Anthropic does not have a new-grad program, does not run internships, and hires almost nobody straight out of school: a June 2026 analysis of 1,680 Anthropic engineers' LinkedIn profiles found only 50 with under three years of experience. That is the honest starting point. The rest of this guide is about the three doors that do exist in 2026, how the Research Engineer and Software Engineer loops differ, and which of those doors are closed to you on a student visa.
Who they actually hire
The careers page says what it wants you to hear: "We care about what you can do, not where you learned to do it," and "about half have PhDs, but plenty of brilliant colleagues never went to college" (anthropic.com/careers, retrieved Oct 2026). The LinkedIn data says what actually happens.
| Signal | What the data shows | Source |
|---|---|---|
| Experience at hire | Median 12.2 years; middle 50% between 8.8 and 16.5 years | AI Engineering, 1,680-profile analysis, Jun 2026 |
| Juniors | 172 of 1,680 under 6 years; 50 under 3 years | same |
| PhDs among engineers | 13.7% (about 1 in 7); careers page's "about half" covers all technical staff including researchers | same; anthropic.com/careers |
| Most common prior background | Infrastructure engineering on 40.4% of resumes; ML/DL 28.3%; RL only 3.3% | same |
| Where juniors came from | 50% listed an elite internship (16 Meta, 10 Google, 6 DeepMind); 9% from quant firms; 6% from alignment fellowships (MATS, SERI, Redwood, ARC) | same |
| Open roles | 644 on the Greenhouse board; no title containing "intern", "new grad" or "graduate" | job-boards.greenhouse.io/anthropic, 13 Oct 2026 |
| Headcount growth | 686 engineers added in 2025, 455 more by June 2026; median tenure 10 months | AI Engineering, Jun 2026 |
Two things follow. First, the median Anthropic engineer is an infrastructure veteran from Google, Meta, Amazon, Microsoft or Stripe (the top five prior employers in that dataset), not a researcher. Second, the fifty juniors who did get in substituted a hard credential for the missing years: a top internship, a competitive-programming record, published research, or a named alignment fellowship. The analyst's own archetype is "an MIT graduate with an IOI silver medal, a 2900+ rating on Codeforces". The junior cohort was also more international than the senior one, with Tsinghua (7), Shanghai Jiao Tong, NUS and ETH Zürich on the school list.
The company is also mid-sprint. Bloomberg reported on 1 Oct 2026 that Anthropic is aiming to list as soon as mid-November at a valuation of up to $2 trillion, and Reuters, citing the S-1, put 2025 revenue at roughly $4.6 billion, up 12x (both via Yahoo Finance, 1 Oct 2026). Headcount is growing fast, but the hiring bar described above has not loosened for juniors.
The three doors, and which ones a visa closes
| Door | Length and pay | Who is eligible | Visa | Converts to full-time? |
|---|---|---|---|---|
| Full-time RE or SWE posting | Permanent; RE RL Engineering $500,000-850,000, RE Interpretability $315,000-560,000, Full-Stack SWE RL $300,000-405,000 base (Greenhouse, Aug-Oct 2026) | "Bachelor's degree or an equivalent combination of education, training, and/or experience"; years "will correlate with the internal job level" | "Yes. We sponsor visas and green cards for eligible roles" (careers page); postings add "we aren't able to successfully sponsor visas for every role and every candidate" | It is the full-time role |
| Anthropic Fellows Program | 4 months; $3,850/week (£2,310 / C$4,300); ~$15K/month compute; next cohort January 2027 | Python fluency; no PhD or ML experience required; full-time availability | "We are not currently able to sponsor visas for fellows"; must hold US, UK or Canada work authorization | "25-50% of fellows received a full-time offer" in previous cohorts; "over 80% of fellows produced papers" (Greenhouse posting) |
| Claude Corps | 12 months; $85,000/year, employed by CodePath; cohorts of ~100 (19 Oct 2026), 300 (8 Mar 2027), 600 (Aug 2027) | Anyone 18+, no minimum education; current full-time students discouraged | Must be "authorized to work in the US without employer sponsorship"; F-1 CPT, OPT and STEM OPT "do not qualify" (Claude Corps FAQ) | Places you in a nonprofit, not at Anthropic |
For an international student this table is brutal. Claude Corps is closed outright: OPT does not count. The Fellows Program is open only if you already have work authorization that does not depend on Anthropic, which in the US means OPT or STEM OPT (the Fellows FAQ excludes sponsorship, not student work authorization) and in the UK means a Graduate visa. The only door with sponsorship attached is the full-time posting, which is also the one with the 12.2-year median. So the realistic international-student route is Fellows on OPT or the UK Graduate visa, then conversion, and the posting says a quarter to a half of fellows convert.
RE vs SWE: the same gate, different rooms
Both tracks share the front half of the process and the two filters that reject most people. The back half is where they split.
| Stage | Research Engineer | Software Engineer | Source |
|---|---|---|---|
| Recruiter call | 30 min; mission alignment is tested here and "candidates can fail it" | same | interviewing.io, 2026 |
| Coding screen | 90-min CodeSignal take-home most common, 60-min live for some roles; one engineer: 2 hours, four levels, black-box tests, pass all to advance | same | interviewing.io, 2026 |
| Hiring manager | 45-60 min on past projects; may include a code review | same | Aced, Oct 2026 |
| Onsite (4-5 hours, five sessions) | ML coding and debugging, systems design, research discussion, paper discussion, values | coding, system design, second role-specific coding, values, plus a technical project presentation on some teams | Sundeep Teki, 2026; interviewing.io; Aced |
| References | "Often conducted during the loop rather than at the end" | same | Sundeep Teki, 2026 |
| Total time | 6-10 weeks (Teki) | 3-4 weeks (interviewing.io) to 4-6 weeks (Aced) | as listed |
Three details matter more than the round names.
The CodeSignal gate is pass-all-or-nothing. The engineer quoted by interviewing.io had to pass every black-box test at a level before the next level unlocked. The example task is not an algorithm puzzle: implement a bank with multiple transaction types, with the spec getting more complicated at each level. Candidates "often run out of time". Anthropic's candidate AI guidance (Jul 2025) is explicit that take-homes are done "without Claude unless we indicate otherwise" and live interviews are "all you". Aced reports that the performance-engineering team's separate take-home, which since early 2024 asked candidates to optimise code for a simulated accelerator, was redesigned after Claude Opus 4.5 matched the strongest human applicants on it; treat that as a signal that any task a model can solve will be rotated out.
Scoring well is not enough. A Blind poster in March 2025 reported a CodeSignal rejection email that said they "scored well but didn't have the right experience for the role". Another Blind account, from a Google engineer who interviewed for Research Engineer in July 2024, had the CodeSignal skipped because of a closely related background, passed an ML round on binary classification with LLMs, then was rejected after a hiring-manager call "with a generic reason and no specific feedback". For a new grad, the "right experience" problem is the whole problem, which is why the fifty juniors in the LinkedIn data all carried a credential that reads as experience.
The values round rejects people who passed everything else. The same Google engineer wrote that a recruiter told them candidates often pass the technical rounds and fail the final behavioural round, which is heavily safety-focused. In the Blind interview megathread (retrieved Oct 2026), one onsite candidate was quizzed on Dario Amodei's book and had not read it. Aced's guide says interviewers for this round often come from non-engineering teams and that "uncritical enthusiasm and blanket criticism both read poorly". The round is not a loyalty test; it is a test of whether you have an actual position.
What gets new grads rejected
1. Applying as a "pure researcher" with no shipped systems. 40.4% of Anthropic engineers came from infrastructure; RL appears on 3.3% of resumes (AI Engineering, Jun 2026). A resume built around coursework and one Kaggle notebook competes against people who ran GPU clusters.
2. Partial credit on the take-home. Levels unlock only on full correctness (interviewing.io). A clean level 2 beats a half-done level 4, and the clock is the enemy.
3. Using Claude on the take-home. Prohibited unless stated (anthropic.com/candidate-ai-guidance). The redesigned performance take-home shows the team watches for model-shaped solutions.
4. Applying to the Fellows cohort without work authorization. The posting says it cannot sponsor fellows; the application is wasted.
5. No position on Anthropic's own arguments. Being asked about Dario's writing with nothing to say ended at least one onsite (Blind megathread). Read the Constitutional AI work and one recent Interpretability paper, and come with a disagreement you can defend.
6. Ghosting is real; do not read it as a decision. An ex-Databricks candidate on Blind passed the full loop, submitted references, and was told two weeks later there was "no headcount". Keep other processes running; the hiring-manager round on some teams happens before the onsite, so you can lose weeks waiting.
A 4-week plan
- Week 1, pick the lane and the credential. If your strongest artefact is a training-infra or serving project, apply SWE (Full-Stack SWE RL is the one posting with a plain "Software Engineer" title on the board). If it is a reproduced paper with ablations, apply RE, and in parallel apply to the January 2027 Fellows cohort if you hold OPT, STEM OPT or a UK Graduate visa. Put independent research, a blog post or open-source contributions at the top of the resume; the careers page says to.
- Week 2, the gate. Daily 90-minute timed sessions on multi-level CodeSignal-style tasks: a bank with transaction types, an LRU cache with evolving constraints, a job queue with retries. Correctness first, refactor as the spec grows, no AI.
- Week 3, the lane-specific room. RE: multi-head attention and a training loop from scratch in PyTorch, then "loss spikes during pretraining, is it data, learning rate or hardware" style debugging (Sundeep Teki, 2026). SWE: design a Claude chat service and a request-batching API for serving LLMs, which interviewing.io lists as recurring system-design prompts.
- Week 4, the values room and references. Write one page on where you disagree with Anthropic's safety approach and why you still want to work there. Warn two references that Anthropic calls mid-loop, not at offer stage.
Practice
- Anthropic Research Engineer mock interview: the four-level take-home format, PyTorch-from-scratch round and values round.
- Anthropic Software Engineer mock interview: the practical coding and system-design loop without the ML rounds.
- Polish your resume for Anthropic: rewrite research bullets into shipped-systems language, which is what 40% of their engineers' resumes look like.
- OpenAI vs Anthropic vs DeepMind interview guide: the cross-lab comparison if you are applying to all three.
- New Grad 2027 application tracker: the labs and quant firms that do run graduate programs, for the parallel applications you should be making.
Sources
- Anthropic, Careers page (interview tools, PhD share, internships, visa policy), retrieved Oct 2026
- Anthropic Greenhouse, Research Engineer, RL Engineering ($500,000-850,000), updated 1 Oct 2026
- Anthropic Greenhouse, Research Engineer, Interpretability ($315,000-560,000, 5-10+ years), updated Aug 2026
- Anthropic Greenhouse, Full-Stack Software Engineer, Reinforcement Learning ($300,000-405,000), updated Aug 2026
- Anthropic Greenhouse, Anthropic Fellows Program, ML Systems & RL (Jan 2027 cohort, stipend, no visa sponsorship, 25-50% conversion), retrieved Oct 2026
- Anthropic, Claude Corps fellows FAQ (cohorts, $85,000, work-authorization rule), retrieved Oct 2026
- Anthropic, Candidate AI guidance, Jul 2025
- Anthropic Greenhouse job board (644 open roles), 13 Oct 2026
- AI Engineering (Medium), "What Kind of Engineers Does Anthropic Actually Hire? Insights from 1680 Resumes", 16 Jun 2026
- Levels.fyi, Anthropic Software Engineer salaries (55 submissions), updated 13 Oct 2026
- interviewing.io, Anthropic's Interview Process & Questions, 2026
- Aced, Anthropic Software Engineer Interview Guide, updated Oct 2026
- Sundeep Teki, Anthropic Research Engineer Interview 2026
- Blind, Anthropic Interview Megathread, retrieved Oct 2026
- Blind, "Anthropic interview experience" (Google engineer, Research Engineer loop), Jul 2024
- Blind, "Early 2025 Anthropic CodeSignal rejects", Mar 2025
- Yahoo Finance (citing Bloomberg and Reuters), "Anthropic reportedly looking to IPO as early as mid-November", 1 Oct 2026
Frequently asked questions
- Does Anthropic hire new grads in 2026?
- Rarely, and never through a named program. Anthropic's careers page says it does not currently offer internships, and the Greenhouse board had no intern or new-grad titles among 644 open roles on 13 Oct 2026. A June 2026 analysis of 1,680 Anthropic engineers' LinkedIn profiles found only 50 with under three years of experience; the median hire had 12.2 years.
- Research Engineer or Software Engineer: which is easier for a new grad?
- Neither posting sets a years-of-experience number; both say a Bachelor's or equivalent and that experience 'will correlate with the internal job level'. The Interpretability RE posting asks for 5-10+ years, while the RL Engineering RE posting lists no number. The practical difference is the loop: RE adds ML coding, debugging and a paper discussion; SWE is practical coding plus system design. Pick the lane where your strongest artefact lives.
- What does Anthropic pay a new engineer?
- Per Anthropic's own postings (Aug-Oct 2026): Research Engineer, RL Engineering $500,000-850,000; Research Engineer, Interpretability $315,000-560,000; Full-Stack Software Engineer, RL $300,000-405,000, all base. Levels.fyi (55 submissions, updated 13 Oct 2026) puts the entry-level Software Engineer median at $483K total, $266K base plus $216K stock.
- Does the Anthropic Fellows Program sponsor visas?
- No. The ML Systems & RL Fellows posting says Anthropic is 'not currently able to sponsor visas for fellows' and requires work authorization in the US, UK or Canada. Full-time roles are different: the careers page says Anthropic sponsors visas and green cards for eligible roles, with the caveat that it cannot do so for every role and candidate.
- Can I use Claude in the Anthropic interview?
- Only on the application. Anthropic's candidate AI guidance (Jul 2025) says to write the first draft yourself and use Claude to refine it, to complete take-homes without Claude unless told otherwise, and that live interviews are 'all you'.
- What is the Anthropic CodeSignal assessment like?
- Per interviewing.io (2026), most candidates get a 90-minute CodeSignal take-home; some roles use a 60-minute live assessment. One engineer described a two-hour session with four progressively harder levels judged by black-box tests, where you must pass every test at a level to unlock the next. Example task: a bank with multiple transaction types.