OpenAI shipped GPT-6 Astra on Thursday and since then my feed is about 70% Astra. Screenshots of it driving a browser for forty minutes without getting lost, Greg Brockman saying it is “not unreasonable” to call this the AGI era, some guy rebuilding a Blender scene from one photo. Underneath that, the usual fight about whether the numbers mean anything. On the aggregate intelligence index it barely moved past GPT-5.6. On agent and computer-use tasks it jumped a full generation. Ten dollars per million input tokens, fifty per million output. Paid users got it a day later, in a version that refuses to discuss cybersecurity.
I have a habit with new models. Before reading a single benchmark table I ask them to draw a walrus playing a cello. Same prompt every time, one attempt, no retries. Two tusks, whiskers, a cello with two f-holes and four strings, standing on its endpin, bow in the flipper. Then I look at the picture. That’s the whole method.
Astra’s walrus, next to fourteen others
Every model on that page failed at the same spot: the bow. It hovers in the air, or fuses with the flipper, or goes straight through a tusk. Models knew what a walrus and a cello look like and still had no idea how one touches the other.
Astra’s walrus is the first one where nothing is wrong. Both tusks, whiskers, a real cello with two f-holes, four strings, a bridge and a scroll, standing on its endpin. One flipper on the neck, the other holds the bow, and the bow is on the strings. First try, full checklist.
And I looked at it and felt nothing. It’s a tidy cartoon, the kind you get from a stock icon pack. Gemini’s walrus from February has a spotlight and closed eyes and clearly feels the music. Astra’s just sits there doing everything correctly.
So, hype. On my walrus it’s a good student. Not a new era.