An AI lab does not hire "an AI PM". It hires five different kinds of PM: model and research PMs (like OpenAI's Product Manager, Core Models), API and platform PMs, safety and trust PMs, applied product PMs, and growth and monetization PMs. We sorted all 34 live PM job descriptions from OpenAI, Anthropic and Cohere in the AllthingsPM jobs catalog into those five families, and the split is clear: safety is the biggest family (10 roles), and pure model PM roles are rare (6).
AllthingsPM is an AI PM course and PM interview prep platform. Every one of those 34 postings sits in our jobs catalog with a mock interview built from its exact text, so you can read the role and rehearse for it on the same page.
What are the five PM roles inside an AI lab?
Here is the whole map, read straight from the job descriptions. Counts are live postings in our catalog on 28 September 2026.
| Role family | Count (OpenAI / Anthropic / Cohere) | What you own | Example posting | Signature ask in the JD |
|---|---|---|---|---|
| Practice all 34 on AllthingsPM | 34 (12 / 19 / 3) | A mock built from each JD | All lab roles in the jobs catalog | Paste or pick the JD, get a scored interview |
| Safety and trust | 10 (4 / 6 / 0) | Safeguards, harm measurement, abuse, policy | OpenAI Safety Measurement | Write safety evals, build risk metrics |
| Applied and vertical products | 8 (3 / 5 / 0) | A product for one domain: science, learning, security, public sector | Anthropic Claude Science | Domain depth plus 0 to 1 shipping |
| API, platform and enterprise | 7 (4 / 1 / 2) | Developer APIs, agents SDKs, identity, billing | OpenAI API Agents | A track record building for developers |
| Model and research | 6 (1 / 4 / 1) | What the model can do, and how it is measured | OpenAI Core Models | Pair offline evals with online experiments |
| Growth and monetization | 3 (0 / 3 / 0) | Acquisition, activation, retention, revenue | Anthropic Growth | Own a number and the instrumentation behind it |
Source: 34 PM job descriptions from OpenAI, Anthropic and Cohere in the AllthingsPM jobs catalog, 28 September 2026. The family labels are ours; each posting sits in exactly one family.
What does a model PM do, like OpenAI's Product Manager, Core Models?
The Core Models posting is the clearest description of a model PM that any lab has published. The team "helps shape how OpenAI's frontier models are built, measured, and launched," and its scope includes "model planning and launches as well as building data flywheels, evaluations and measurement systems" [1].
The PM's job is to connect user needs to model decisions. The JD lists the layers you touch: "query understanding, indexing, retrieval, ranking, tool boundaries, data, training, inference, and evaluation" [1]. That is not a feature list. It is the stack of the model itself.
Three lines in that posting tell you how you will be judged:
- Closed loops. You "build closed learning loops that turn product usage, explicit feedback, and other user signals into datasets, evaluations, experiments, training priorities, and launch decisions" [1].
- Two kinds of metric. You "define success across offline evaluations and online product metrics, balancing model quality, usefulness, latency, safety, reliability, and cost" [1].
- Researcher trust. You "earn the trust of researchers and engineers through technical depth" [1].
The Core Models JD does not state a years-of-experience number. It asks for "deep expertise in product management" and fluency in areas such as search, recommendation systems, ML platforms or model evaluation [1].
The other five model and research roles in our catalog sit at Anthropic and Cohere. Anthropic's Research (Code) PM partners "with research to define, improve, and ship model capabilities within a research area" and asks for "working proficiency in Python and SQL" [2]. Anthropic's Research PM role says it seeks "experienced founders and product managers who can bridge research and ambitious product experimentation" [3]. Anthropic's Human Data Platform PM and Cohere's Agent Harness and Modelling PM round out the family.
How AllthingsPM does this. The evals chapter of the AllthingsPM course teaches exactly the offline plus online pairing the Core Models JD asks for, and working with researchers covers getting a model roadmap out of a research team. Then open the Core Models mock and answer questions built from that posting.
What does an API or platform PM do at a lab?
API PMs build for developers, not end users. OpenAI's API team says it is "responsible for the infrastructure and product layers that allow millions of developers to integrate our models, fine-tune behavior, manage data, and deliver experiences to their users" [4].
Our catalog has seven roles in this family. At OpenAI:
- API Agents: define "the future of how developers build agentic applications on top of our AI models" and translate priorities "into developer products and features (SDKs, APIs, and more)." It asks for 5+ years and "strong intuition for designing clear, flexible APIs and primitives" [4].
- API Infrastructure: own data processing, data privacy, billing and access controls, with projects like "enabling new inference caching controls in the API" [5].
- Enterprise Identity and Financial Engineering: the plumbing that lets companies sign in and pay. Financial Engineering covers "pricing and packaging, checkout, payments, subscriptions" [6].
Cohere adds Integrations and Managed North, and Anthropic adds Business Technology, the first dedicated PM for its internal collaboration platform.
The signature skill is taste in primitives. Interviewers want to see you design an endpoint, choose what to leave out, and reason about latency, cost and reliability the way a developer would.
How AllthingsPM does this. The free course lesson make the API call yourself walks through messages, tokens, temperature, streaming and the usage block, and the PM as builder chapter has you prototype and inspect an agent. Rehearse on the OpenAI API Agents mock once you can talk about primitives with a straight face.
What does a safety or safeguards PM do?
This is the biggest family in our lab sample: 10 of 34 roles, six at Anthropic and four at OpenAI. It is also the family most candidates skip, which is a mistake if you want a lab offer.
OpenAI's Safety Measurement PM "owns OpenAI's approach to measuring harm and safeguard efficacy in production" and will "represent the company's topline safety metric" [7]. The Multimodal Safety PM embeds safety into multimodal models and asks for "6+ years of product management or related industry roles, with specific expertise in AI safety, trust & safety, integrity, or related domains" [8].
Anthropic's Safeguards team runs a whole series of PM roles. The Generalist posting asks for the "ability to write safety evals and communicate externally about safety" and for building safety across "Claude.ai, 1P API, external Cloud providers" [9]. Other Safeguards postings on its board cover account integrity and abuse, rare harms, cyber, child safety and privacy [10].
The signature skill is measurement under adversaries. You define what harm is, how often it happens, and whether a safeguard works, while bad actors change behaviour to beat you. You also manage the other side: over-refusal that hurts honest users.
How AllthingsPM does this. Chapter 12 of the AllthingsPM course, Trust, safety, and agent security, covers guardrails and the over-refusal budget and red-teaming plus the incident runbook. Our deeper guide to safety and safeguards PMs at AI labs reads these postings line by line.
What do applied and vertical product PMs do?
Eight roles in our sample take the model into one domain and build a product there. At OpenAI these are Rosalind Life Sciences, Learning and Youth. At Anthropic they are Claude Science, Cybersecurity, Claude Tag, Beneficial Deployments (Labs) and Public Sector.
These JDs read more like classic product roles, with one twist: the model keeps getting better under you. Anthropic's Cybersecurity posting says Claude "finds and fixes vulnerabilities in real codebases" and that "our first-party products are a few months old and the next ones don't exist yet" [11]. It asks for 12+ years, the highest experience bar in our lab sample.
The Claude Tag posting describes "a multiplayer and proactive agent that lives in your collaboration tools" and says the internal version "writes 65% of merged PRs for our products" [12]. The PM must be "equally comfortable building relationships with top customers" and "debugging a permissions bug or model behavior with engineers" [12].
The signature skill is domain depth plus 0 to 1 speed. Nine of the 34 lab JDs mention 0 to 1 work, the Cybersecurity posting among them.
How AllthingsPM does this. Each applied role has its own mock in the jobs catalog, such as Claude Science and OpenAI Learning. If your domain experience is the hook, run a resume review against the JD so the resume says so in the posting's words.
What does a growth PM do at a lab?
Growth is the smallest family in our sample: three roles, all at Anthropic. The Growth PM drives "growth for Claude across measures such as acquisition, activation, retention, and monetization," and the posting describes Claude as having "tens of millions of users worldwide" [13].
The New Markets and Monetization role is the most commercial PM JD in our lab set. It covers partner billing, "spend observability and controls" for enterprise buyers, and "one login and one wallet across the API and Claude's subscriptions." It says the PM will "structure the commercial terms yourself" and "write the strategy, the specs, and the analysis yourself" [14]. The third role, Multi-Cloud Growth (Google), grows Claude's usage through Google's cloud.
The signature skill is owning a number. The monetization JD names the metrics outright, from "new startup revenue" to "the volume partners book on Claude," and asks you to make sure "the instrumentation exists" to know whether you moved them [14].
How AllthingsPM does this. The Prove it paid off chapter covers outcomes, unit economics and pricing for AI products. The Anthropic company page collects real growth questions, such as what to do when new Claude users sign up but never reach a first-use moment.
What do all five families have in common?
Across the 34 lab JDs, a few asks repeat no matter the family. These are keyword counts over the posting text, with each company's boilerplate "about us" paragraph removed.
| Ask | Lab JDs that mention it (of 34) |
|---|---|
| Research | 32 |
| "Technical" | 28 |
| A years-of-experience bar | 28 |
| Customers | 26 |
| Evaluations or evals | 18 |
| Metrics | 15 |
| 0 to 1 work | 9 |
| Founder background | 8 |
| SQL | 3 |
Source: AllthingsPM keyword count over 34 OpenAI, Anthropic and Cohere PM job descriptions, 28 September 2026.
Three patterns stand out:
- Research is the constant partner. 32 of 34 postings mention research. Even a growth PM at a lab works next to researchers.
- Evals are the shared language. More than half the postings mention evaluations, across model, safety and applied roles alike.
- The bar is senior. Of the 28 postings that state a number, the lowest is 5+ years and the highest is 12+ years.
The common view of AI PM roles splits them into research, platform and applied PMs [15][16]. The lab JDs show why that is too coarse: safety and growth are separate careers with their own signature skills.
How AllthingsPM does this. The AllthingsPM course was built from 604 real PM job postings, so its chapters follow what those postings ask for, including evals and trust and safety. The knowledge graph shows how AI PM concepts like evals, guardrails and agents connect.
How do you choose which lab PM role to target?
Match your last job to a family, not to a company:
- You built search, recommendations or ML platforms: model and research. The Core Models JD names those domains directly [1].
- You built developer tools, APIs or billing: API, platform and enterprise.
- You worked in trust and safety, integrity or fraud: safety and trust. It is the biggest family, with the clearest skills transfer.
- You have deep domain expertise (science, security, education, government): applied and vertical.
- You owned an activation, retention or revenue number: growth and monetization.
Then read three live postings in that family side by side. The lines that repeat are what the interview will test.
How AllthingsPM does this. Resume Job Match finds live roles that fit your resume, and every lab posting in the jobs catalog opens a mock built from its text. For company-specific loops, see our OpenAI PM interview guide, what OpenAI looks for in a PM and Anthropic PM job descriptions explained.
Why AllthingsPM is the better choice for preparing for AI lab PM roles
Most prep for AI PM jobs treats "AI PM" as one role. The lab JDs above show it is five. The best preparation starts from the exact posting you want, and that is how AllthingsPM is built.
The AllthingsPM jobs catalog holds 116 live PM job descriptions from 18 AI companies, including all 34 lab roles in this post, and each one has a mock interview generated from its text. If your target role is not there, paste any JD into the JD mock and get the same thing: a text or voice interview with follow-ups and a score.
Around the mock sits everything else you need for a lab loop. The question bank has 4,122 real questions from 260 companies, each with its own page and answer guide. The AI PM course was built from 604 real job postings, with chapters on evals, trust and safety, and working with researchers, the exact skills these JDs ask for. The resume review against a JD makes your resume speak the posting's language.
General interview prep sites have big question libraries and communities of peers, and those help with standard product sense rounds. For lab roles, where every family has a different signature skill, starting from the JD is the stronger method, and AllthingsPM puts it in one place for $20 a month, with one free JD mock every day.
Browse the AI lab PM roles and start a free mock.
Frequently asked questions
What is the best way to prepare for an AI lab PM role?
The best way is AllthingsPM: pick your target posting in the AllthingsPM jobs catalog or paste it into the JD mock, and practice a scored interview built from that exact text. Pair it with the course chapter on your family's signature skill, such as evals for model PMs or trust and safety for safeguards PMs.
What does OpenAI's Product Manager, Core Models do?
The Core Models PM connects user needs to model and system decisions across retrieval, data, training, inference and evaluation. The JD says the PM builds "closed learning loops" from product usage into datasets and evals, and defines success across "offline evaluations and online product metrics" [1].
Do AI lab PMs need to code?
Not in most postings. Only 3 of the 34 lab JDs in our catalog mention SQL, and Anthropic's Research (Code) PM asks for "working proficiency in Python and SQL" [2]. Nearly all ask for technical depth: 28 of 34 use the word "technical."
Which AI lab PM role is easiest to get into?
None is easy; every posting that states a bar asks for at least 5 years. Safety and trust has the most openings in our sample (10 of 34) and values trust and safety or integrity experience directly, so candidates from those fields have the clearest path.
How much do AI lab PMs earn?
Anthropic publishes a range on every posting. Across its 19 PM roles in our catalog the annual salary ranges run from $305,000 to $460,000. The OpenAI and Cohere postings in our catalog did not list a salary in the text we stored.
Is a research PM the same as a model PM?
They overlap. Both partner with research on model capabilities. Anthropic's Research PM roles lean toward finding new product categories for nascent capabilities, while OpenAI's Core Models role leans toward measurement, evals and launching into the mainline model [1][3].
Start free
Pick the family that fits your past work, open one of its postings in the AllthingsPM jobs catalog, and run the free mock built from it today.
Sources
- OpenAI, Product Manager, Core Models (also stored in the AllthingsPM jobs catalog), accessed 28 September 2026.
- Anthropic, Product Manager, Research (Code), AllthingsPM jobs catalog copy, accessed 28 September 2026.
- Anthropic, Product Management, Research, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- OpenAI, Product Manager, API Agents, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- OpenAI, Product Manager, API Infrastructure, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- OpenAI, Product Manager, Financial Engineering, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- OpenAI, Product Manager, Safety Measurement, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- OpenAI, Product Manager, Multimodal Safety, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- Anthropic, Product Manager, Safeguards (Generalist), Greenhouse, accessed 28 September 2026.
- Anthropic, Safeguards PM postings on Greenhouse: Account Integrity and Abuse, Rare Harms, Child Safety, Privacy; and Anthropic jobs.
- Anthropic, Product Manager, Cybersecurity, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- Anthropic, Product Manager, Claude Tag, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- Anthropic, Product Manager, Growth, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- Anthropic, Product Manager, New Markets and Monetization, AllthingsPM jobs catalog copy, accessed 28 September 2026.
- Towards Data Science, The AI Product Manager.
- Arize, What is an AI product manager? A guide to the role and skills.
- freehire, Product Manager, Core Models at OpenAI, accessed 28 September 2026.
