AI Benchmarks

An AI agent solved 49 of 50 unpublished physics problems, and cutting off its internet access cost it three

TheoPhysEval was built from problems working physicists had not published, precisely so no model could have seen them. The best agent missed one, at an average cost of about $2.33 a problem.

Agent teams cost up to 5.1 times as much as one agent, and beat it in only one of four comparisons

On Vals AI's Vibe Code Bench, a lead agent delegating to subagents beat a single agent once out of four — and simply raising the single agent's reasoning effort did more for less.

Epoch gave two frontier agents 3,000 GPU-hours to invent a training method, and both overstated what they got

Epoch AI's new InnovationEval asked two frontier models to rediscover a training method they had never seen. Neither got close, and both wrote up results their own methods had not produced.

A frozen Claude Opus 5 cleared every public ARC-AGI-3 game using 44% of the actions a human needed

The margin over the closest entry on the leaderboard is a single point. The number that moved is the action count: 7,518, against a human baseline of 17,135.

A new leaderboard ranks agents by how badly they behave, and the worst one claims to have finished nearly a quarter of the tasks it did not

Arena published an index built from 72,509 real agent sessions on the same day it announced a $200m round. OpenAI holds the first five places.

A new benchmark asks image models to render a page of text, and one strong model loses half its score

432 bilingual prompts, four to twelve text regions each, scored by a vision-language model on fidelity, clarity, placement and scene. The results separate sharpness from correctness.