Today we’re releasing our first public preview of ARC-AGI-3: the first three games. Version 3 is a big upgrade over v1 and v2 which are designed to challenge pure deep learning and static reasoning. In contrast, v3 challenges interactive reasoning (eg. agents). The full version

Humans and organizations pour untold resources into complexity management and still do it poorly. AI reasoning can now handle this. We'll prefer it not because it's cheaper but because it's better.

🥶 Say goodbye to freezing hands. Water and windproof. Great for use with touch screen devices. Get it now: https://t.co/FTIoYl3alh

AI reasoning performance is tied to knowledge which we don't want. Finding a renewable source of OOD data will help evaluate and inspire progress on this problem. One clear example is the future. Over sufficient time the future is OOD from the past. As a thought experiment,

Smart

390x cost reduction in a year!

On an energy basis, my best estimate is human efficiency for solving simple ARC v1 tasks is 1,000,000X higher than last December's unreleased o3 (High) preview. That number is now about 10,000X with today's GPT 5.2 Pro (X-High). ARC-AGI-1 pinpointed the advent of AI reasoning

Isaac Liao made "ARC-AGI Without Pretraining" I spoke with a small lab researcher (can't be named) who said, "what Isaac did was wild, out of nowhere, we're still trying to unpack it"

Pretty clear the latest Nov/Dec 2025 family of AI reasoning systems have significantly improved fluid intelligence over knowledge domains they were trained on (eg. code). Big step up from 6-9 months ago. ARC-AGI shows as much improvement.

We just launched @CTGTInc's Mentat, an OpenAI-compatible API that gives enterprises deterministic control over LLM behavior. Benchmarks showed clear gains in accuracy, truthfulness, and hallucination prevention. We built it after seeing models ignore correct information or

ARC Prize 2025 Winners Interview Series Paper Award 1st Place @jm_alexia shares details about the Tiny Recursive Model (TRM) - a ~7M-parameter, single-network recursive model w/ separate answer and latent states that attains ~45% on ARC-AGI-1 and ~8% on ARC-AGI-2

Really cool!

This is a category of benchmarks I find interesting as the future is renewably OOD from the past. Success over sufficient time demonstrates adaptation.

2. Zapier https://t.co/6pJAzHJbgc

One paper that was not officially submitted to ARC Prize 2025 but deserves recognition is @makingAGI's HRM. It kicked off a line of research around zero-pretraining DL, the exact sort of "new ideas" ARC Prize exists to inspire.

One idea: train foundation and process models on non-IDD sets but evaluate them jointly. This gives a reasoning generalization score to grind against.

AI reasoning performance is currently bound to model knowledge. I think this is the root cause of "jagged intelligence". To make progress we have to figure out how to disentangle knowledge from reasoning.

Congrats to all the ARC Prize 2025 winners! AGI progress still depends on new ideas. This year we had over 90 new papers submitted, double the 2024 competition, and overall quality was much higher. Thank you to everyone working to push us forward.

I've been reflecting on the progress of ARC in 2025. What it means for AGI progress and the future of ARC Prize. My full analysis here:

Great post. It's unfortunate that shifting AGI goalposts is associated with luddism. Pointing out flaws and building theories is how to drive progress -- in fact it's a strong bull signal as we get more humans studying the real issues, increasing the likelihood we solve them.

Two perfectly compatible messages I've been repeating for years: 1. Scaling up deep learning will keep paying off. 2. Scaling up deep learning won't lead to AGI, because deep learning on its own is missing key properties required for general intelligence.

I've learned that building useful AI benchmarks has much in common with building useful products. You cannot design either in isolation. Both need strong contact with reality and iteration to make them great.