supply ratio (LinkedIn Workforce Report 2026). However, the premium is concentrated: entry-level AI engineer hiring at the top 15 tech firms fell 25% from 2023 to 2024, while senior-level demand surged. The AI premium over non-AI engineers is 18% at the staff level but only 6% at entry level (Levels.fyi 2026). The frontier-lab compensation tier discussed in this article applies to a few hundred new hires per year across all three companies combined.