2026-09-09 · WednesdayPublished daily at 08:00
AI Daily
- 1OpenAI releases the ChatGPT Images 2.5 image modelOpenAI: Website updates
- 2
- 3
01Model Releases/Updates
OpenAI releases the ChatGPT Images 2.5 image model
OpenAI released the ChatGPT Images 2.5 image model, reducing generation latency by up to 50% compared with Images 2.0 and improving detail, editing precision, reference photo fidelity, and consistency across multiple rounds of editing.
Original text02Product Releases/Updates
Runway released an Adobe plugin that can directly generate and edit content within Premiere Pro and After Effects.
Runway released Runway Plugins, a new panel that can be embedded directly in Premiere Pro and After Effects to generate images and videos within the timeline, repaint clips, and place the results. Edit Studio can use Aleph 2 to re-render existing clips at their original duration. The plugin is free to download (macOS and Windows), and its generation features are available on all paid plans and consume existing credits.
Original textOpenAI rolls out Astra to all Plus, Pro, Business, and Enterprise users
OpenAI announced that Astra has been fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work. Users can use it directly and watch a live Astra demonstration at openai.com/gpt-tv/.
Original text03Industry News
NYU mathematician accuses OpenAI of unfair competition in the Millennium Prize Problems contest; Bubeck denies it
NYU mathematics professor Tristan Buckmaster and Anthropic mathematician Levent Alpöge announced preliminary results for three proofs addressing the Navier-Stokes existence and smoothness problem, a Millennium Prize Problem. They said information about their research progress had leaked to OpenAI, which then used substantial computing power to catch up with the proof along their distinctive approach.
Original textMistral completes €3 billion Series D financing round, with valuation exceeding €21 billion
Mistral announced the completion of a €3 billion Series D funding round at a post-money valuation of more than €21 billion. The company said this was the largest equity financing ever raised by a European technology company, just three years after its founding.
Original text04Paper research
OpenAI announces a solution to the Millennium problem on Navier–Stokes, produced by an internal AI system
OpenAI announced that its internal AI system had produced a solution to the Navier–Stokes existence and smoothness problem, proving that an initially smooth fluid can develop a singularity in finite time. It also released the proof manuscript and a Lean formal verification.
Original textDwarkesh Patel research: Pretraining progress comes mainly from data improvements
Dwarkesh Patel published an experimental analysis training representative model recipes and data corpora from each year between 2019 and 2025 under compute budgets of up to 1e19 FLOPs. He found that data improvements produced a 12.0x increase in compute efficiency, while model improvements produced a 3.7x increase, making the contribution from data approximately 3.24 times that of models.
Original text05Tips and perspectives
Mathematician Buckmaster announces results on blowup for several equations and publicly recounts communications with OpenAI
Original textNoam Brown comments on OpenAI using agents to solve the Navier-Stokes Millennium Prize Problem: costs will fall rapidly
Noam Brown commented on OpenAI’s announcement that a combination of agents had solved the Navier–Stokes Millennium Prize problem. He acknowledged that the proof cost several million dollars but believed the cost would fall rapidly.
Original textA beginner-friendly tutorial on controlling Blender with GPT-6 Astra: three approaches and a record of pitfalls.
The author Kazike, known for Digital Life, published a beginner’s guide to using GPT-6 Astra to control Blender. It covers initial setup, configuring the official MCP and Computer Use plugins, and testing three workflows. Computer Use took about 4 hours to build Beijing’s Temple of Heaven Hall of Prayer for Good Harvests, but consumed nearly half the allowance of a $200 Pro membership. MCP was faster and could complete motorcycle modeling and assembly animation, but complex tasks could time out during a single run. Breaking the work into steps or using the CLI to run Python scripts can address this.
Original textSam Altman responds to the controversy over the release of a Navier-Stokes proof by an Anthropic researcher
Sam Altman responded to the dispute with Anthropic over the release of proofs for Euler/Navier–Stokes, saying that the other side had only an Euler result, that coordination between the two sides had failed, and that the other side had threatened them with plagiarism allegations.
Original textAnthropic explains three ways to reduce costs and improve performance with the Claude Platform.
An article from the Anthropic team says costs can be reduced without sacrificing performance by optimizing prompt-cache hit rates, removing prompt anti-patterns after upgrading to a frontier Claude model, and calibrating effort.
Original textTom Tunguz analyzes whether OpenAI’s 3x AI productivity gain is simply because machines don’t sleep
Tom Tunguz cited internal OpenAI data to analyze the source of its 3x research productivity gain: each researcher’s 8-hour shift corresponds to 3.14 agent workdays, with 4 agents typically running in parallel.
Original text