Job search
Anthropic Fellows Program: Choose a Research Project and Check Funding, Location, and Employment Status
Review the research contribution and current fellowship arrangement separately, with two completed fictional decision records.
The Anthropic Fellows Program is a funded research commitment with a project and mentor, and its current workstreams have different application instructions. Choose a research contribution first, then check the intake, four-month commitment and engagement terms. A fellowship listing on a careers board does not itself establish a permanent staff job.
This guide reviews official public pages checked October 22, 2026. We did not apply, contact mentors, upload private work or test a candidate assessment. The completed cases below are original fictional planning exercises. LandOffer.ai publishes this guide; you can browse recent roles while comparing a research fellowship with an ordinary job search.

Find the current workstream before choosing an application route
The official program announcement describes empirical research aimed at a public output, with fellows helping choose and shape projects alongside mentors. Its current body is dated September 28, 2026, despite the older year in its URL. Read the page itself rather than dating an opportunity from its address.
There is a deadline difference to preserve. The announcement says to apply by October 18 for January consideration; that date had passed at inspection. The current AI Safety & Security listing instead says applications are open for a January cohort with an October 30 deadline. Recheck that specific listing before acting; do not assume the blog's earlier date closes every stream or that one stream's later date applies to all of them.
The ML Systems & Reinforcement Learning listing and Anthropic Institute Economics & Policy listing describe rolling applications for a cohort expected in January 2027. “Expected” is a planning description, not your start-date commitment.
Safety applicants are directed to Constellation's linked form, and the page explicitly says the Greenhouse application is unnecessary. The other two listings direct applicants to their own bottom-of-page forms. Follow the instructions for your selected stream; copying an old application route could leave you preparing material for the wrong destination.
Choose a research contribution you can explain
The three listings describe different kinds of work. Safety and security include empirical model-safety questions and security research. The ML workstream covers systems and reinforcement-learning projects. The Institute workstream includes empirical economics and policy research and communicating unexpected or null findings. Read each detailed description before treating a shared program name as a shared research role.
Start with evidence of something you personally built, investigated or measured. Identify the question, implementation, comparison and limitation. Software experience can support a research contribution without proving that you already know the whole field. Conversely, an interest in a topic does not supply evidence of experimental ownership.
The program uses project selection and mentor matching. A list of potential mentors is useful context, not an agreement that a named person will supervise you. The ML listing explicitly directs applications through the official process rather than to mentors. Prepare a project preference that can change after discussion instead of presenting an inflexible agenda as a confirmed placement.
Worked example: Lea selects a defensible project direction
Lea is a fictional Python infrastructure engineer. She has built a small workload generator, diagnosed inconsistent timing measurements and maintained an open-source test runner. She has read introductory interpretability material but has not trained a large model or conducted economics research. These are invented teaching facts, not an applicant evaluation or a successful fellowship story.
Her completed comparison narrows three possible directions:
| Direction | Evidence Lea can explain | Gap she must retain | Preparation decision |
|---|---|---|---|
| ML systems experimentation | Workload generator, test runner and investigation of timing variance | No demonstrated accelerator-training infrastructure ownership | Investigate this stream first; present a bounded measurement project |
| Mechanistic interpretability | Introductory reading and Python programming | No completed interpretability experiment | Keep as an interest, not a claimed research specialty |
| Economics and policy | Experience documenting technical work | No empirical economics project or suitable research writing sample supplied | Do not relabel an engineering incident note as economics research |
Her choice is provisional: investigate the systems direction because she can explain relevant implementation and measurement work. This is an editorial comparison, not Anthropic's selection decision. A mentor could propose a different question, and the actual program might require capabilities she has not demonstrated.
Her complete project-preference brief reads:
Question: How much does workload ordering change measured completion time in a small reproducible inference benchmark? Contribution I can support: build the Python harness, retain raw measurements, document the environment and investigate variance. Proposed comparison: fixed model and hardware configuration, saved workload set, then compare two ordering rules across repeated runs. Boundary: this is a proposed small benchmark, not experience scaling Anthropic's training systems. Failure criterion: if configuration or workload differences confound the comparison, I will not attribute the timing change to ordering. Deliverable: a runnable harness, methods note and report including inconclusive results. Mentor discussion: whether this question is useful to the stream and how its scope should change. Status: proposed only; no experiment or fellowship application has been executed.
The brief contains a question and a way to detect a weak conclusion. It leaves research direction open for discussion while making Lea's contribution inspectable. She can explain the controlled comparison without giving herself an invented research title.
Her saved next step is equally specific: select one shareable repository example, explain her own changes and write down the variance issue she found. She will not submit confidential employer logs. She also will not claim that this proposed project was requested, approved or supervised by a listed mentor.
Separate your stipend from research resources
Current listings describe four months of full-time research. The Safety listing specifies forty hours weekly and lists a weekly stipend of 3,850 USD, 2,310 GBP or 4,300 CAD, with benefits varying by country. These are listed currencies and amounts, not a comparison of equivalent take-home compensation. Obtain the terms that apply to your arrangement before budgeting.
Safety and ML describe approximately 15,000 USD per month for compute, alongside other research expenses. The Institute listing describes substantial research funding without promising that same numeric allowance. Keep separate budget entries for personal living costs and approved project resources so you can assess the arrangement without counting the same funding twice. Compute funding is not an extra stipend to add to rent money.
Do not convert “four months” into a guaranteed sixteen payments or calculate an annual salary from the headline weekly figure. Actual start and end dates, payment schedule and deductions need their own record. An extension is described as possible, not a confirmed addition to your funded period.
Before accepting an arrangement, ask what the benefits include, who administers payment and what rules govern research expenses. Those questions make the financial commitment concrete without assuming that a careers-page stipend uses the same structure as a staff compensation package.
Verify location and work authorization together
The ML listing requires fellows to have work authorization in the US, UK or Canada and be located in that country during the program. It says fellows must already have or independently obtain the relevant full-time authorization because fellow visa sponsorship is unavailable. This article does not determine an individual's authorization or immigration options.
The location arrangements also differ. Safety describes shared workspaces in Berkeley and London; ML and the Institute describe shared workspace in San Francisco. Their fellow logistics describe openness to remote fellows within the specified countries; this does not confirm an individual remote arrangement. “Remote” therefore does not mean work from any country, nor does a workspace address alone tell you which arrangement has been agreed for you.
Read the fellowship-specific section carefully. The listings include generic staff policies below a notice saying those policies do not apply to the Fellows Program. Do not import the staff visa-sponsorship language or office-percentage policy into your fellowship plan. If you need an exception or a particular attendance arrangement, retain it as unresolved until the program confirms it.

Worked example: Morgan keeps the arrangement on hold
Morgan is a fictional researcher considering the Institute workstream. The exercise stipulates that Morgan currently lives in the US, has not verified authorization for this program, can reserve forty weekly hours for four months, and plans a six-week stay in Spain during that period. Morgan has no engagement agreement, payment calendar or written location confirmation.
The completed record prevents enthusiasm about a funded project from filling those gaps:
| Item | Evidence in this exercise | Recorded decision |
|---|---|---|
| Research route | Institute Economics & Policy listing; January 2027 expected, rolling applications at inspection | Recheck this listing and its own form; do not use Safety's October 30 date as its deadline |
| Time | Forty hours weekly reserved for four months | Candidate planning assumption only; actual dates remain unconfirmed |
| Country and authorization | US residence stipulated; authorization not verified | Residence alone does not establish authorization; resolve separately |
| Spain travel | Six weeks planned outside the listed participation countries | Do not assume remote permission covers the trip; arrangement remains unresolved |
| Personal funding | Published weekly stipend; no applicable payment schedule supplied | Record stipend separately; no take-home or total-period amount established |
| Research funding | Institute promises substantial resources, not a specified personal cash allowance | Do not add 15,000 USD monthly to living money |
| Engagement | No agreement or named employing entity supplied | Do not record “Anthropic employee” or an accepted fellowship |
Morgan's full unsent clarification draft is:
I am reviewing the Institute fellowship arrangement before making commitments. Which entity would engage me and administer the stipend and benefits, and what engagement status would the agreement specify? Could you confirm the actual program dates, payment schedule and applicable country benefits? I also have a planned six-week stay in Spain. Does that conflict with the requirement to be located in an authorized US, UK or Canadian location during participation? I will not assume the remote description permits that travel. Please also distinguish the approved research-expense process from the stipend. I have not accepted an arrangement or resolved my authorization requirements.
This is a complete question draft for an invented scenario, not a message sent to the program or a request to decide Morgan's legal status. The recorded decision is hold: time appears available in the exercise, but location, authorization and engagement terms are not settled. For example, confirmation of program dates would update the time record without resolving Spain travel, work authorization or the employing entity.
Confirm engagement status instead of borrowing old terms
The original pilot announcement explicitly said fellows would not officially be employed by Anthropic. It also described a six-month pilot with older funding terms and now points readers toward the newer announcement. It is historical evidence, not a current contract for every stream.
The current pages reviewed here establish fellowship logistics and distinguish them from staff policies; they are not a signed engagement agreement. They do not conclusively identify the employing or contracting entity and engagement classification for every reader's eventual agreement. Ask for those details directly in the actual terms. A recruiting partner, workspace operator and employing entity need not be interchangeable roles.
Likewise, possible subsequent staff work is a separate outcome. Past-cohort figures describe selected earlier participants and do not establish your chance of an offer. Evaluate whether the research period itself fits your goals without counting on conversion to make an uncertain arrangement acceptable.
Lea's project brief and Morgan's hold record serve different decisions: what research you can contribute and whether the arrangement works. Write the project contribution and unresolved arrangement questions in separate records before committing time or making plans. If you need a conventional engineering role sooner, browse recent roles on LandOffer.ai and verify the employer's current application route separately.
Sources and Further Reading
Official public sources checked October 22, 2026; no application or personalized program process was tested.


