GPT-6 Astra just cracked a German Enigma cipher from World War II.
A message encrypted in 1942. Eighty years. Professional cryptographers. Government programs. Couldn't break it.
Someone handed Astra the ciphertext in a chat window. It decoded it.
No fine-tuning. No research team. No special setup. Just a conversation.
I keep thinking about what this means. Not as a benchmark. As a signal. The bar for "only humans can do this" just dropped again. Hard.
GPT-6 Astra just directed a 30-second film shot.
No cuts. Street to a private casino table, one continuous take. It mapped the global space, locked the camera blocking in low-poly previz, then handed the whole thing to Seedance 2.5 to render.
That used to be a location scout, a director, a DP, and a post-production team. Now it's a prompt.
Everyone is racing to build the fastest flash model.
DeepSeek, Qwen, GLM all dropped new ones this week. Speed benchmarks everywhere.
Meanwhile Jev just launched quietly. It's not an LLM. It doesn't generate answers. It makes decisions. 200x faster than any of these flash models. 400x cheaper.
Someone ran 5,000 requests and spent $2. The model speed race everyone is watching just became the wrong race to follow.
A 13-year-old uploaded a trading script to GitHub during summer vacation.
For free. No pitch. No promotion.
Went back to school.
Two months later, a trader found it and sent him $20,000.
He didn't build a startup. He didn't raise funding. He just shipped something useful and put it out.
That's the whole playbook now.
I put Opus 5 and GPT-6 Astra head to head.
Same prompt. Build a 3D kart racer. One shot each.
Opus 5 came out smoother. Astra had a lens flare issue.
When I removed it, they were almost identical.
A few months ago there was a clear winner in every comparison. Now the gap is basically gone.
The model doesn't matter as much as what you do with it.
I showed Astra the 2025 Super Bowl ad and told it to rebuild it.
One shot. No design brief. No templates.
It watched the original, wrote the code in Remotion, generated the visuals with GPT-Image-2, and matched the camera cuts.
A few months ago that's a two-week agency project.
Now it's a single prompt and a few minutes.
The cost of production just went to zero.
Astra's quality issues just got fixed.
The same model that was glitching last week now watched a Seedance clip and rebuilt it in Blender from scratch.
Camera angles. Lighting. Depth of field. Motion blur. All set by the model.
It's rough around the edges. But a few months ago this required a studio.
The barrier to entry just dropped again.
A 39-year-old programmer built a stock trading system with Claude Code.
No trading background. 2 months.
Week 1: 4x return on his starting capital.
The barrier to building isn't technical anymore.
The only thing standing between "I have an idea" and "I built it" is whether you actually start.
This is what that looks like.
GPT-6 Astra just recreated Craig Federighi in Blender.
One prompt. No 3D experience required.
I used to pay a 3D artist $200/hr for this kind of work.
The skill isn't dead. The barrier to entry just hit zero.
Anyone can now produce senior-level 3D work on demand. The question is what you do with that.
SWE-2 just dropped.
98% of GPT-6 Astra's performance. At 70% lower cost.
3 months ago I was paying $50 to run this task. Today it's $0.70.
Every quarter the frontier model becomes the commodity model.
This is exactly what I wrote about in my last article: the model gap didn't just close — it's now irrelevant.
Your model isn't your advantage. Your workflow is.
I asked Claude Fable 5 to simulate real physics from scratch.
Double pendulum. Galton board. Water in a rotating drum.
No physics engine. No library. Just a prompt and HTML5.
The water simulation stopped me cold. The fluid volume stayed cohesive. The particles didn't scatter. It looked real.
Cost: $3.35. Time: 14 minutes.
This is what I mean when I say we're not talking about chatbots anymore.
GPT-6 Astra just took over your computer.
Anything you used to do manually — browsing, coding, research, email — Astra does it on command.
Not "it helps you." It does it FOR you.
Computer use was a gimmick 6 months ago. Now it's the product.
The scary part: most people still don't know this shipped.
Humanoid robots just ran the 100m at the World Humanoid Robot Games in Beijing.
Some fell. Most kept going.
Four years ago this was CGI.
Now it's a stadium. 20+ models competing in front of a crowd.
The gap between lab prototype and athletic competition closed faster than any benchmark predicted.
What closes next: endurance. Then agility. Then you stop being able to justify why the gap between humans and machines still matters.
A humanoid robot just ran the 400m faster than any human in history.
38 seconds. At the World Humanoid Robot Games in Beijing.
The human world record: 43 seconds. Set in 2016. Never broken.
But here's the part that actually matters:
Nobody programmed how it moves.
The arm swing that makes it faster wasn't designed. The robot discovered it during training — a physics trick, rotating the arms to stabilize the hips at max speed. The system invented it without instruction.
Last year, the fastest robot finished the same race in 1 minute 28 seconds.
That's not a 10% improvement. That's a different category.
We're not watching engineered machines anymore. We're watching machines that teach themselves to move the way humans learned — through millions of failed attempts until something clicks.
The crowd in Beijing was watching like it was the Olympics.
It kind of is.