Breaking

StudentBench Study Finds AI Tutors Matched Expert GRE Tutors as Gemma 4 31B Hit 918x Lower Cost Per Point

October 1, 2026 at 13:38 EDT

Education
Foundation Models

A September 2026 StudentBench study reports that pooled AI tutors produced immediate GRE learning gains statistically equivalent to expert human tutors, with Gemma 4 31B emerging as the lowest-cost model to pass an individual human-equivalence test. The result is narrow but consequential: one hour of AI tutoring improved unaided GRE performance by 6.15 percentage points over a control condition, while human tutoring improved it by 7.04 points.

September 2026 · GRE tutoring

AI tutors matched expert GRE gains

A one-hour StudentBench experiment found pooled AI tutoring statistically equivalent to expert human tutoring on immediate GRE gains. Gemma 4 31B stood out on cost, not as the highest-scoring tutor.

2,383

adult participants

6.15 pp

AI gain over control

$0.067

Gemma 4 31B session cost

918x

lower cost per point

Immediate learning gains

Each stacked block is about one percentage point of pre-to-post score gain.

7.1 pp

Control

13.9 pp

AI pooled

15.6 pp

Human tutor

Cost curve

Gemma 4 31B reached human-equivalent gains at $0.00524 per percentage point.

Gemma

918x

Human

Human stack is compressed: each orange segment represents about 51 Gemma cost units.

Study flow

27-question pre-test→1-hour condition→27-question post-test

The comparison used AI tutors, live human tutors, and a no-tutoring control.

Why it matters

The result points to cheaper one-on-one practice, especially when low-cost open-weight models are good enough.

What remains open

The study measured immediate GRE gains in English-speaking adults, not long-term retention or classroom outcomes.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.