This New AI System Cuts Research Costs by 90% — And Does It Better

Researchers built an AI teammate that doesn't just run experiments — it learns from them. Meet Praxist, a system that lets AI research agents track which ideas worked, which failed, and why, instead of starting from scratch each time. Today's autonomous agents often repeat the same mistakes across long projects because they treat every attempt as isolated. Praxist fixes this by creating an "evidence graph" — a connected map of findings and failures that future agents inherit, so lessons actually stick.

The results are striking. On MLE-bench, a standardized test of 75 research tasks, Praxist earned 60 medals (49 gold) versus 55 medals (34 gold) for a Claude Code baseline using the same underlying AI model. More impressive is the cost: Praxist spent about $3,054 on computing, while the baseline burned $38,370. That's a 12x savings — the difference between a research budget that's sustainable and one that isn't.

Praxist also worked on real-world engineering challenges, improving results in algorithmic stock trading, robotic navigation (SLAM), nuclear fusion reactor control (tokamaks), and rocket landings. In each case, it beat the task's existing baseline while leaving a clear audit trail — so you can see exactly which design choice made the difference.

For everyday people, this matters because scientific and engineering breakthroughs often stall on cost and repetition. If AI can do R&D at a fraction of today's price while keeping better records, we could see faster development of better batteries, safer self-driving cars, cleaner fusion energy, and more. The catch: this is early research, not yet available as a product, but it hints at a future where top-tier research capability is no longer a luxury.