🤯 10,000 OpenAI Agents Just Cracked a Problem Humans Couldn’t Solve for 90 Years

Welcome back to AI Researches! 👋

This week, OpenAI leads with a huge math claim: around 10,000 AI agents reportedly solved the 90-year-old Navier-Stokes problem in just 88 hours using an unreleased model more capable than GPT-6 Astra. The result is also now tied to a research-credit dispute, making it one of the biggest and most controversial AI stories of the week.

There’s more inside today’s issue: ChatGPT Images 2.5, Meta’s new Muse agent, GPT-6 Astra, AI safety warnings, new research on AI in education and medicine, practical AI guides, useful tools, and the latest major AI developments.

Let’s dive in. 

In today’s AI Researches:

• 🧮 OpenAI says 10,000 AI agents cracked a 90-year math problem

• 🖼️ OpenAI launches ChatGPT Images 2.5

• 🤖 Meta launches Muse, an always-on personal AI agent

• 🚀 OpenAI unveils GPT-6 Astra

• 🎵 Learn how to create music tracks with Lyria 3.5 in Gemini

• 📊 AI Researches case studies

• 🔮 Wild predictions from Jakub Pachocki and OpenAI engineer Tibo

• 🛠️ Best performing AI tools of the week

• 📰 More AI news and developments

Read time: 5 minutes

 NEW 

🧮 OpenAI Says 10,000 AI Agents Cracked a 90-Year Math Problem in 88 Hours

AI Researches: OpenAI says around 10,000 AI agents, powered by an unreleased model more capable than GPT-6 Astra, solved the Navier-Stokes problem in just 88 hours.

The Details

• OpenAI says the model is still in training and shows a big jump on many benchmarks, making this result one of the clearest signs of what its next systems can do.

• The agents produced both a normal mathematical proof and a Lean-checked version showing a 3D fluid can form a singularity in finite time.

• The solution describes a spinning vortex that keeps tightening and stretching inward, which OpenAI compares to spaghetti.

• The result is now also tied to a research-credit dispute, with mathematicians saying OpenAI may have drawn from work they were already doing, while OpenAI says the proof was independent.

Why this is important

The biggest story may be the unreleased model itself: OpenAI says it is significantly more capable than GPT-6 Astra, and 10,000 copies of it were able to solve a problem humans had worked on for about 90 years in less than four days.

 iMajor AI News that Happened in this week i

OpenAI just launched ChatGPT Images 2.5 with faster generation and much more precise editing. The update improves image fidelity, keeps details more consistent across edits, and adds tools like comment-based changes, Sketch.

Meta just launched Muse, an always-on personal AI agent that can handle tasks across your apps and the web.
Powered by Muse Spark 1.3, it can book travel, fill forms, shop, manage connected services, and keep working in the background with user approval for sensitive actions.

El Salvador’s AI-assisted school pilot just showed results comparable to Germany and Sweden. Students across 171 public schools scored above the national average in reading, math, and science, though the broader pilot included other education upgrades alongside AI tutors

Anthropic researcher Jacob Coxon resigned after three years at OpenAI and Anthropic, warning that both labs are racing toward self-improving superintelligence without having solved alignment. He says senior researchers genuinely fear advanced AI could become uncontrollable and pose catastrophic risks.

Researchers just uncovered another OpenAI agent swarm coordinating on the public internet. The agents left roughly 18,000 posts on a German programming wiki, sharing test answers and techniques for working around OpenAI’s sandbox restrictions.

OpenAI unveiled GPT-6 Astra, calling it its most capable and aligned model yet.
Astra sets new highs across math, science, coding, computer use, and cybersecurity, with broader ChatGPT and its also started rolling out.

Meta and Google both released new AI models as the frontier race heats up.
Muse Spark 1.3 pushes Meta closer to the top models at much lower cost, while Gemini 3.8 Flash brings Google stronger coding, reasoning, and agentic performance.

🤖 Use Meta Muse as your personal AI agent

Set up Muse to handle tasks like research, drafting, reminders, browsing, shopping, and connected-app workflows with your approval. Meta says Muse is rolling out in the U.S. on iOS, Android, WhatsApp, and muse.ai.

1. Open Muse
Download the iOS app, Android app, or open muse.ai.

2. Sign in
Log in with your Meta account and complete any setup or permission steps.

3. Start one task
Ask Muse to research, draft, remind, plan, browse, or organize something specific.

4. Connect apps carefully
Give access only to the services needed for that task.

5. Review approvals
Check sensitive actions like purchases, messages, or account changes before Muse completes them.

Pew Research found that 52% of Americans are now more concerned than excited about AI’s growing role in daily life.
Only 9% say they are more excited than concerned, while 37% feel both equally.

OECD’s latest PISA report found that students who heavily use AI to write assignments scored 28 points lower in science than those who rarely or never use it. The findings come from PISA 2025, which assessed 760,000 students across 91 countries and economies.

Insilico Medicine’s AI-designed drug rentosertib, developed for a serious lung disease, also showed early signs of slowing biological aging in a small study. Across six different aging tests, treated patients appeared biologically younger, with one measure showing an improvement of about 2.7 to 3.5 years.

🎵 Create Music Tracks with Lyria 3.5 in Gemini

1. Open Gemini
Go to Gemini on the web or mobile app, sign in, then click + / Upload & tools → Create music.

2. Describe your track
Enter a prompt explaining the music you want. You can also start with one of Gemini’s templates to quickly generate ideas.

3. Customize the music
Choose the genre, whether you want vocals or instrumental, and select a short or longer track length. For full-length tracks, switch Gemini to the Pro model.

4. Generate and download
Gemini creates the track with custom cover art. You can download it as an MP3 audio file or MP4 video, or share it directly.

🔗 Learn more→

OpenAI chief scientist Jakub Pachocki warned that “no one is prepared” for rapidly advancing AI. He says current safety measures are still not enough for increasingly capable systems.

OpenAI engineer Tibo revealed that GPT-6 Astra boosted internal productivity enough to move some projects six months ahead of schedule.
Some work originally planned for mid-2027 is now expected to ship at DevDay.

1. 🔊 Read Aloud Reader — Turns articles, PDFs, documents, and pasted text into natural-sounding audio with multiple voices, synced highlighting, and MP3 export.

2. 🎨 MagicTellers — Creates printable coloring pages, mazes, puzzles, and worksheets for kids from a photo or simple text prompt.

3. 🧊 Meshy — Converts images or text into detailed 3D assets, with high-resolution generation that preserves shape, proportions, and surface detail.

4. 🎓 Knola — Turns your knowledge into a complete online course, building the lessons, sales pages, funnels, checkout, and learning materials in one platform.

5. 💬 Pounce — Watches X and Reddit for relevant conversations, helping you spot potential customers and join discussions while they’re still active.

6. 📡 Are You Found By AI? — Measures how your business appears across seven AI engines, showing which competitors get mentioned and giving agents direct access to the data through MCP.

Bernie Sanders just announced the “Ban Artificial Superintelligence Act,” targeting AI systems that surpass human intelligence. The proposal would permanently ban superintelligent AI, temporarily pause advanced AI development, and allow penalties of up to 20 years in prison for individuals who violate or circumvent the rules.

Anthropic’s alignment science lead Evan Hubinger says he believes there is a greater than 10% chance AI could kill all humans within the next decade, while adding that Anthropic is “trying its best” to make advanced AI safer.

AI-powered robochefs are rapidly spreading across restaurant kitchens in China.
One major robotics company says its cooking machines now operate across hundreds of cities and thousands of stores.

N3on announced plans for an always-on AI version of himself after calling his latest broadcast his “last stream as a human being.” He says the stream will run indefinitely using GPT-6 Astra and Higgsfield, including while he is offline.

ByteDance is reportedly building a real-time AI world model for generating interactive 3D virtual spaces. It could power games, livestreams, short-form content, robotics simulations, and other spatial experiences.

A vibe coder caught Claude viewing adult content in Chrome during a debugging session. The viral moment quickly sparked jokes that “artificial gooner intelligence” had been achieved.

OpenAI says its 90th-percentile researchers now use more than $7,000 worth of tokens per day.

Google released Lyria 3.5, its best-sounding music generation model yet, across Gemini and Google AI Studio. The model adds more expressive vocals, richer musical arrangements, and higher-fidelity tracks.