By skyforbes The basic story is that a nine-researcher startup just developed a way of virtually instantaneously (within a few hours) layering a meta-system architecture onto virtually any AI that can handle Python, often doubling reasoning performance to the extent that a model like GPT 5.1 or Gemini 3 can move from scoring about 30% on ARC-AGI-2 to scoring over 60%, a score that surpasses even human performance on this benchmark! Additionally, instead of this fitting taking weeks or months, it can be fully implemented within hours of a model's launch.
It can also achieve this performance acceleration at six times less cost than it would take Gemini 3 or other top models. But that's just the beginning. To frame this in terms a layman can understand, it immediately transforms an AI that scores 13O on the Norway Mensa IQ test offline to one that scores 170 or higher.
Poetiq announced its benchmark results based on public ARC-AGI-2 data, and the official verification will probably be completed by December 5th. Given the stature of the researchers on the team, we can be confident that their results will pass the private data verification as well.
This breakthrough will accelerate AI across every domain, but especially within the fundamental domain of AI reasoning, from where it can further accelerate every other aspect of AI development.
One way to understand how this will come about is to realize that boosting top AI IQ from 130 to 170 is just the beginning. Whereas model IQ increases have been limited to 2.5 points per month over the last 18 months, it's reasonable to expect that moving into 2026 this rate will increase to perhaps 4 or 5 points per month. So imagine unleashing millions of 200 IQ level AIs on our hardest problems across every scientific, medical and enterprise domain before the end of 2026!!!
But perhaps the most amazing part of this advancement is that the scaffold is recursively self-improving. It will continue to improve itself with each iteration so that the numbers cited above will only get stronger and stronger, perhaps exponentially, at a faster and faster rate.
Something else to note about Poetiq is that it works by bringing together top models like Gemini 3 and Claude 4.5 to achieve these world-changing results. In fact, there's no theoretical limit to how many models Poetiq can pull together to work as a team, increasing the power and efficiency of the mix far beyond what each of the models could achieve on their own.
This is an inflection point in AI that we can hardly begin to understand and appreciate. Recursive self-improvement means that ASI may be just months away. Imagine AIs that are 10 or 20 times more intelligent than the most intelligent person who has ever lived. Imagine the problems these AIs will solve. Right now we are way too amazed to really understand what this inflection point really means, but as December unfolds it will become crystal clear as our top AI researchers step up to the plate to explain to the world what has just happened.