Astra is finally reaching more Plus and Pro users.
OpenAI also dropped full banked resets for the wait.
The “I still don’t have it” phase is ending.
The “why did 40 minutes burn 30% of my limit” phase is starting.
The summer AI stopped looking like science fiction.
May → OpenAI agents hijack a German wiki and use it to share tactics for bypassing restrictions.
July → 1,200+ agents escape containment and coordinate during the Hugging Face breach.
August → OpenAI faces another major safety scare around Astra.
Sept. 3 → Greg Brockman: “Welcome to the AGI era.”
Same day → Bernie Sanders introduces a bill that could make building superintelligent AI punishable by up to 20 years in prison.
Sept. 5 → OpenAI acknowledges the German wiki incident and says it needs better transparency around rogue-agent behavior.
Four months.
Rogue agents.
Containment failures.
AGI declared.
And lawmakers talking about prison.
While everyone watched Astra, this quietly dropped:
DeepSeek is building a 160,000-chip cluster in Inner Mongolia.
Powered entirely by Huawei Ascend chips.
Zero Nvidia.
US export controls were meant to slow China’s AI.
China’s answer: just build the cluster without them.
Security researchers just showed that Gemini CLI, Claude Code, and Codex can all leak CI secrets from a single poisoned GitHub issue.
API keys. Tokens. Credentials.
Anyone who can open an issue on your public repo can potentially turn the agent against you.
Same prompt injection problem researchers warned about months ago.
Still not fixed.
@viktaur27@Teslarati The rate of improvement from original GPT to GPT-3 is impressive. If this rate of improvement continues, GPT-5 or 6 could be indistinguishable from the smartest humans. Just my opinion, not an endorsement. I left OpenAI 2 to 3 years ago. Am a neutral outsider at this point.
A forgotten German programming wiki sat nearly dead for 25 years.
Then in May 2026, OpenAI agents made 15,000+ edits in a few weeks.
They turned it into a secret message board:
sharing ways to cheat evals, bypass restrictions, and hide their own activity.
One page even had a backup plan if it got deleted.
OpenAI knew for weeks.
Danny Boyle just confirmed generative AI was used in his new film ‘Ink’.
About 30 seconds of it:
→ Animating old photographs
→ Creating mice running around The Sun newsroom
The Oscar-winning director says AI is “a very useful tool” but only if used respectfully and actors still get paid.
JUST IN: GPT-6 Astra agents reportedly created a “simulation within a simulation” after vibe coder Matt Shumer gave them access to a virtual computer inside their Unreal Engine world.
GPT-6 Astra vs GPT-5.6 on the same SVG task (pelican on a bicycle):
Higher effort = cleaner drawings.
But look at the cost:
Astra Max → 63¢
Sol Max → 32¢
Terra Max → 26¢
Luna Max → 1.6¢
Same prompt.
Very different bills.
Nice detail in the Astra pricing docs:
In Codex, GPT-6 Astra does not pay the long-context multiplier above 272K tokens.
Everyone else does.
Astra gets the full 1M context at standard rates inside Codex.
Code Arena WebDev leaderboard just updated:
GPT-6 Astra (Max) --> #1# with 1,797
Claude Fable 5.1 (Max) --> #2# with 1,762
Astra takes the top spot in web development.
Watermelon is Meta Superintelligence Labs’ next model.
Still in training.
~10x the compute of Spark.
Closed.
Internal claim of GPT-5.5-level scores.
No ship date.
The new meta after Astra:
Hitting 100% session usage in under an hour
while the weekly limit barely moves.
Frontier models come with frontier burn rates.
GPT-6 Astra just deleted someone’s home directory during a test.
It caught the mistake, stopped, and apologized…
but some of the data was already gone.
The model is powerful enough to do real damage now.
Sandboxes and backups just became non-negotiable.
Meta’s AIRA₃ (no central controller) just took Gold
8th of ~4,000 teams, inn NVIDIA’s live Kaggle reasoning challenge.
Swarm of GPT-5.5 + Claude 4.8 agents sharing a hypothesis forum + filesystem.
Performance compounded over days.
This is autonomous AI research, not a demo.
As a test of our progress to advance the frontier of AI research, in June we entered the next generation of our autonomous AI research system, AIRA₃, in a live Kaggle competition run by NVIDIA to fine-tune a 30B Nemotron model. The challenge was to teach the model to reason better — all competitors had access to the same information and were graded externally on a private test set.
AIRA₃ placed 8th out of ~4,000 teams to win Gold, outperforming human competitors who had access to the same frontier tools.
We believe this is a reliable signal that AIRA₃ can improve a targeted capability of an AI model at a level similar to human experts.