8news

Tech • AI • Robotics

VIDEO
ENFR
TodayShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI GPT-6 Astra, DeepSeek V4.1 Flash, Mistral €3B

AIWednesday, September 9, 2026· 18 videos

Briefing

Audio player
0:00 / 0:00

OpenAI unveils GPT-6 Astra

OpenAI formally introduced GPT-6, codenamed Astra, after a September 3 launch that it framed as a step-change in workplace automation and multimodal reasoning. The model is pitched less as a chatbot than as a general work assistant able to handle voice, screen, camera and broad task delegation in natural language. A defining feature is computer use, with agents able to operate software and execute multi-step workflows on a user’s machine. Early reactions center on stronger handling of ambiguous research, project structuring and coworker-like autonomy.

OpenAI warns on self-improving systems

OpenAI chief scientist Jakob Pachocki said frontier models remain partly understood "alien minds" whose capabilities emerge from scaling rather than elegant top-down design. He argued internal evidence suggests progress may continue into recursive self-improvement, with AI increasingly helping build better AI over the next few years. The warning sharpens the view that no major lab is fully prepared for that transition and that voluntary safety measures will be insufficient. It also suggests the long-running argument over whether AGI has arrived may be less urgent than the governance gap around fast-improving agents.

Navier-Stokes claim fuels capability debate

OpenAI said one of its systems found a counterexample implying the Navier-Stokes equations can develop a singularity under certain conditions, touching one of the seven Millennium Prize Problems. The claim matters symbolically because it would indicate AI can contribute genuinely new scientific knowledge, not just summarize or assist. At the same time, debate remains over practical impact, since engineers already solve many fluid problems numerically without a full proof. The episode extends a broader race in elite mathematics after OpenAI and Google DeepMind reportedly each hit 35/42 on the International Math Olympiad in 2025.

DeepSeek tests V4.1 Flash

DeepSeek opened a short public test of DeepSeek V4.1 Flash, a two-day build available through its API until September 10. The company describes it as a new architecture with native multimodal support, while withholding official benchmarks and parameter counts. Early testers reported generation speeds around 350 to 400 tokens per second, with one run peaking at 427 tok/s, putting throughput at the center of the release. DeepSeek also signaled lower V4 Flash pricing starting September 10, reinforcing a speed-and-cost play.

Grok 4.7 rumors gather pace

Signs tied to Grokbot suggest xAI may be preparing a near-term release of Grok 4.7, following a pattern that reportedly preceded Grok 4. Rumors place the model at roughly 2.1 trillion parameters, which would put it among the largest frontier systems discussed publicly. Purported side-by-side outputs linked to a hidden checkpoint suggest improvements over Grok 4.6, particularly in ambience, gradients and visual creativity. Even without confirmation, the leaks highlight how tightly compressed the top-tier race has become among xAI, OpenAI, Anthropic and Google.

Mistral lands record €3B round

Mistral secured a record €3 billion funding round, valuing the French company at about €21 billion and marking one of Europe’s biggest AI financing events. Rather than chase the consumer assistant war directly, Mistral is selling control: on-premise deployment, modifiable open-weight models, dedicated compute and auditable outputs. That strategy is aimed squarely at enterprises and public-sector buyers worried about compliance, sovereignty and vendor dependence. The raise signals that European investors still see room for a large regional AI champion despite US and Chinese dominance.

ChatGPT Images 2.5 expands creation

OpenAI began rolling out ChatGPT Images 2.5, broadening image generation across the web app, desktop, chat workflows and the API. New features include sketch-to-image, faster rendering, stronger in-image editing, better reference-photo consistency and structured templates such as logos. Two API variants were highlighted: GPT Image 2.5 Flare for speed and GPT Image 2.5 Sunburst for higher-detail work. Early comparisons suggest improved visual quality and multi-step editing consistency, though some tests found generation slower than Gemini Nano Banana Pro.

Insilico drug shows younger blood

Insilico Medicine reported an unexpected signal from Rentosertib, its AI-designed treatment for idiopathic pulmonary fibrosis. In a 42-patient trial, routine blood samples analyzed by six separate aging clocks suggested treated patients appeared 3 to 6 biological years younger than placebo at peak effect. Researchers stressed this does not mean patients literally became younger, but that blood protein patterns shifted toward profiles associated with younger biology. The finding adds to evidence that AI’s impact is spreading beyond software into drug discovery and biomarker interpretation.

Videos covered

Previous briefings · AI