🤖 AI HOT Daily · Sep 9, 2026

OpenAI launches ChatGPT Images 2.5 with up to 50% lower generation latency, rolls Astra out to all Plus/Pro/Business/Enterprise users in Codex and ChatGPT Work, and says an internal AI system has solved the Navier-Stokes Millennium Problem — smooth flows can form singularities in finite time, backed by a proof manuscript and Lean formal verification; an NYU mathematician accuses OpenAI of dirty play on the problem (a leak plus brute-force pursuit), Bubeck denies it, and Altman counters that the other side only had an Euler result while threatening plagiarism claims; Noam Brown says the proof cost millions but costs will fall fast; Mistral closes a €3B Series D above €21B valuation — Europe's largest-ever equity raise; Runway ships an Adobe plugin that generates and edits right inside Premiere Pro and After Effects; Kazike publishes an Astra-in-Blender tutorial (Computer Use rebuilt the Temple of Heaven in ~4h at half a $200 Pro quota); Dwarkesh Patel's experiments show pretraining progress is mostly data (12.0x vs 3.7x for modeling); Anthropic shares three ways to cut Claude costs without losing performance; and Tom Tunguz dissects OpenAI's 3x research productivity as parallel agents.

1. 1. OpenAI Unveils ChatGPT Images 2.5

   Generation latency drops up to 50% vs Images 2.0, with gains in detail, edit precision, reference-photo fidelity, and multi-round consistency — its most advanced image model yet.

2. 2. Runway Plugins Bring Generation Into Premiere & AE

   New panels embed directly in Premiere Pro and After Effects — generate images/video in the timeline, repaint clips, and drop results in; Edit Studio re-renders with Aleph 2 at the original duration, free on macOS/Windows.

3. 3. Astra Now Wide Open to All Paid Tiers

   OpenAI rolls Astra out to Plus, Pro, Business, and Enterprise across Codex and ChatGPT Work, with live demos watchable at gpt-tv.

4. 4. NYU Mathematician: OpenAI 'Fought Dirty' on Millennium Problem

   Buckmaster and Anthropic's Alpöge claim their research leaked to OpenAI, which then chased the same unique route with heavy compute; OpenAI's Bubeck denies it.

5. 5. Mistral Raises €3B Series D, Valuation Exceeds €21B

   With a post-money valuation above €21B, Mistral calls it the largest equity raise ever by a European tech firm — just three years after founding.

6. 6. OpenAI: Internal AI System Cracks Navier-Stokes

   OpenAI says an internal AI system proved smooth fluid flows can develop singularities in finite time, publishing the proof with Lean formal verification.

7. 7. Pretraining Progress Is Mostly Data: Dwarkesh Patel

   Across 2019-2025 recipes at up to 1e19 FLOPs, data improvements delivered 12.0x compute-efficiency vs 3.7x from modeling — data matters ~3.24x more.

8. 8. Mathematicians Publicize Blowup Results and OpenAI Thread

   Buckeraster and Alpöge release three finite-time blowup results — including incompressible porous-media flows with smooth forcing — and the full OpenAI correspondence.

9. 9. Noam Brown: Proof Money Burn Will Drop Fast

   He admits the proof cost millions of dollars but argues such costs will plunge quickly as agentic pipelines scale and compute cheapens.

10. 10. A Hand-Held Astra-Steers-Blender Tutorial

    Kazike tests three modes: Computer Use rebuilt the Temple of Heaven in ~4h but burned ~half a $200 Pro quota; MCP is faster for vehicle modeling/animation yet complex tasks hit timeouts.

11. 11. Altman Responds to the Proof-Launch Spat

    Altman says the other side only had an Euler result, coordination failed, and the rival team threatened plagiarism accusations.

12. 12. Three Ways to Cut Claude Costs, Keep Performance

    Raise prompt-cache hit rates, purge anti-patterns left over from older models, and calibrate effort — cheaper without giving up quality.

13. 13. Tom Tunguz Decodes OpenAI's 3x Research Productivity

    Internal data shows each researcher's 8-hour shift maps to 3.14 agent workdays, usually with 4 agents in parallel — is the '3x' just machines that never sleep?