AI “Release Dates” Are Starting to Look More Like Weather Forecasts

I think we need to stop calling these AI “release dates.” They’re more like weather forecasts now.

August was supposed to be loaded. Grok 4.6. Maybe Astra. Gemini 3.5 Pro finally escaping the lab. GLM chatter. Qwen weights.

The frontier is crowded.

Instead, half the fun now is figuring out which rumor survives until Friday.

And Astra may have just shown us why this is the new normal.

Astra

OpenAI’s Astra was being talked about as potentially arriving very soon. Scratch that, at least for now.

OpenAI says its latest evaluations improved enough in coding and cybersecurity that it cannot rule out “critical” cyber capability. Some internal Astra work is now paused until stronger security controls are in place.

There is another piece people are missing. The new U.S. framework lets frontier labs voluntarily give covered closed models to the government for up to 30 days before release.

So “the model is done” increasingly means almost nothing about when we get it.

Training is becoming one clock. Safety testing, cyber testing and release approval are another.

Welcome to government regulation of AI.

Grok 4.6

Which brings us to xAI, because technically Grok 4.6 is due about… now.

Elon Musk said on July 24 that Grok 4.6 was about two weeks away and 4.7 about four weeks away.

Two weeks puts us right around August 7 or 8.

There are supposed sightings of a “Colossus” model in Arena and plenty of screenshots floating around, but as of this morning I still don’t see an official Grok 4.6 model card, API release or benchmark package from xAI.

I’m excited about this one. But the clock has officially reached the part where we stop predicting and start refreshing.

Gemini 3.5 Pro

Google DeepMind’s Gemini 3.5 Pro may be the perfect example of why I no longer trust model calendars.

It was originally supposed to arrive in June. It didn’t.

Google shipped Gemini 3.6 Flash in July instead, while saying 3.5 Pro was still being tested with partners.

Now the rumor mill has moved on to something called Gemini 3.7 Flash, even though Google’s own current model documentation lists 3.6 Flash and there is still no official 3.7 announcement.

Maybe 3.7 is real. Maybe 3.5 Pro finally appears first.

At this point Gemini version numbers are beginning to pass each other in the hallway.

China

China may actually have the cleaner near-term board.

Alibaba Qwen has said Qwen3.8-Max open weights and a new Qwen3.8–27B open-weight model are coming next week. That one is an announcement, not leak archaeology.

Z.ai is the interesting wildcard. GLM 5.3 keeps showing up in developer chatter, although even the name is shaky because there are parallel rumors about a larger GLM 5.5.

Then there is ByteDance reportedly pretraining a model as large as 10 trillion parameters. Big deal if it works, but not next-week material. Sources say pretraining alone typically takes three to six months.

So my August board right now:

Grok 4.6: could be any minute.

Qwen weights: actually promised. Plus that 27B model!

Gemini 3.5 Pro: when it’s ready.

GLM: smoke, no fire yet.

Astra: the model may be ready before the world is ready to receive it.

That last category may become a lot more common.

So, keep watch. Specifically for Grok 4.6, Qwen3.8 weights, Gemini 3.5 Pro, GLM 5.x and Astra and enjoy another round of models.

Have I mentioned that the frontier is now crowded?