GoKawiilIndependent researcher trains small transformer to 44% on ARC-AGI-1 for 67 cents
A researcher built and trained a small transformer model from scratch in 1.5 hours on a single RTX 5090 GPU, achieving 44% accuracy on the ARC-AGI-1 public evaluation set and 7% on ARC-AGI-2. The model, an upgraded and cheaper version of a previous open-source release, matches the performance of existing test-time-training approaches like TRM and HRM while using far less compute.
GoKawiil's interpretation of the reporting above, not reported fact.
ARC-AGI is a notoriously difficult benchmark because it demands sample-efficient reasoning rather than pattern matching from massive datasets, making this result significant for the broader push toward data-efficient AI. Achieving competitive scores for under a dollar in training cost dramatically lowers the barrier for researchers experimenting with generalization and reasoning, potentially accelerating open research outside large AI labs.
- A small transformer trained from scratch in 1.5 hours on one GPU hit 44% on ARC-AGI-1 for about 67 cents.
- The model matches specialized test-time-training approaches like TRM and HRM despite being simpler and cheaper.
- The work is open source and part of a series aimed at improving sample efficiency in AI research.
NVIDIA GeForce RTX 5090 Founders Edition — The whole experiment ran in 1.5 hours on a single RTX 5090 — proof that serious from-scratch transformer training no longer needs a datacenter. If you want to reproduce this ARC-AGI work or train your own small models at home, this is the card doing the heavy lifting.
See NVIDIA GeForce RTX 5090 Founders Edition on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: mvakde.github.io — Mithil Vakde, 2026-09-01
Published there as: “44% on ARC-AGI-1 in 67 cents”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.