A new model called ๐๐๐๐ฟ๐ฎ ๐ฒ has reportedly been released, and this time the big story isnโt just about chat or reasoning.
Itโs about computer use.
So what does that actually mean?
Previous AI models could tell you how to do something. Astra is designed to actually use your computer and do the task for you.
On the ๐ข๐ฆ๐ช๐ผ๐ฟ๐น๐ฑ ๐ฏ๐ฒ๐ป๐ฐ๐ต๐บ๐ฎ๐ฟ๐ธ, Astra reportedly scored ๐ณ๐ฎ.๐ฒ%, compared with ๐ฒ๐ฑ.๐ณ% for ๐๐ฃ๐ง-๐ฑ.๐ฒ ๐ฆ๐ผ๐น.
But the more interesting part is the speed.
The same tasks that took Sol around 75 minutes took Astra about ๐ฐ๐ฌ ๐บ๐ถ๐ป๐๐๐ฒ๐ on average.
And cybersecurity is where things get even more serious.
According to OpenAIโs Preparedness Framework, Astra is the first model to reach the โCriticalโ level for cybersecurity capabilities.
During testing, it reportedly discovered two previously unknown security vulnerabilities and scored 100% on ExploitBench.
That context window is huge. You could potentially give it an entire series of books and still have room to work.
Thereโs one important catch, though.
You canโt just start using Astra today.
Access is initially limited to OpenAIโs Trusted Access and Daybreak programs, with wider access expected later across ChatGPT Plus, Pro, Business, Enterprise, API, AWS, and Azure.
And for advanced cybersecurity tasks, the model will reportedly refuse certain requests for general users.
One more interesting point:
ARC Prizeโs team has pointed out that a 100% ARC-AGI-3 score isnโt completely new. NVIDIAโs AVO architecture, running on Claude Opus 5, has reportedly achieved it before.
So maybe the bigger story isnโt just the model itself.
Itโs the combination of the model + agent system + tools.
And honestly, thatโs the part I find most interesting.
AI is slowly moving from โHereโs how you can do itโ to โIโll do it for you.โ
That shift could change a lot about how we work with computers. ๐
Of course, AI now can do more than basic understanding of rough sketches or shapes, it is far beyond what the public frontier AI models we are able to access can even do, we are far too limited from the use if its full power, no complaints from me though, we should strive to support AI safety-measures to avoid catastrophe.
ChatGPT just added a new sketch feature to its new Images 2.5 model, so I tried it with extremely simple sketches and short prompts to see how well it could understand my intention.
The city learned to pray to the Mantis Queen. // a.exot
A short AI visual experiment exploring surreal worldbuilding, character design and cinematic motion.
Created with generative AI, focusing on atmosphere, composition, movement and visual storytelling.
Heron AI is an agent that reviews architectural models and flags building code problems. The hardest part of the brief was making that legible in the first few seconds.
We built the hero as a small interaction instead of a statement. The visitor drags across a black and white building sketch and violations surface where they drag, annotated with real IBC clauses. It puts the person in the agent's position for a moment.
Everything after that section can then talk in normal product language, because the core idea has already landed without needing to be described.
Really smart way to make a technical AI capability tangibleโthe drag interaction lets the user discover the value instead of reading a claim. The architectural visual language supports that clarity beautifully.