Headlines

1. OpenAI Launches GPT-5.5 “Turing” with Two-Million-Token Context and Native Tool Use
OpenAI’s newest flagship model, unveiled on September 30, targets autonomous agent workflows, supporting a 2-million-token context and long-horizon planning. Early evaluations report a 78% success rate on the GAIA agentic benchmark, up from 51% for GPT-5. Source: The Information, TechCrunch

2. EU AI Act’s “High-Risk” Compliance Deadline Shakes Up Enterprise AI Market
As the EU’s landmark regulation takes effect for high-risk AI systems, providers including Google, Microsoft, and Aleph Alpha race to certify models and document training data. Analysts expect a wave of compliance-driven model updates, as well as modest price increases for enterprise customers. Source: Financial Times, Reuters

Model Releases

3. Google DeepMind Unveils Gemini Ultra 2.8 with Multimodal Active Perception
Gemini Ultra 2.8 improves real-time video understanding and robotic control, achieving state-of-the-art results on MMMU-Pro and a 31% error-rate reduction on the ALOHA manipulation benchmark. It debuts in Vertex AI and the Google Colab research tier. Source: DeepMind Blog, TechCrunch

4. Anthropic Releases Claude 4.6 Opus with Self-Reflective Memory
Claude 4.6 Opus introduces session-persistent memory across long-running tasks, cutting repeated instruction failures by 40% on long-horizon writing and coding workflows. Anthropic emphasizes “steerable humility” and fine-grained privacy controls for enterprise deployment. Source: Anthropic Newsroom, CNBC

Industry

5. NVIDIA Blackwell Ultra Begins Mass Shipment, Powering New AI Hyperscaler Deployments
NVIDIA announced that its B300 “Blackwell Ultra” GPUs are now shipping to Azure and CoreWeave data centers, delivering 2.2x inference performance over H100 for Mixture-of-Experts models. Supply-chain sources describe allocation as “tight but manageable”. Source: Reuters, Data Center Dynamics

6. OpenAI and Microsoft Retool $15B Compute Deal Amid Sovereign AI Push
Under a renegotiated agreement, Microsoft gains cloud-exclusive rights for select OpenAI models in non-EU regions, while OpenAI may use additional Oracle and Google Cloud clusters for training. The restructuring responds to antitrust scrutiny and expanding national AI infrastructure initiatives. Source: Bloomberg, The Verge

Research Highlights

7. “Seed-Distillation for Weak-to-Strong Reasoning Transfer”
Researchers from DeepMind and Stanford show that compact 7B-parameter student models can outperform large teacher models on specialized reasoning tasks by distilling chain-of-thought trajectories from an ensemble of “weak” teachers. The paper reports 97% accuracy on MATH-500 without additional reinforcement learning from human feedback. Source: arXiv, MIT Technology Review

8. “Locally Interpretable Neural Features Revealed by Sparse Autoencoders in Vision Transformers”
Anthropic and UC Berkeley researchers release a large-scale analysis of vision-language models, mapping 50M sparse autoencoder features to human-interpretable concepts such as object parts, lighting, and 3D geometry. The findings offer a potential safety mechanism for reducing visual compositional hallucinations. Source: arXiv, AI Safety Newsletter

Compiled from multiple media reports.