AI Tutors Achieve Human-Level GRE Gains at 918x Lower Cost

💡 A randomized study of 2,383 students published in September 2026 found that AI tutoring produces GRE learning gains statistically equivalent to expert human tutors - at 918 times lower cost per point gained. This is the first large, controlled head-to-head comparison, and the results matter for anyone who learns, teaches, or pays for tutoring.

What Actually Happened in the StudentBench Study

September 2026 brought one of the most carefully designed educational AI comparisons yet. Researchers from Handshake AI Research published StudentBench, a randomized controlled trial comparing AI tutoring to expert human tutoring on GRE preparation. The finding: statistically equivalent learning gains, in a properly randomized trial with over two thousand participants.

The study assigned 2,383 adult participants to one of three conditions: a 60-minute session with one of 13 AI tutors built on frontier language models, a 60-minute session with an expert human GRE tutor, or no tutoring at all. Every participant sat a standardized test before and after. The pre-to-post gap was the measure of learning.

How Did the Researchers Set Up the Test?

The trial collected over 175,000 student-AI messages from a publicly accessible platform. Participants were randomly assigned; the pre- and post-test questions were equivalent but distinct, so memory alone could not inflate scores. A secondary study had expert human tutors complete 2,028 pairwise evaluations of AI-generated lesson plans and practice problems, giving an independent quality signal beyond test scores alone.

Thirteen AI tutors were tested, drawing on frontier models from multiple vendors. The specific model names were not publicly disclosed. One result stands out sharply: the most cost-effective AI tutor in the trial delivered equivalent gains at $0.0052 per percentage point, compared to $4.81 for a human tutor. Faster AI response times also correlated with better student engagement and learning outcomes (all p < .002).

What Does Statistically Equivalent Actually Mean?

Statistically equivalent does not mean identical. It means the gap between average AI and human learning gains fell within the range plausibly attributable to chance, given the sample size. The study used a p = .015 threshold, tighter than the conventional p = .05 standard. That is worth noting: the researchers set a higher bar and the result still held.

In 5 of 7 GRE subject domains, the best-performing AI tutor outperformed the human tutor on average. Across all seven domains combined, neither side had a statistically meaningful edge. Both findings are true simultaneously. Read the result precisely: one AI tutor, in certain domains, in one narrow context, in one study.

What This Means for You: Learning, Tutoring, and Affordability

This is where the finding has real-world weight. If you are preparing for a standardized test - or helping someone who is - a well-designed AI tutor is now a credible option for the preparation stage. The cost gap is not marginal. Moving from $4.81 to $0.0052 per point gained shifts high-quality prep from a premium service into something that can reach learners who could not previously afford expert one-on-one coaching. That is a genuine change in access.

For anyone working in education, language learning, or professional certification training, the signal is practical: AI tutoring tools are worth testing seriously in structured, goal-defined contexts. The existing science on working memory and how we retain new material is relevant here too. The best AI tutors in the secondary study generated strong lesson plans and retrieval practice problems, which aligns with what learning science says about how knowledge sticks.

For teachers and professional tutors, the picture is more nuanced. The study shows AI matches human tutors on short-term, measurable gains in a specific bounded context. It does not show AI is a superior long-term educator or that students prefer it for the many things human teaching involves beyond content delivery.

What Can This Study Not Actually Show?

The study is well-designed, but any honest reading needs these caveats.

• One session, not a semester. A 60-minute GRE prep session is not a full learning relationship. The study cannot show whether gains persist, accumulate over time, or compound with repeated AI tutoring.
• A well-defined task. GRE questions have known right answers and clear rubrics. That makes them a favorable setting for AI. Open-ended tutoring in creative subjects, interpersonal skills, or long-form writing is a different challenge.
• Adults only. Participants were adults preparing for graduate school admission. The findings say nothing about children, K-12 education, or developmental learning contexts.
• Group averages, not every student. The authors note directly: "statistical equivalence at the study level does not mean that every student or topic receives the same quality of support."
• Funding source matters. Handshake AI Research is the data arm of Handshake, a jobs platform with commercial incentives in AI-in-education. The methods appear sound and the data is open, but independent replication by researchers with no stake in the result is the right next step before treating this as settled.

What Should You Watch Next?

The StudentBench team made their evaluation platform freely available and published their code on GitHub, a good-faith move that invites independent scrutiny. Independent coverage from educators and researchers with no stake in the outcome has generally found the methodology credible.

The questions that matter most now: do AI tutoring gains persist over weeks or months? Do they generalize beyond exam prep to creative or long-form learning? As AI models improve and costs fall further, where does the remaining human tutor edge actually lie? For now, AI tutors have passed a meaningful first test in one well-defined context. The limits are real, and so are the implications.

FAQ

Source(s): StudentBench, arXiv:2609.28470 (September 2026); AI Weekly (2026)

About the author

Dao Huy (Lucas) is a professional translator working across English, Vietnamese, Chinese, and French with over 7 years of experience. He follows the frontier of AI, learning science, and language technology because these fields directly shape how communication, knowledge transfer, and education work - and staying current is part of doing translation well.

He offers English-Vietnamese, technical, and software localization translation for clients who need accurate, clear work across languages. To request a quote, visit daohuy.com.