Anthropic CEO Dario Amodei this week proposed that independent AI safety evaluators — organizations like METR that assess frontier models for dangerous capabilities before they ship — should be paid roughly $687,000 a year, an attempt to address a structural problem that's quietly undermined third-party AI safety oversight all year: the people checking frontier labs' work are paid a fraction of what the labs themselves pay comparable technical staff.
TL;DR — what people are asking
| Question | Answer |
|---|---|
| What was proposed? | ~$687,000 salaries for independent AI safety evaluators like METR |
| Who proposed it? | Anthropic CEO Dario Amodei |
| What is METR? | An independent nonprofit conducting pre-deployment dangerous-capability evaluations for frontier labs |
| Is Anthropic funding this directly? | Not specified — framed as an ecosystem-wide proposal, not a confirmed Anthropic funding commitment |
| Why does pay matter here? | Frontier labs can outbid nonprofit evaluators for the same technical talent, undermining evaluation quality |
| What's the broader context? | Part of Amodei's ongoing public framing that safety infrastructure is a collectively underfunded public good |
The problem this proposal is actually addressing
Third-party AI safety evaluation depends on organizations like METR being able to recruit researchers with genuinely deep technical skill — people capable of stress-testing a frontier model's autonomous-agent capabilities, cyberoffense potential, or self-replication risk in ways that actually reveal something a lab's own internal red team might miss or might be structurally incentivized to soft-pedal. That kind of technical talent is exactly what frontier labs themselves compete hardest to recruit, often with total compensation packages — salary plus equity — reaching well into seven figures for senior technical roles.
Nonprofit safety evaluators operate on a completely different financial footing, typically funded through philanthropic grants or modest lab contributions rather than venture-scale capital or equity upside. That gap creates a predictable talent asymmetry: the organizations meant to independently check frontier labs' safety claims are systematically outcompeted for hiring by the very labs they're supposed to evaluate. Amodei's proposal names a specific number — $687,000 — as an attempt to close that gap meaningfully rather than incrementally.
Why this matters more in 2026 specifically
This proposal lands in a year where independent evaluation has become a recurring, load-bearing part of the AI safety conversation, not a theoretical nicety. explainx.ai has tracked METR and similar evaluators appearing repeatedly across major incident and capability disclosures this year — pre-deployment cyber-capability assessments, agentic-risk evaluations ahead of frontier model launches, and now OpenAI's own new misalignment disclosure framework, which explicitly separates internal reporting from the role independent evaluators play in catching what a lab's own teams might miss or underweight.
Amodei's own broader framing this year — the Dario Amodei "pace the frontier" position on embedding evaluators directly into training pipelines — treats evaluation capacity generally as infrastructure the industry needs more of, not less. A compensation proposal for external evaluators is a direct extension of that same argument applied to the nonprofit ecosystem specifically rather than internal lab processes.
What "independent" actually requires
There's a structural tension worth naming plainly: if frontier labs themselves end up funding the salary increases for the evaluators checking their own models, does that compromise the independence the whole arrangement depends on? Amodei's proposal didn't specify the funding mechanism — whether it would come from a lab consortium, government grants, or dedicated philanthropic funding specifically insulated from any single lab's influence. That detail matters enormously for whether raised evaluator pay actually strengthens independent oversight or simply shifts the funding relationship without changing the underlying incentive structure. This is the kind of "who pays the auditor" question that recurs across regulated industries generally, and AI safety evaluation isn't exempt from it just because the proposal comes with good intentions attached.
Why $687,000 is a specific and telling number
It's worth dwelling on why Amodei proposed a specific dollar figure rather than a vaguer call for "better funding" or "competitive compensation." A specific number is a checkable, falsifiable commitment in a way that vague language isn't — it invites direct comparison against what frontier labs actually pay comparable technical roles internally, and it gives METR and similar organizations a concrete target to advocate for rather than an open-ended aspiration. That specificity also signals Amodei likely has some real reference point in mind — plausibly a rough approximation of senior research engineer or safety researcher total compensation at a frontier lab, adjusted to a nonprofit context without equity upside, which independent evaluators structurally can't offer the way a venture-backed or publicly traded lab can.
The number is also politically useful in a different way: it's precise enough to sound serious and researched rather than performative, which matters for a proposal aimed at influencing philanthropic funders, government grant programs, or a potential lab consortium — all audiences that respond better to a specific, defensible ask than an open-ended plea for "more resources."
The broader talent-pipeline problem this proposal is trying to solve
Pay parity alone doesn't fully solve the deeper structural challenge independent AI safety evaluation faces: even with competitive salaries, nonprofit evaluators are competing against frontier labs not just on compensation but on access to compute, access to the very models being evaluated (often under restricted, pre-release terms), and the prestige and career-trajectory appeal of working directly on frontier model development rather than on external evaluation of someone else's model. A researcher choosing between a frontier lab role and an independent evaluator role, even at equal pay, may still lean toward the lab role simply because it offers more direct influence over what gets built, rather than an auditing role checking someone else's work after the fact.
Solving that fuller talent-pipeline problem likely requires more than compensation parity — it may also require independent evaluators building their own compelling research agendas, publishing findings that meaningfully shape the field's understanding of AI risk (which METR and similar organizations have already done to some degree this year), and cultivating a professional identity around independent evaluation as prestigious and impactful in its own right, not merely a lower-tier alternative to working at a frontier lab directly. Amodei's compensation proposal is a necessary piece of that puzzle, but likely not sufficient on its own to fully close the talent gap it's aimed at.
Honest limitations
- No confirmed funding source. The proposal names a target salary figure without specifying who would actually pay it or how independence would be preserved if labs themselves fund it.
- Not an Anthropic-specific commitment. This reads as an ecosystem-wide proposal, not confirmation that Anthropic itself will fund METR salaries at this level.
- METR's own position on the proposal wasn't reported alongside Amodei's remarks — whether the organization itself endorses this specific figure or funding approach is unclear.
- Other frontier labs' reactions weren't detailed. Whether OpenAI, Google DeepMind, or other major labs have signaled support for, or reservations about, this specific proposal wasn't part of available coverage, leaving unclear how broadly shared this view is across the industry rather than being an Anthropic-specific position.
- No timeline for implementation was suggested. Even if the proposal gains support, a proposal-to-funded-reality timeline for something requiring new funding commitments from philanthropic or industry sources could plausibly take many months to materialize, if it happens at all.
- No comparison to what other independent evaluators beyond METR currently pay their own technical staff was provided, making it unclear whether $687,000 is being proposed as a METR-specific figure or as a broader industry benchmark applicable across all similar organizations.
Why this proposal is worth revisiting in a few months
Proposals of this kind are easiest to evaluate in retrospect, once it becomes clear whether they translated into an actual funding commitment or faded as a one-time talking point. Readers tracking AI safety infrastructure specifically should treat this as a placeholder worth checking back on — has METR's own compensation structure changed, have other labs echoed similar proposals, and has any philanthropic or consortium funding mechanism actually materialized to support it. That follow-through, more than the initial proposal itself, will determine whether this represents a genuine shift in how the industry funds independent oversight or simply a well-intentioned statement that didn't translate into changed practice.
- Doesn't address evaluator capacity beyond pay — recruiting and retaining talent also depends on career trajectory, research freedom, and organizational stability, none of which a salary figure alone resolves.
What this means for what you build or pay
AI safety researchers and evaluators: if this proposal gains traction, it's a meaningful signal that compensation parity with frontier labs is becoming a live industry conversation rather than an accepted structural gap — worth watching for concrete follow-through, not just the initial statement.
Teams relying on third-party safety evaluations to inform procurement or deployment decisions: the quality and depth of independent evaluation is directly downstream of whether organizations like METR can actually recruit and retain top technical talent — a funding gap here is a real quality risk for the evaluations your own risk assessments may lean on.
Policy watchers: track whether any concrete funding mechanism follows this proposal, and specifically whether it preserves evaluator independence from the labs being evaluated — that detail will determine whether this strengthens or merely reshapes the current oversight relationship.
Related on explainx.ai
- OpenAI ships a misalignment disclosure framework — and six reports
- Dario Amodei: "pace the frontier" with embedded evaluators
- What is an embedded evaluator in AI safety?
- Sanders' Superintelligence Ban Act and Anthropic's extinction-risk framing
- JD Vance: AI labs should build defenses, not seek regulation
- AI regulation: EU AI Act and US policy, complete guide
Details reflect public remarks reported as of September 17, 2026. No confirmed funding mechanism or Anthropic-specific financial commitment was disclosed at time of writing.