OpenAI
OpenAI5d ago
Tech

GPT-6 Astra with Ben Davis

2 min video2 key momentsWatch original
TL;DR

Astra solved three DEF CON puzzles that humans couldn't, including one never solved before — Ben Davis says its ability to coordinate parallel agents and stay on track is what sets it apart.

Key Insights

1

First model to crack itAstra solved a Rubik's cube arrangement puzzle three times in a row that humans spent days failing — the first model ever to crack it unsupervised.

2

The model spawns up to 10 parallel sub-agents to test theories while the main agent orchestrates them — it stays on track better than prior versions when paths go wrong.

Want this for every new video OpenAI posts? Brevyd summarizes each upload automatically, the morning it drops.

Deep Dive

The DEF CON test

Ben brought Astra to the DEF CON puzzle challenge, where he and his friends had spent days struggling with some of the hardest problems on offer. He fed the model three of the toughest ones. Astra solved a Rubik's cube arrangement puzzle (three by four pattern hiding a message) three times consecutively — something humans hadn't managed. It also cracked a beaded dress puzzle that required parsing a message from poorly-shot images across multiple photos. Both times it needed the official hint from puzzle creators, same as humans do, but then executed flawlessly. The kicker: Astra was the first model anywhere to solve one specific puzzle in the entire competition.

Why Astra stays on track

The architecture is what makes the difference. Astra proposes a theory, then deploys up to 10 parallel sub-agents to test different approaches at once while the main agent coordinates them. With no verification between starting point and solution, it's easy for models to hallucinate and wander into dead ends. But Ben noticed Astra maintains focus — it doesn't lose itself on bad assumptions. That consistency is why it keeps solving problems that would normally trip up earlier models. He calls out multi-agent swarms as the real unlock for tackling bigger, harder problems than were previously possible.

Takeaways

  • Test Astra on your unsolved hard problems — it's outperforming humans on DEF CON-level puzzles for the first time.
  • If you're building with agents, the swarm architecture with parallel instances coordinating on a single task is worth prototyping now.

Key moments

0:11Astra solves unsolved puzzle

It was able to solve one that we did not get, actually three that we did not get, as well as one puzzle that no one else in the world has solved. So it was the first one to get it.

2:10Parallel agent swarms unlock bigger problems

I really think that the sub-agent workflows, the swarm workflows, all of these like putting many instances of this agent together to work on much bigger, more complex problems that you probably feel like are currently possible. It is now possible and it's worth trying.

You just read one. Brevyd does this for every upload.

Follow OpenAI and every new video comes back as a summary like this, in your morning briefing. No watching required.