It’s time to stop worrying about our AI overlords, at least for the time being. A new benchmark test from the ARC Prize Foundation, a non-profit tracking the development of Artificial General Intelligence (AGI), has shown that humans still have the edge over machines, with none of the current wave of frontier AI models able to match our brainpower. 

Proving our opposable thumbs are yet to be defeated, ARC’s latest test, the ARC-AGI-3, challenges AI agents with a series of “games”, rather than asking them to solve fixed puzzles, evaluating their ability to learn from experience.

The test sees an AI agent dropped into unfamiliar turn-based environments with no prompts, goals, or pre-set rules, forcing them to fend entirely for themselves to evaluate how well they interrogate their environment, acquire skills, and adapt strategy without instruction.

A 100% score means that the AI can beat every game as well as any human. So far, none have scored close to 1%.

According to ARC’s leaderboard, Google’s Gemini 3.1 Pro and Anthropic’s Opus 4.6 have both managed 0.2%, while OpenAI’s GPT-5.4 has squeezed ahead with 0.3%. xAI’s latest flagship model, Grok 4.20, is currently sitting at 0.0% – a little awkward, given Musk’s recent dismissal of his competitor’s products as “weak sauce”. 

But while frontier AI models haven’t covered themselves in glory, the good news is that 100% of human testers solved every level of the benchmark’s environments on their first try, with ARC saying that, as long as there is a gap between human and AI’s “fluid intelligence”, we haven’t achieved AGI – no matter what Jensen Huang thinks. 

While it would be easy to dismiss the ARC-AGI-3 as just one new AI benchmark in an already crowded field, its developers argue that it offers a far more rigorous measure of genuine reasoning ability.

As with other benchmarks, previous iterations of ARC’s test involved static pattern-matching and puzzle‑solving tasks, leaving them susceptible to AIs attempting brute force solutions by throwing masses of compute at the tests and simply memorising patterns through training.

Recommended reading

• AI Giants ‘Fundamentally Unprepared’ for AGI Safety
• World’s Leading AI Firms Falling Behind Global Safety Standards
• Beyond DeepSeek: 3 Critical Questions For The Future of AI

ARC-AGI-3, however, has no dataset to memorise, since most of its environments have been kept private by design, and any brute force attempts come with severe penalties, even if an AI eventually beats the level.   

“Beating ARC-AGI-3 will be achieved when an AI system matches or exceeds human-level action efficiency on all environments, upon seeing them for the first time,” said François Chollet, co-founder of the ARC Foundation and creator of the ARC-AGI benchmark.

“Any sudden score jump will mean something important has changed about AI capabilities. ARC-AGI-3 will be your early warning signal.”