Summer 2026 is the first intern season where every major AI lab is simultaneously running a structured program, competing for the same 200-level PhD students, and paying compensation that rivals full-time packages at mid-tier tech companies two years ago. OpenAI is in San Francisco running 12- to 26-week placements across its Applied Emerging Talent track. Anthropic is converting a portion of its AI Safety Fellows into full-time researchers at a rate its own program page describes as exceeding 40 percent. Scale AI is paying $60 per hour for software engineering interns with a known pipeline to full-time offers. The question for Class of 2026 is no longer which lab has a program — they all do. The question is which ones are actually converting.
The honest answer, drawn from public program disclosures and compensation data filed through Levels.fyi and Glassdoor through early 2026, is that conversion rates vary by a factor of more than two between the most aggressive and most passive pipelines. That spread is not talent scarcity. It is a deliberate organizational posture.
The Conversion Landscape, Lab by Lab
OpenAI runs its intern pipeline under the Emerging Talent brand. The Summer 2026 software engineering internship is a 12-week in-person placement in San Francisco, with a systems research track that can extend to 26 weeks based on team needs and performance. OpenAI does not publish a conversion rate, but the program architecture — named "Emerging Talent," structured around production teams rather than standalone research projects, based physically in Mission District headquarters — is explicitly designed as a talent funnel. The full-time "Emerging Talent" role for 0-to-3-year engineers is the stated downstream destination on the same careers page that lists the internships. That structural link is a stronger signal than any stated percentage. Intern hourly pay, per Glassdoor data submitted through August 2025, runs between $56,530 and $96,067 annualized ($27 to $46/hr), placing OpenAI at the high end of AI lab intern compensation for non-residency tracks.
Anthropic does not run a traditional software engineering internship. Its primary early-career pipeline is the AI Safety Fellows Program: a structured 3-to-6 month funded fellowship paying $3,850 per week plus $15,000 per month in compute funding, targeted at researchers investigating alignment and safety questions. The conversion data Anthropic has published is the clearest of any lab: over 80 percent of fellows from the first cohort produced published research, and more than 40 percent joined Anthropic full-time following their fellowship. That 40-percent-plus figure is a floor, not a ceiling — Anthropic opened a second Fellows cohort for May and July 2026 with expanded capacity, suggesting the program is performing as a hiring engine. The catch is selectivity: the fellows program is PhD-track or equivalent, not a generalist undergraduate pipeline. A rising senior without a research record is not the target population.
Google DeepMind presents a structural split that Class of 2026 needs to understand clearly before applying. The Student Researcher Program — which runs 12 to 24 weeks, pays at the Research Intern rate ($72.12 per hour per Levels.fyi data for that designation; the separate Student Researcher stipend sits materially lower, in the range of $20–$33 per hour per Glassdoor), and places students across DeepMind, Google Research, and Google Cloud AI teams — carries an explicit non-conversion designation on Google's own careers site. The program description states it is "best suited for students who will not be seeking full time employment following this role." That is an unusual and important disclosure. DeepMind's Student Researcher track is built as a credential and research output mechanism, not a hiring pipeline. PhD candidates who want a DeepMind publication credit before going on the academic or industry job market benefit substantially. A graduating senior expecting a return offer should treat it as a separate process requiring a separate application to full-time roles.
The Research Ready Programme — the UK-anchored undergraduate track that places students at DeepMind for 8-week summer placements — operates similarly: high-prestige, low direct conversion. For the US market, DeepMind full-time new-grad hiring runs through standard Google recruiting processes rather than program conversion.
Meta AI operates two parallel tracks with different conversion logics. The Software Engineering intern program — running 12 to 24 weeks, paying $51.35 per hour plus $2,600 per month in housing stipend, per Meta University compensation data published as of 2026 — functions as a direct conversion pipeline. Meta's general intern-to-full-time rate for SWE, as referenced in internal hiring forums and confirmed by multiple Glassdoor reviews, has historically run in the 50-to-65 percent range for performers. The AI Research track at FAIR (Fundamental AI Research) operates differently: Meta's own careers page describes research scientist internships as a "non-convertible role," designed for PhD candidates who are explicitly expected to return to their programs. That said, former FAIR interns who receive positive evaluations are encouraged to re-interview for full-time roles — a common practice that means the conversion pathway exists but requires a second hiring event rather than a direct offer.
The practical implication: if you are a May 2026 undergraduate seeking a Meta AI full-time offer, the SWE intern route is the one with a direct conversion mechanism. FAIR is a credential, not a pipeline.
Microsoft AI is the most transparent of the hyperscaler cohort about conversion intent. Internal data cited across hiring forums, and referenced in our May 9 analysis of rotational programs, indicates 75 to 80 percent of Microsoft Explore participants who perform well receive return-offer invitations. Explore is specifically a first- and second-year undergraduate program, which means the Class of 2026 seniors who ran through Explore in summer 2024 are the cohort most directly in line for full-time offers now. Entry-level AI Engineer roles at Microsoft — a distinct job family within Azure AI and Copilot — carry a median total compensation of $282,000 at current Levels.fyi submissions, significantly above the $161,000-to-$240,000 general SWE new-grad band. That premium reflects Microsoft's willingness to pay for demonstrated AI systems experience at the point of hire.
Scale AI is the most conversion-direct of the non-hyperscaler labs. The company's Summer 2025 Software Engineering Intern job posting — which carried a hiring target of Fall 2025 or Spring 2026 graduation explicitly — is a pipeline disclosure by definition: Scale posts intern roles referencing graduation dates rather than academic terms, which means it is mapping intern cohorts to anticipated full-time headcount by quarter. Intern hourly pay is $60.00 for software engineering, per Levels.fyi, with research tracks in post-training and frontier reasoning carrying the same graduation targeting. Scale does not publish a conversion percentage, but the hiring architecture — role-specific, graduation-anchored, with posted benefits matching those offered to full-time employees — is among the more explicit conversion structures of any lab in this cohort.
Cohere runs a continuous internship and co-op cycle out of Toronto, with Spring/Summer 2026 research positions active as of this writing. The program targets PhD candidates in machine learning, NLP, and related disciplines, with exceptional non-PhD candidates considered case by case. Cohere's hiring velocity slowed in 2025 relative to its 2023 peak, a dynamic covered in our prior reporting on the company. The intern-to-full-time conversion at Cohere reflects that constraint: the research internship is a genuine collaboration, but class-of-2026 candidates should check for active full-time postings in parallel rather than assuming conversion.
Sector Benchmarks: What NACE's 2025 Data Shows
NACE's 2025 Internship and Co-op Report, published March 2025 and covering the 2023-24 intern cohort, sets the cross-industry benchmark: employers extended full-time offers to 62 percent of their 2024 intern class, down from 67 percent for the 2023 cohort — the lowest offer rate in five years. Among employers running in-person internships specifically, the offer rate climbed to 72 percent. Among hybrid programs, it dropped to 56 percent.
Every AI lab internship program listed above is in-person or required-in-person. That modality choice is not incidental. Labs competing for the same research and engineering talent have settled on office-based formats because the conversion economics favor them: interns embedded in production teams, physically present in daily syncs, are evaluated on contribution rather than output documents. The labs that are converting most aggressively — OpenAI, Microsoft, Scale AI — are in-person in the markets where their teams actually work: San Francisco, Bellevue, and San Francisco respectively.
The 63.1 percent cross-industry conversion rate reported for 2024-25 in NACE's most recent data, published April 2026, is the highest in five years. AI labs are running above that benchmark if they are running structured programs at all. The question is not whether labs convert interns. It is whether a specific program was designed to convert this cohort.
What Class of 2026 Should Know Before Accepting
Three decisions matter more than the conversion rate percentage itself.
Program architecture over brand name. A Google DeepMind Student Researcher placement carries institutional prestige that will help on the academic job market and in future PhD applications. It does not carry a conversion mechanism. An OpenAI Emerging Talent internship carries less prestige per name but sits directly adjacent to full-time hiring. If your goal is a full-time offer by October 2026, program architecture matters more than the lab's ranking in a benchmark list.
Ask for the offer timeline explicitly. The strongest leverage a summer intern has runs from week six to week ten of a twelve-week program. That is when performance reviews crystallize and when headcount decisions for fall and winter hiring are being made. At Microsoft, return offers are typically extended in the final two weeks of the internship. At OpenAI, the 13-week systems research track includes explicit language about potential 26-week extension, which is a conversion signal: extensions are how labs hold candidates they want to eventually hire. Ask, in week six, whether your team has headcount for a return offer. You will get an honest answer more often than you expect.
The expiration risk is real. Return offers at AI labs typically carry 2-to-4 week acceptance windows in the September-October period, which creates a compression problem for candidates also holding other offers. A candidate holding a Scale AI return offer and waiting on a Google full-time recruiting process will likely face an expiration before the second process concludes. Plan for parallel timelines rather than sequential ones.
What to Watch in the Next 90 Days
Three indicators will clarify whether this intern cohort converts at the pace the hiring architecture suggests.
OpenAI's headcount trajectory through Q3 2026 is the primary signal. The company's expansion into Systems Research with extended-term internships implies it needs full-time engineers in that function. If the 26-week extensions materialize for the Summer 2026 cohort, that is a conversion before the formal offer process even opens.
Anthropic's Fellows Program May and July 2026 cohort size relative to its headcount growth rate determines how many fellowship-to-full-time conversions are actually funded. The 40-percent conversion floor from cohort one held when Anthropic was scaling from 300 to 500 employees. At current headcount of approximately 1,000, the absolute number of fellows-to-FT slots grows, but whether the percentage holds depends on whether hiring velocity continues or moderates.
Scale AI's fall 2026 full-time posting volume will be the clearest lagging indicator of summer intern conversion. Scale anchors its intern programs to specific graduation cohorts. If the fall 2026 job board reflects openings calibrated to Spring 2026 graduates, the summer pipeline is working. If postings thin out, the conversion rate for this cohort dropped.
The intern class of 2026 is entering the strongest single-year AI hiring cycle in the five-year history of this competition. The labs that built structured pipelines — and told candidates what they were — are the ones doing the actual hiring. The ones that built credential programs are building the next generation of competitors.