Ask these six questions first
- What are the exact round names and durations?
- Is coding in a blank editor, an existing codebase, a notebook, or a take-home?
- Which AI tools are allowed in each round?
- Is there a project presentation, research discussion, or work-sample review?
- Which domain will the technical depth target?
- What artifact or behavior does each interviewer score?
Current process evidence
OpenAI
Prepare for a role-specific work sample and a long final loop. Do not assume one standard engineering or research process.Officially confirmed
- Application and resume review, then an introductory recruiter or hiring-manager call.
- One or more skills-based assessments. Official examples include pair coding, take-home projects, and technical tests.
- A final loop of roughly 4 to 6 hours with 4 to 6 people over 1 or 2 days.
- Engineering work is judged on solution design, code quality, performance, tests, communication, and collaboration.
Currently reported
- Current engineer reports add practical multi-part coding, architecture or system design, a technical project presentation, and separate leadership or collaboration discussions.
- A June 2026 report describes a beta agentic-coding round in an existing codebase. Treat this as a pilot, not a universal round.
- Some research-engineering reports include ML debugging, statistics, probability, or information-theory work. Team scope matters more than the company label.
AI policy
Official guidance says tool rules vary by interview. Some formats intentionally allow AI, while others assess independent problem solving without it. The preparation material for each assessment states what is allowed. Ask the recruiter when anything is unclear.
Practice only what matches
Caution: OpenAI explicitly says experiences vary by team. Ask for the assessment format, editor, AI policy, presentation requirement, and domain before choosing a drill.
Sources and evidence level
- OpenAI interview guide ↗OfficialChecked August 28, 2026. Stages, assessment examples, assessment-specific AI rules, final-loop length, and engineering criteria.
- interviewing.io OpenAI process ↗Method-based reportChecked August 28, 2026. Practical coding, presentation, system design, and agentic-coding pilot.
- Frontier Evals and Environments role ↗OfficialChecked August 28, 2026. Current work in RL environments, graders, synthetic data, and evaluation systems.
Current process evidence
Anthropic
Expect practical Python, explicit values evaluation, and team-dependent work samples. Performance and research tracks can look nothing like a standard SWE loop.Officially confirmed
- Technical interviews use live coding tools such as Colab and CodeSignal. Looking up syntax is allowed, but fluency still matters.
- AI is not allowed in take-homes or live interviews unless the instructions explicitly permit it.
- The performance team has used timed take-homes based on a simulated accelerator and now publishes the original challenge.
- The Fellows process includes an application and reference check, technical assessments and interviews, and a research discussion.
Currently reported
- A June 2026 report describes a progressive CodeSignal task, hiring-manager project depth, practical coding, system design, and a standalone company-values round for many SWE candidates.
- A 2025 Fellows candidate reported a short research brainstorm and a longer black-box model investigation followed by a presentation.
- Round count and composition vary substantially across software, research, safety, and performance roles.
AI policy
Complete take-homes and live interviews without AI unless Anthropic explicitly says otherwise. The public performance challenge is an exception where AI use is allowed, but weakening tests invalidates the result.
Practice only what matches
Caution: Do not infer a full-time Anthropic loop from the Fellows process or the performance team challenge. Use each as evidence for a format only when the recruiter confirms it.
Sources and evidence level
- Anthropic careers ↗OfficialChecked August 28, 2026. Live coding tools and hiring philosophy.
- Candidate AI guidance ↗OfficialChecked August 28, 2026. AI rules for applications, take-homes, and live interviews.
- AI-resistant technical evaluations ↗OfficialChecked August 28, 2026. Performance take-home design, simulated accelerator, profiling, and time limits.
- Anthropic Fellows Program ↗OfficialChecked August 28, 2026. References, technical assessment, research discussion, and current workstreams.
- interviewing.io Anthropic process ↗Method-based reportChecked August 28, 2026. Current reported SWE stages, CodeSignal format, values round, and system design.
Current process evidence
Google DeepMind
The process is role-specific, but Research Engineer preparation still needs executable coding, mathematical ML depth, model design, and a defensible project history.Officially confirmed
- A 30-minute recruiter introduction, with a possible hiring-manager interview.
- Two or three skills interviews calibrated to the role.
- Final conversations with team leads and leadership through the lens of team plans, culture, mission, and values.
- Candidates receive role-specific preparation because the exact steps differ by position.
Currently reported
- A June 2026 synthesis of candidate reports describes two executable coding rounds followed by mathematical ML fundamentals and ML design for many Research Engineer loops.
- Reported ML prompts reward derivation and intuition, not definition recall. Constraints often change after the first answer.
- General algorithms remain a real requirement for some roles, but this site deliberately leaves that curriculum to dedicated resources.
AI policy
The official interview guide permits AI for preparation, but says not to use AI during live interviews or interview tasks unless the instructions explicitly allow it. Role-specific recruiting instructions control.
Practice only what matches
Caution: Do not substitute generic Google interview guidance for DeepMind role instructions. Research Engineer, Research Scientist, and product-facing ML roles use different mixes.
Sources and evidence level
- Google DeepMind careers ↗OfficialChecked August 28, 2026. Initial, skills, final, and decision stages.
- Google DeepMind interview guide ↗OfficialChecked August 28, 2026. Interview stages, preparation, and AI-use policy.
- 2026 Research Engineer synthesis ↗Method-based reportChecked August 28, 2026. Reported coding, ML fundamentals, ML design, and team interviews.
Current process evidence
Meta AI
AI-assisted coding and design are now first-class formats for selected roles. The signal is control, review, debugging, and judgment, not prompt volume.Officially confirmed
- Select roles now include an authorized AI assistant inside CoderPad during technical interviews.
- The assistant is built into CoderPad and offers Claude, ChatGPT, Gemini, and Meta models.
- For selected AI-native design interviews, candidates use Mermaid Markdown in CoderPad with the authorized assistant available.
- Meta recommends practicing in the provided environment and getting comfortable reading, debugging, and extending existing code.
Currently reported
- A 2026 report describes a 60-minute, multi-file onsite round that replaces one coding interview for some roles.
- Candidates are judged on planning, codebase navigation, bounded delegation, critical review, testing, and explanation.
- This format does not remove the need for unaided algorithmic or role-specific technical rounds.
AI policy
Use only the assistant inside the authorized interview environment. Outside AI tools are not permitted. Select the model in the environment and remain responsible for every change.
Practice only what matches
Caution: Meta says selected roles use the format. Confirm whether your loop includes AI-native coding, AI-native design, both, or neither.
Sources and evidence level
- Meta hiring process and AI FAQ ↗OfficialChecked August 28, 2026. AI expectations, models, languages, design environment, and preparation.
- AI-assisted coding report ↗Method-based reportChecked August 28, 2026. Reported duration, onsite placement, and multi-file task shape.
Current process evidence
xAI
The public process is sparse but unusually direct: technical staff screen the application, ask short technical questions, and then go deep on relevant expertise.Officially confirmed
- Applications are evaluated by technical team members rather than recruiters for assessment.
- The screening interview covers background, fit, and short technical questions.
- Technical interviews examine complex problem solving and critical thinking in the candidate’s domain.
- Applications ask for a statement of exceptional work in 100 words or fewer.
Currently reported
- Public evidence is not strong enough to publish a stable ML or Research Engineer round-by-round recipe.
- Use the job description and recruiter guidance to identify model training, infrastructure, product, or domain depth.
AI policy
No universal public interview AI rule was found. Ask before using any tool that is not explicitly provided.
Practice only what matches
Caution: Avoid SEO question lists that claim exact xAI prompts without a disclosed source. The official process supports format-level preparation, not a question dump.
Sources and evidence level
- xAI careers ↗OfficialChecked August 28, 2026. Technical review, screening, technical interviews, and application artifact.
- Model Training role ↗OfficialChecked August 28, 2026. Current role scope and exceptional-work application prompt.
Cross-lab changes worth preparing for
Work samples are replacing trivia
Expect existing code, failing tests, traces, notebooks, or multi-stage specifications. The task is longer, and the scoring surface includes judgment and verification.
AI policy is now part of the format
Select Meta roles use an authorized assistant. Anthropic and DeepMind prohibit interview-time AI unless a round explicitly permits it. OpenAI says the rule varies by assessment. Never infer permission.
Senior candidates defend artifacts
Technical presentations, project depth, research discussions, exceptional-work statements, and references test whether the claimed scope is real.
Questions probe one level deeper
Practical coding still needs algorithms. ML theory still needs derivation. Design still needs failure behavior. AI-generated surface fluency raises the value of follow-up depth.
Maintenance and corrections
The internal source registry is reviewed quarterly. Company pages and recruiter instructions outrank summaries. Method-based and first-person reports are labeled because they are not policy.