The most important AI benchmark ever built just launched. It's called ARC-AGI-3. It dropped today at Y Combinator in San Francisco with a fireside chat between François Chollet and Sam Altman. The results? The best AI agent scored 12.58%. Frontier LLMs scored under 1%. Humans score 100%. That gap tells you everything about where we actually are on the road to AGI.

ARC-AGI has been the gold standard for measuring real AI intelligence since 2019. Versions 1 and 2 tested static pattern recognition. Show the AI a grid pattern, ask it to complete a new one. Frontier models eventually cracked version 1 at over 90%. Version 2 was harder but still the same format.

Version 3 changes everything. It's no longer about looking at patterns. Now AI agents are dropped into interactive video-game-like environments. There are no instructions. No descriptions. No stated rules or goals. The agent sees a visual state, takes an action, sees the result, and has to figure out what it's supposed to do and how to do it. On the fly. With zero prior exposure.

The benchmark includes over 1,000 levels across 150+ handcrafted environments. Each game has 8 to 10 levels where new mechanics are introduced as you progress. The agent needs to explore, learn, plan, and adapt in real time.

For humans this is surprisingly easy. Most people figure out the games within minutes. For the best AI systems on the planet it's a different story. 12.58% from the top agent. Under 1% from frontier LLMs like GPT and Claude. That's not a small gap. That's a canyon.

The scoring doesn't just measure whether an agent can complete a game. It measures how efficiently it does it compared to a human. This is the first time anyone has formally measured human versus AI learning efficiency in a benchmark.

Because this answers the question everyone keeps asking. Are we close to AGI?

Based on this data the honest answer is no. Not yet.

AI can write code. Generate images. Control your computer. Manage your calendar. Build entire applications from a prompt. But when you drop it into an unfamiliar environment with no instructions and ask it to figure things out on its own, it falls apart.

That's what general intelligence actually is. Not memorizing patterns from training data. Not following instructions really well. It's encountering something completely new and adapting in real time. A child can do this. The most advanced AI systems in the world currently cannot.

This doesn't mean AI isn't useful. It's incredibly useful. The tools shipping right now from Anthropic, OpenAI, Google, and others are genuinely transforming how people work. But there's a difference between a powerful tool and a generally intelligent system. ARC-AGI-3 makes that difference impossible to ignore.

OpenAI killed Sora. The AI video generation platform that broke the internet when it launched is officially gone. OpenAI's goodbye was brief: "We're saying goodbye to the Sora app." No detailed explanation. No pivot announcement. Just done. Meanwhile Sam Altman announced the OpenAI Foundation with $1 billion in initial funding aimed at tackling AI threats and accelerating scientific discovery. One product dies, a billion dollar foundation is born. OpenAI is clearly reshuffling what matters to them heading into the second half of 2026.

ARC-AGI-3 launched today. AI agents dropped into video-game-like environments with no rules and no instructions. Best AI scored 12%. Humans scored 100%. Over $2 million in prizes to anyone who can close the gap. OpenAI killed Sora and launched a $1 billion foundation on the same day. AI is moving faster than ever. But real intelligence? The hardest test ever built says we're not there yet.

Stay building. 🤖