Headlines

1. OpenAI Unveils GPT-5.5 “Orion” with Autonomous Agentic Workflows
OpenAI released GPT-5.5 on September 8, 2026, featuring a new “agentic reasoning layer” that enables long-horizon planning, tool use orchestration, and self-correction in live enterprise workflows. Early benchmarks show a 40% improvement on GAIA-style real-world task suites.
Source: TechCrunch

2. United Nations Adopts Global AI Compute Governance Framework
On September 9, 2026, the UN General Assembly endorsed a binding framework for frontier AI compute oversight, including real-time reporting of large training runs and mandatory safety audits for clusters above 10^25 FLOPs. The agreement, cosponsored by the US and EU, comes amidst growing calls for enforceable international rules.
Source: Reuters

3. Anthropic Completes First Coordinated RSP 3.0 Stress Test Across Frontier Models
Anthropic, in partnership with OpenAI and Google DeepMind, ran a multi-organization “alignment red-team” on Sept 7, simulating coordinated failures across deployed agents. The test revealed no critical incidents, but triggered a new protocol for truncating runaway tool-use loops within 200ms.
Source: The Verge

Model Releases

4. Google DeepMind Launches Gemini 3 Ultra with 10M-Token Context
Gemini 3 Ultra, released September 8, 2026, introduces a 10-million-token context window with sparse attention and a new “episodic memory” module. It scores 92.7% on the new Multi-Domain reasoning benchmark and supports real-time multimodal streaming from live video feeds.
Source: Google Blog / Bloomberg

5. Meta Releases Llama 4.5 MoE — Largest Open-Weight Model to Date
Meta open-sourced Llama 4.5 on September 5, a 2.1-trillion-parameter mixture-of-experts model with 210B active parameters. It matches closed-weight competitors on code and math tasks, and includes a permissive license allowing commercial use for organizations under 700M active users.
Source: Meta AI / VentureBeat

Industry

6. Microsoft and Quantinuum Showcase Quantum-AI Hybrid for Electrolyte Design
On September 9, 2026, Microsoft and Quantinuum announced a fault-tolerant quantum circuit combined with a large language model for predicting electrolyte degradation in solid-state batteries. The system identified two novel candidate materials, accelerating a discovery process that usually takes years.
Source: Microsoft Newsroom / IEEE Spectrum

7. NVIDIA B300 “Rubin” Ramp Cuts LLM Inference Costs by 60%
NVIDIA’s B300 GPU, in full production since August, is now deployed in major cloud regions, driving token prices below $0.10 per million for mid-size models. Analysts credit the 288GB HBM3e memory and 5.5x sparse FP4 throughput for the steep cost decline.
Source: CNBC / SemiAnalysis

Research Highlights

8. “AlphaFold 3.5” Predicts Intrinsically Disordered Proteins with Fast-Folding Dynamics
A preprint from DeepMind on September 8 describes AlphaFold 3.5, which extends structure prediction to disordered states and folding pathways at millisecond resolution. The method integrates diffusion models with molecular dynamics, reported in Nature.
Source: Nature / arXiv:2609.08472

9. New Attention Architecture “Linear-Interleaved Transformers” Achieves 1B Context on Single GPU
Researchers from ETH Zurich and Mistral posted a paper showing an interleaved linear-attention layer that reduces memory complexity to O(T) while retaining full attention accuracy on 300B-token pretraining. The architecture enables a 1-billion-token context on a single 80GB H100.
Source: arXiv:2609.08241

Compiled from multiple media reports.