In [22]: # openai_anthropic.ipynb
Direct answer: OpenAI vs Anthropic salary
At senior (L5-equivalent) level, OpenAI ML engineer total compensation runs $700k to $1.1M and Anthropic $650k to $1.0M. OpenAI typically pays 5 to 10 percent higher headline total comp at the senior (L5) level, reflecting its larger revenue base; at staff (L6) the two are near parity. The difference is small; the choice between the two is usually driven by research focus and equity structure (OpenAI Group PBC stock since the 28 October 2025 recapitalization vs Anthropic capped options) rather than the comp gap.
$700k-1.1M
OpenAI L5 TC
$650k-1.0M
Anthropic L5 TC
+5-10%
OpenAI premium
Abstract
OpenAI and Anthropic are the two largest frontier AI labs by revenue and employee count, and the two highest-profile destinations for senior ML engineer compensation in 2026. Following its October 2025 recapitalization, OpenAI now grants conventional equity in OpenAI Group PBC (a public benefit corporation), replacing the capped Profit Participation Unit (PPU) structure it used previously. Anthropic uses capped pre-IPO option grants with mechanics designed to balance employee upside against the company's mission-control governance. Senior L5 total compensation at OpenAI is approximately $700,000 to $1,100,000; at Anthropic approximately $650,000 to $1,000,000 [1].
1 L4 to L6 bands are anchored to Levels.fyi OpenAI (L4 median ~$612k, L5 ~$941k, L6 ~$1.23M, open-ended top ~$1.34M+) and Levels.fyi Anthropic (senior ~$591k, lead ~$863k, staff median ~$1.27M, open-ended top ~$1.41M), both checked September 2026, with cross-reference to public reporting in The Information and Bloomberg. Neither aggregator publishes a distinct L7/principal band, so L7 figures are floors extrapolated from the open-ended top of the L6/Staff data, not benchmarked ranges. Equity and option terms vary by cohort and offer; candidates should request explicit terms.
table oa-1 : total comp
| Level | OpenAI TC | Anthropic TC | Note |
|---|---|---|---|
| L4 / Mid-senior | $520k - $750k | $480k - $700k | Both use non-standard titling; entry-to-mid track |
| L5 / Senior | $700k - $1.1M | $650k - $1.0M | Equity-driven; refresh cadence active |
| L6 / Staff | $1.0M - $1.35M | $950k - $1.4M | Slot-limited; medians near parity at this level |
| L7 / Principal | $1.5M+ | $1.4M+ | No distinct L7 band published; floors from open-ended L6/Staff top |
Figure oa-1. OpenAI vs Anthropic total compensation by L-equivalent level. L4 to L6 anchored to Levels.fyi medians (checked September 2026); L7 shown as a floor because no aggregator publishes a distinct principal band. Both labs use non-standard internal titling, so mapping to the L4-L7 ladder is approximate.
section oa-2 : from PPUs to PBC stock
Until late 2025, OpenAI's compensation was unusual among frontier AI labs: the company operated under a capped-profit structure (a for-profit operating entity with a cap on investor returns, governed by a non-profit parent), and employees received Profit Participation Units (PPUs) rather than conventional equity. A PPU was a contractual right to a share of OpenAI's profit distributions, subject to a stated per-unit cap, rather than an ownership stake.
That structure ended on 28 October 2025. After approval from the California and Delaware attorneys general, OpenAI converted its for-profit entity into OpenAI Group PBC (a public benefit corporation), with the renamed OpenAI Foundation retaining control. The OpenAI Foundation holds roughly 26 percent of the PBC, Microsoft roughly 27 percent, and employees and other investors the remaining roughly 47 percent. Outstanding PPUs converted into PBC shares and the prior return cap was removed, so equity value now tracks the company's valuation without a contractual ceiling.
For ML engineers evaluating OpenAI offers in 2026, the practical implication is that new grants are conventional equity (restricted stock units or options in OpenAI Group PBC) on a standard 4-year vesting schedule, rather than capped PPUs. Those grants are struck against a steeply rising reference valuation: OpenAI's March 2026 round raised $122 billion at an $852 billion post-money valuation, up from roughly $500 billion at its October 2025 employee tender offer. Realisation still depends on a liquidity event: OpenAI has run periodic employee tender offers (most recently a $7 billion buyback that closed in August 2026 at the same $852 billion valuation) and filed confidentially for an IPO in June 2026. On 19 August 2026, CFO Sarah Friar told employees that OpenAI "will be a public company in 2027," or sooner if its business continues to inflect, but vested equity remains illiquid until that listing or the next tender offer.
Tax treatment now follows conventional equity rules rather than the old PPU distribution model. RSUs are taxed as ordinary income at vesting (or at the vesting-plus-liquidity event for double-trigger grants); NSO options are taxed on the spread at exercise. For an ML engineer in California, the combined marginal federal-plus-state rate at these compensation levels approaches 50 percent, so the timing of vesting events and tender-offer sales is the main tax-planning lever.
section oa-3 : option mechanics
Anthropic uses pre-IPO option grants with capped value mechanics. The cap structure is designed to balance employee upside against the company's long-term mission-control governance (Anthropic's stated long-term governance model emphasises continued employee and founder alignment with its safety-focused mission). The detailed cap mechanics are non-public, but in functional terms, options vest over 4 years with conventional cliffs.
The cap structure means that even if Anthropic's eventual public-market valuation grows substantially, the realisable value per option is bounded above. For most ML engineers, the cap is set at levels that still allow meaningful upside (the cap binds in extreme high-valuation scenarios), but the headline maximum realisable value is lower than an uncapped pre-IPO option at a peer lab. The cap structure also affects how Anthropic equity should be valued relative to OpenAI's PBC equity or to standard pre-IPO options at smaller frontier labs.
Vesting and exercise mechanics are conventional. Options typically have a 10-year expiration from grant date and require exercise before that date or before separation from the company. Anthropic has run periodic tender offers that let employees sell vested equity at then-current secondary-market valuations, and those reference valuations have risen steeply: its Series G round in February 2026 priced the company at $380 billion post-money ($350 billion pre-money), and its Series H round in May 2026 at $965 billion post-money, so recent grants and secondary sales are struck against a rapidly rising valuation. These offers allow employees to sell a portion of vested equity at secondary-market valuations, providing intermittent liquidity without requiring full company exit. Tender-offer participation is at the company's and outside investors' discretion and is not guaranteed.
Anthropic filed confidentially for a US IPO on 1 June 2026 and has since set a target listing window of autumn 2026, with September 2026 reports pointing to an October Nasdaq listing, managed by Morgan Stanley, Goldman Sachs and JPMorgan. Public-market valuation estimates run well above the $965 billion Series H mark, but no shares have priced, so any trillion-dollar figure is speculation until the listing completes. A completed IPO would be the first broad liquidity event for employee equity (replacing the discretionary tender offers as the main exit) and would typically carry a lock-up of around six months before employees can sell, but it does not remove the capped-option mechanics, which still bound the maximum realisable value per option.
Tax treatment: Anthropic options are typically NSOs (non-qualified stock options), which means exercise triggers ordinary income tax on the spread between strike price and current fair-market value. For options exercised at high spread, the tax bill at exercise can be substantial and is owed in cash even though no actual cash has changed hands. Exercise timing matters; many Anthropic employees exercise vested options at tender-offer events to align tax with cash availability.
section oa-4 : common questions
What is the average salary at OpenAI for ML engineers?
Senior (L5-equivalent) ML engineer total compensation at OpenAI in 2026 is approximately $700,000 to $1,100,000, comprising base salary $280,000 to $400,000 and equity grants whose annualised paper value is $400,000 to $700,000. Since OpenAI's October 2025 recapitalization into OpenAI Group PBC, these grants are conventional equity (restricted stock units or options); they replaced the capped Profit Participation Unit (PPU) structure OpenAI used before the restructuring.
What is the average salary at Anthropic for ML engineers?
Senior (L5-equivalent) ML engineer total compensation at Anthropic in 2026 is approximately $650,000 to $1,000,000, comprising base salary $260,000 to $380,000 and pre-IPO equity option grants whose annualised paper value is $390,000 to $620,000. Anthropic uses a capped-option structure designed to balance employee upside against the company's long-term mission-control governance. The cap mechanics are not fully public; candidates should ask for the specific cap structure on any offer.
What is the L4 salary at Anthropic and OpenAI?
At the L4 (mid-senior) equivalent level, ML engineer total compensation is approximately $520,000 to $750,000 at OpenAI and $480,000 to $700,000 at Anthropic. These bands are consistent with Levels.fyi aggregated self-reported data (checked September 2026), which shows an OpenAI L4 software-engineer median of about $612,000 and an Anthropic senior-engineer median of about $591,000. Both labs use non-standard internal titling, so the mapping to a conventional L4 band is approximate; L4 is the entry-to-mid track above new-graduate hires. These are total-compensation figures (base salary plus the annualised paper value of equity), not base salary alone.
What is the L7 (principal) salary at OpenAI and Anthropic?
No aggregator publishes a distinct L7 (principal / distinguished) band for either lab. Levels.fyi self-reported data for the highest published levels shows OpenAI L6 at a median of about $1,230,000 (open-ended top around $1,340,000) and Anthropic Staff at about $1,270,000 (open-ended top around $1,410,000), so a principal-level ML engineer sits above those tops rather than inside a benchmarked range. The population at this band is very small, packages are negotiated individually, and total compensation is dominated by equity, so realised value depends heavily on each lab's valuation trajectory and the timing of liquidity events (periodic tender offers or an eventual IPO). Treat any single principal-level number as an individual data point, not a published market rate.
How does OpenAI equity work after the 2025 restructuring?
On 28 October 2025 OpenAI converted its capped-profit for-profit entity into OpenAI Group PBC, a public benefit corporation controlled by the renamed OpenAI Foundation, after approval from the California and Delaware attorneys general. The conversion removed the previous return cap and converted employees' Profit Participation Units (PPUs) into conventional PBC shares; new grants are stock-based (RSUs or options) vesting over a standard 4-year schedule. Equity is taxed under normal rules (ordinary income for RSUs at vesting, NSO spread at exercise) rather than the old PPU model of ordinary income at profit distribution. The equity remains illiquid until a tender offer or IPO; OpenAI has run periodic tender offers, and CFO Sarah Friar told employees on 19 August 2026 that OpenAI will be a public company in 2027, or sooner if its business continues to inflect.
Are Anthropic options exercisable like standard ISOs?
Anthropic options are pre-IPO instruments with vesting and exercise rules similar to standard private-company options, but with specific cap-structure modifications that affect the maximum realisable value. The detailed terms are negotiated per offer and per cohort. Candidates evaluating Anthropic offers should ask explicitly: what is the strike price, what is the vesting schedule and cliff, what is the cap structure, what is the recent history of tender offers or liquidity events. For tax purposes, options are typically NSOs (non-qualified stock options) at this scale of private company.
How does Anthropic's planned IPO affect equity compensation?
Anthropic filed confidentially for a US IPO on 1 June 2026 and has set a target listing window of autumn 2026, with September 2026 reports pointing to an October Nasdaq listing, managed by Morgan Stanley, Goldman Sachs and JPMorgan. Its most recent private round (Series H, May 2026) valued the company at approximately $965,000,000,000 post-money; public-market estimates run higher, but no shares have priced, so treat any trillion-dollar figure as speculation until the listing completes. For an ML engineer holding Anthropic options, an IPO matters in three ways: it creates the first broad liquidity event, replacing the periodic and discretionary tender offers as the main way to sell vested equity; it typically comes with a lock-up of around six months after listing before employees can sell; and it does not remove the capped-option mechanics, which still bound the maximum realisable value per option regardless of how high the public valuation goes. On any Anthropic offer, ask for the strike price, the vesting schedule and cliff, the cap structure, and the current expectation on lock-up.
Which pays more, OpenAI or Anthropic, for senior ML engineers?
OpenAI typically pays a modest premium at the senior (L5) level, on the order of 5 to 10 percent, reflecting its larger revenue base and longer history of established PPU distribution. At the staff (L6) level the two labs' crowd-sourced medians are effectively at parity (within a few percent, and the ranking moves with sampling), so the OpenAI premium is a senior-level phenomenon rather than a fixed gap across the ladder. The choice between the two for most senior ML engineers is rarely driven by the comp difference; it is driven by research focus, team fit, organisational culture, and the candidate's view of which lab's strategy is more likely to succeed long-term. Both are competitive top-of-market offers.
How risky is OpenAI or Anthropic equity compared to public-company RSUs?
Substantially riskier. Both companies are pre-IPO with illiquid equity instruments. Realisation depends on either tender-offer participation (which both companies have run periodically), an eventual IPO, or an acquisition. The downside risk is the company failing to monetise foundation-model capability at scale sufficient to support current valuations. OpenAI and Anthropic both have substantial revenue and broad commercial deployments, which reduces this risk relative to earlier-stage frontier labs, but the equity remains illiquid and contingent until a defined liquidity event.
How do offers from xAI, Mistral, and Cohere compare?
xAI offers are comparable to OpenAI and Anthropic for senior ML engineers, with base salaries in similar ranges and pre-IPO equity grants tied to xAI's current private-market valuation. Mistral offers in Paris are materially lower in base salary (EUR-denominated, France labour-market norms) but with European-tax-favourable equity structures; total compensation in euro terms is competitive. Cohere offers in Toronto and SF are below US-frontier-lab averages on base salary but competitive on equity. All four labs operate in the same talent pool and compete for similar candidates; offer packages converge on rough parity in risk-adjusted total comp.
Anthropic salary deep dive
Single-lab: level bands, capped options, IPO
Frontier-lab tier overview
T1 deep dive, equity forms
Bay Area metro
Where both labs HQ
LLM engineer salary
The specialisation premium
RLHF engineer salary
Post-training at frontier labs
Negotiation playbook
Frontier-lab offer tactics
Total comp breakdown
How to value equity packages
In [ ]: # reader_questions.ipynb
Write in. A person opens every message, and the ones we can answer come back with the survey release or the public disclosure the number is drawn from, stated plainly enough that you can go and check it. We put no clock on a reply. Should an answer be worth reading twice we might set it out here stripped of anything identifying, and we will write and ask you before that happens.
Out[ ]:
Treat everything on this site as compensation reference assembled from published surveys and public filings, not as financial advice and not as a benchmark any single negotiation should be anchored to. Your own offer turns on the employer, the level it is pitched at, the equity instrument behind it and the market on the day it is made.