Frontier labs have spent four years concealing their compensation data from the engineers who generate it. Six platforms — Levels.fyi, Glassdoor's AI Salary Intelligence, the ENTRA Job Signal Index, Radford AI Analytics, Merc, and LinkedIn Salary Insights — are now deploying machine learning against that concealment, with measurable but partial success. The data battle is not close to won: frontier labs retain structural advantages in pay opacity that no algorithm has yet neutralized, and the $43K negotiation gap that separates candidates who know their number from candidates who do not is the clearest proof that the information asymmetry remains load-bearing.
Six Platforms, Six Architectures
Six platforms entered Q3 2026 with meaningfully different architectures, data sources, and structural biases. Understanding what each is and is not capable of is a prerequisite for using any of them correctly.
Levels.fyi reaches approximately 40 million annual visitors. Its 2026 AI-enhanced verification layer cross-references submitted comp offers against known posting parameters, HR metadata patterns, and equity structure signals, flagging fabricated submissions with a 94% precision rate per Levels.fyi's published methodology. The ENTRA corpus now tracks 3,800+ verified AI-specific offers from that dataset, covering research-track and engineering-track roles across Anthropic, OpenAI, DeepMind, Mistral, xAI, and Cohere. The 2025 addition of frontier-lab equity vesting-schedule tracking — converting private company grants to "4-year annualized" total comp using last-round valuations — represents the most consequential upgrade to candidate-facing comp data infrastructure since the platform launched. The tool is not perfect; its private-company equity assumptions carry real uncertainty. It is, still, the best single source of AI compensation intelligence available to a candidate without institutional data access.
Glassdoor's AI Salary Intelligence was retrained in 2025 on an 85M+ salary corpus and now disaggregates base, RSU, and signing bonus for AI-sector roles — directly addressing the historic conflation that allowed labs to report a $320K base ceiling while concealing $305K in equity and signing structures above it. Glassdoor's "true take-home" calculator, launched Q1 2026, adjusts for state marginal tax rates and metropolitan cost-of-living indices, producing a take-home equivalent figure that is more analytically useful to a candidate choosing between an Anthropic SF offer and a DeepMind London offer than either company's published band.
LinkedIn Salary Insights draws on 900M+ member data points but carries a structural bias toward disclosed base salary, because employer-reported data skews toward the compensation components that legal disclosure requires or employers choose to share. The full-TC picture for an L5 or L6 research role is systematically underrepresented. LinkedIn Salary Insights is most useful as a volume signal — revealing band distributions and hiring-direction trends — not as a total comp benchmark.
Radford AI Analytics occupies the institutional tier. Radford's survey-grade data covers 800+ technology companies and is the compensation benchmark used by major HR functions, including at several frontier labs. The problem is structural: Radford sells to employers, not candidates. An Anthropic People team subscriber knows what the L5 research scientist market pays across twelve competing labs, disaggregated by level, function, and geography. The candidate sitting across from them does not have that data. The information asymmetry is not incidental; it is the product's design.
The ENTRA Job Signal Index monitored 47,000+ AI job postings in Q2 2026. Beyond volume, the index now rates individual postings on a "narrowness score" — a band-width metric on which a $60K spread earns C- and a $30K spread earns A, making disclosure quality comparable across employers for the first time. OpenAI's Q2 2026 postings averaged C+. Anthropic averaged B-. xAI averaged D+. Mistral and Hugging Face, both European-headquartered, clustered in the B-to-A range — a regional disclosure pattern with competitive implications discussed below.
Merc (formerly Mercor) operates at the candidate-facing end, providing comp intelligence specifically for AI/ML roles with primary coverage of US frontier lab markets. Its Q2 2026 dataset holds verified TC data for 6,200+ AI/ML offers, weighted toward research-track and ML engineering roles at Anthropic, OpenAI, and DeepMind. Its limitation is geographic: non-US markets are thin, and Merc's candidate-facing positioning means it indexes more toward applicants than toward the full population of placed engineers.
Opacity by Design
Frontier labs have not resisted pay transparency reactively. The mechanisms through which they maintain pay opacity are architecturally deliberate, and they have adapted as each new transparency tool has emerged.
Band suppression is the most visible tactic. OpenAI's Q2 2026 postings carried ranges as broad as $150K–$300K+ for positions spanning L4 through L6 — a three-level range in which the true total comp variation runs from approximately $200K to $650K. A $150K-wide base band complies with California SB 1162 and tells a candidate almost nothing. It is legal disclosure designed to satisfy the minimum requirement without surrendering any actual information.
Equity packaging is the more sophisticated mechanism. Anthropic posts an L5 researcher base band with a ceiling of $320K. The verified L5 total comp, per ENTRA's Q2 2026 Candidate Survey (n=214 US placements), sits at $625K — a gap of $305K that lives entirely in RSU grants, equity refresh packages, and company-value appreciation on pre-IPO shares. California's SB 1162, amended by SB 642 effective January 2026, closed the absurd-range loophole. It did not and legally cannot require equity disclosure. For an Anthropic L5, the omission describes the majority of compensation.
Geographic arbitrage exploits the interaction between multi-location remote-work hiring and state disclosure laws. xAI posts comp bands of $130K–$310K for roles where the realistic offer to a San Francisco-located senior candidate runs $280K–$310K in base alone, before equity. The wide posted range reflects nominal hiring locations spanning Memphis, Austin, and the Bay Area in a single requisition. A California candidate reading an xAI posting has no way to know which end of the band describes their situation. This is not an error in the posting. It is the architecture.
Signing bonus weaponization operates outside every disclosure framework. Signing bonuses of $50K–$150K are standard at L5 and above across Anthropic, OpenAI, and DeepMind. They do not appear in posted salary bands because no disclosure law requires it. Levels.fyi added a signing bonus normalization field in late 2025 — the retrospective data for 2024 is still being backfilled — but the structural gap between what a posted band states and what a candidate actually receives at offer remains partially invisible to every platform.
Referral-only hiring is the fifth mechanism and the structurally hardest to address. Approximately 28% of Anthropic hires in 2025 came through internal referral networks, per ENTRA's 2025 hiring-channel analysis. Referral-channel candidates receive offers that never hit a job board, never carry a posted band, and never enter the disclosure ecosystem. The $43K transparency dividend that benchmarking tools deliver does not apply when there is no posted role to benchmark against.
What the Algorithms Get Right
The 2026 benchmarking landscape is materially better than 2024's on three dimensions.
Equity normalization has made real progress. Levels.fyi now converts all submitted total compensation to a 4-year annualized figure including RSU appreciation at last-round valuation for private companies. For Anthropic, currently at an $852B post-money valuation from its most recent financing round, the platform applies an annualized equity appreciation factor to convert grant-at-issue RSU values to projected total-comp equivalents. This makes a $625K Anthropic L5 offer and a $580K OpenAI L5 offer comparable in a way that raw base-plus-grant figures were not. The methodology is imperfect — it cannot model liquidation waterfall mechanics — but the output is more useful than any prior approach.
Geographic take-home adjustment is functional for US markets. Glassdoor's "true take-home" calculator covers California (13.3% marginal rate at senior comp levels), Washington (no income tax), Texas (no income tax), and New York City (up to 10.9% combined marginal rate). A senior AI researcher deciding between an Anthropic San Francisco offer and an xAI Austin offer can produce take-home equivalent figures with a precision that was unavailable before 2026. Glassdoor's UK and EU modules are less developed but improving quarterly.
Band-width scoring, via the ENTRA Job Signal Index, provides the first systematized cross-employer metric for disclosure quality. The Q2 2026 distribution across 47,000+ postings is not flattering to US frontier labs: the lowest quartile of AI employer disclosures carries band widths above $80K; the top quartile posts bands below $25K. Cohere, Hugging Face, and Mistral sit in the B-to-A range. xAI and several early-stage AI infrastructure companies sit at D+ or below. The score does not tell a candidate what they will earn, but it reliably identifies which employers are disclosing in good faith and which are complying minimally.
The Remaining Blind Spots
Four problems remain structurally unsolved, and they affect data quality across every platform in the ecosystem.
Private company equity valuation is the hardest. Anthropic's $852B post-money valuation is a single number that describes four different financial realities: the value to early employees (large upside, fully vested, preference-stacking headroom below them), the value to Series D hires (meaningful upside, partial vesting), the value to Series H hires (narrower spread between issue price and current round price, full preference stack above them), and the value to whatever the IPO-day public price turns out to be. No benchmarking platform has a credible model for liquidation waterfall mechanics. Levels.fyi's equity normalization uses last-round valuation as a proxy. That approach systematically overstates late-stage grant value relative to early-stage grant value, and understates the dilution effect of preference stacking. The gap between what a platform says an Anthropic Series H hire's equity is worth and what it is actually worth at liquidity could run $80K–$200K in either direction.
Non-US jurisdictions remain a 90% data desert. Levels.fyi's verified offer corpus is approximately 85% US-weighted. DeepMind London's £240K–£340K total comp range for entry research rests on fewer than 50 verified submissions. Mistral Paris's approximately €280K base plus €240K equity is ENTRA-estimated from recruiter network intelligence, not crowdsourced. The EU Pay Transparency Directive — tracking toward transposition in France and Germany in 2027 — will require full compensation disclosure including variable pay and equity for the first time. When it is fully in force, the data quality gap between US and European markets should narrow materially. Until then, any platform comparison between a US frontier lab offer and a European frontier lab offer carries uncertainty on the European side that is not adequately flagged in the interface.
Level inflation remains algorithmically naive across every platform. "Senior Research Scientist" at OpenAI, at Cohere, and at Scale AI describes three different roles in terms of scope, autonomy, research impact, and compensation. OpenAI's L6 research band runs $480K–$740K total comp per ENTRA's Q2 2026 Salary Survey (n=2,140). Cohere's equivalent senior researcher band sits materially below that. Scale AI's research function is a different category entirely. Benchmarking platforms map by title and seniority keyword. Levels.fyi has introduced employer-specific level mapping for major labs, distinguishing "OpenAI L6" from "Cohere L6" in its normalized outputs. For less data-rich employer pairs, cross-company level equivalence remains manual estimation.
The vibes premium is the most uncomfortable finding in ENTRA's Q2 2026 Candidate Survey (n=214 US placements). Candidates with insider network access — those who received a warm referral, had a prior relationship with the hiring manager, or were known within the team before the formal process began — received offers averaging 15–22% above the platform-median total comp for their level at the same employer. The information asymmetry is not only about knowing what to ask for. It is about being in a position where the employer voluntarily offers more because the relational dynamic changes the negotiation before it starts. No platform can solve this. Benchmarking tools close the information gap. They cannot close the social capital gap.
The $43K Transparency Dividend
The ENTRA Pay Transparency Dividend report, published August 7, 2026, documented a $43K median negotiation uplift for AI-sector candidates who arrived at offer stage with verified, benchmarked salary data, compared to candidates who did not. The methodology tracked 214 US AI/ML placements through ENTRA's recruiter network in Q2 2026, comparing final negotiated total compensation against first-offer total compensation, stratified by benchmarking-tool use. The $43K figure is a median. The distribution is right-skewed: candidates at the senior research track (L6 equivalent) showed uplifts of $80K–$140K. Even candidates at L4 entry bands showed a $12K–$18K median uplift when they entered negotiation with Levels.fyi data in hand.
The figure is, analytically, an argument for mandatory disclosure. If the information asymmetry between employer and candidate is worth $43K in negotiation outcomes — compensation that employers were willing to pay but that candidates without data left on the table — the asymmetry is not a neutral market outcome. It is a structural transfer, repeating with each hiring cycle, from candidates who lack platform access or platform literacy to employers who benefit from opacity. Every benchmarking platform that closes part of that gap is running a redistribution mechanism inside the AI labor market. Every tactic the labs use to undermine that data — wide bands, equity packaging, signing bonuses that never touch the posted range — is, in this frame, a mechanism for preserving a transfer.
2027: Three Regulatory Triggers
Three events will reshape the benchmarking landscape by end of 2027. EU Directive transposition in France and Germany — the markets where enforcement will move the most AI labor — is tracking to Q1–Q2 2027. When implemented, the Directive will be the first regulatory framework in any major market to require disclosure of full compensation packages, including equity and variable pay. California SB 1162's second full annual pay data reporting cycle, required under SB 642's expanded scope, will produce the first multi-year time series comparing posted bands to actual pay outcomes — a dataset that transparency platforms will immediately index. And if Washington state's Branson ruling creates a template for broader WA EPOA enforcement expansion — enforcement risk that both Microsoft and Amazon are actively managing — the pressure on band width at the US's two largest AI infrastructure employers will increase without federal action.
The more proximate force may be competitive. Mistral and Hugging Face's voluntary transparency posture — both maintain higher ENTRA narrowness scores than any US frontier lab — has become a recruiting signal in the European research market, where candidates cite transparency as a factor in lab selection at measurably higher rates than their US counterparts per ENTRA's Q2 2026 Candidate Survey. If the D+ narrowness scores at xAI and early-stage infrastructure labs begin correlating with hiring velocity declines in European research pipelines, the architecture of opacity will carry a talent cost. In 2027, the data will be good enough to measure that cost. Whether the labs decide it is worth paying is a different question — one that the benchmarking platforms, no matter how precise their algorithms, cannot answer for them.
Methodology: ENTRA Job Signal Index Q2 2026 (47,000+ AI job postings monitored). ENTRA Salary Survey Q2 2026 (n=2,140 respondents). ENTRA Q2 2026 Candidate Survey (n=214 US placements). ENTRA Pay Transparency Dividend report (August 7, 2026). Levels.fyi AI offer corpus (3,800+ verified offers tracked by ENTRA). Glassdoor AI Salary Intelligence (85M+ salary corpus). Radford AI Analytics (institutional survey-grade data, employer-facing). LinkedIn Salary Insights (900M+ member data). Merc/Mercor candidate platform (6,200+ verified AI/ML offers, Q2 2026). Comp figures attributed to "ENTRA reporting" or "ENTRA estimate" reflect recruiter network intelligence and carry higher uncertainty than platform-crowdsourced medians. Narrowness scores computed by ENTRA from Job Signal Index posting analysis. All USD figures unless marked. Anthropic post-money valuation per Series H financing documentation. Cross-company level equivalences are illustrative; individual compensation varies by negotiation, performance tier, and hiring-channel. Band-width grade thresholds: A = spread ≤$30K; B = $30K–$45K; C = $45K–$65K; D = >$65K.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.