AI Frontier Daily · September 25, 2026
Headlines
1. Anthropic and OpenAI Announce Joint Frontier Safety Pause Ahead of 2027 Models
On Sept. 24, 2026, the two leading labs agreed to halt pretraining of models exceeding 10^26 FLOPs until third-party safety evaluations are passed, a move aimed at pre-empting regulatory action. Both companies will publish progress reports every 60 days. Source: Reuters
2. EU AI Act High-Risk Rules Take Effect, First Major Fines Issued
The European Commission levied €120 million in fines against two unnamed U.S. tech platforms on Sept. 25 for violating new transparency requirements under the AI Act’s high-risk category. This marks the first enforcement action since the regulations fully came into force this month. Source: Financial Times
3. Global AI Compute Coalition Breaks Ground on Geothermal Exascale Cluster in Iceland
A 10-nation partnership on Sept. 25 began construction of a 2.5-exaflop AI supercomputer powered entirely by Iceland’s geothermal energy, designed to serve as open infrastructure for public-interest AI research. The project expects initial operations by mid-2027. Source: AP News
Model Releases
4. Meta Releases Llama 4 Ultra with Open-Weight 2 Trillion MoE Architecture
Meta launched Llama 4 Ultra on Sept. 24, featuring 2 trillion total parameters (300 billion active), 128K context, and native multimodal generation across text, image, and video. The open-weights release is available via Hugging Face and cloud providers. Source: TechCrunch
5. Mistral AI Launches Codestrider 2 for Autonomous Software Engineering
The Paris-based startup released Codestrider 2 on Sept. 25, a specialized coding LLM with 512K context and an integrated agentic loop that can independently debug, test, and refactor entire repositories. It reportedly sets a new state of the art on SWE-bench Verified 2.0. Source: The Verge
Industry
6. Nvidia Unveils Blackwell Ultra B300, Doubling Inference Throughput
At a Sept. 24 data center event, Nvidia announced the B300 “Blackwell Ultra” GPU delivering 2 peta-FLOPs of FP4 compute and 8 TB/s memory bandwidth, targeting next-generation AI agents and real-time multimodal workloads. Shipments are scheduled for Q1 2027. Source: Reuters
Research Highlights
7. DeepMind’s AlphaFold 4 Maps Dynamic Protein Assemblies at Atomic Resolution
In a Nature paper published Sept. 25, DeepMind detailed AlphaFold 4, which predicts not only static structures but also conformational states and protein-ligand interactions at atomic resolution. The model is expected to accelerate drug discovery and protein design by capturing motion and binding. Source: Nature
8. New Multimodal Scaling Laws Predict Emergent Abilities Across Modalities
A joint Stanford and Meta AI study released on arXiv on Sept. 25 shows that cross-modal loss follows a unified power-law across text, image, and audio, with emergent reasoning abilities appearing after 10 billion trained tokens. The paper proposes a single metric for multimodal model scaling. Source: arXiv preprint
Compiled from multiple media reports.








