What top PM roles pay for agent evals
The real job postings, checked, and an honest read on what the numbers mean.
Here is what the postings say. The two Anthropic rows marked "verified" I re-checked on Anthropic's live job board on 5 October 2026. The rest come from a research pass the same day.
| Role | Company | Posted base | Eval language in the posting |
|---|---|---|---|
| Product Manager, Claude Science (verified) | Anthropic | $305,000 to $385,000 | "Have personally built evals or benchmarks for model capabilities, ideally agentic or scientific ones." Also: "build and shape evals grounded in real scientific workflows, surface failure modes from real usage." |
| Product Manager, Research (Code) (verified) | Anthropic | $305,000 to $385,000 | "Synthesize user insights into actionable requirements and evaluations." |
| Product Manager, Claude Code Model Performance | Anthropic | $305,000 to $460,000 reported | The research pass recorded "Have personally built agentic evals (e.g. SWE-bench-style task suites)." It was no longer on Anthropic's board when I re-checked, so treat it as closed. |
| Product Manager, Core Models | OpenAI | $347K to $490K plus equity | Define success across offline evaluations and online product metrics; "distinguish a useful metric from a convenient one"; build reusable platforms for evaluation. |
| Product Management, Research | Anthropic | $305,000 to $385,000 | None |
| Product Manager, API Agents | OpenAI | $293K to $325K plus equity | None, even though the role is about agents |
| Staff Product Manager, AI Platform | Databricks | $171,000 to $235,200 | None |
Honest read:
- At frontier labs, $250k understates it. Posted base bands start around $293k to $347k before equity.
- Inside the same lab, the eval-heavy PM roles sit in the same band as PM roles with no eval wording. So the band reflects "PM at a frontier lab at this level." Evals are a requirement for some of those seats, not a premium on top.
- Outside the labs, $250k is mostly a total-comp number, not a base number. A June 2026 recruiter guide (KORE1) puts US AI PM base pay at $165K to $238K (25th to 75th percentile) and total comp at $244K to $390K, median $305K.
- Eval requirements cluster in PM roles that sit next to research and model quality. Product-surface roles (like OpenAI's API Agents PM) don't mention them.
- Nobody has a dataset comparing PMs with and without eval skills at the same company and level. Anyone quoting an "eval premium" is guessing.
The sentence I can defend: PM roles at frontier labs that ask for hands-on eval building post base salaries starting around $300k before equity, and senior AI PM roles elsewhere commonly clear $250k in total comp.
Who's saying evals matter, and what they actually said:
- Kevin Weil, then OpenAI CPO, on Lenny's Podcast (April 2025): "Writing evals is going to become a core skill for product managers." His point around it: a model that is right 99.95% of the time and one that is right 60% of the time need completely different products. That is a forecast, and it is about product design, not pay.
- Garry Tan (Feb 2025): "Evals are emerging as the real moat for AI startups." About startup defensibility.
- Greg Brockman (Dec 2023): "evals are surprisingly often all you need."
- A line attributed to Anthropic CPO Mike Krieger ("writing evals is probably the most important thing") circulates through Lenny Rachitsky's April 2025 post. I couldn't find the original venue, so I treat it as secondhand.
- Lenny's Podcast billed its episode with Hamel Husain and Shreya Shankar as "Evals are the new PRDs." Worth noting they teach the most popular paid evals course (5,000+ students at $4,200 as of October 2026), so the signal comes with an incentive.
There is also a dedicated eval job market. One specialist board tracked 1,119 open evaluation, post-training and red-teaming roles at 176 companies on 4 October 2026, median top-of-band $330,000. Almost all of them are engineering and research roles. For PMs, evals are a skill inside the job, not the job title.