AI Frontier Daily — September 5, 2026

Headlines

1. OpenAI Releases GPT-5.5 “Aurora” with Native Agentic Computer Use
OpenAI announced GPT-5.5 Aurora, a multimodal model with a 2-million-token context window and an integrated agentic loop that can plan and execute multi-step browser tasks without external scaffolding. Initial results on the updated GAIA validation split show 73.4%, up from GPT-5’s 65.1%.
(Source: OpenAI Newsroom, The Verge)

2. EU Mandates “Frontier AI Stress-Test” Regime Under AI Act Article 56
The European AI Office published its Article 56 inspection protocol on Sept 4, requiring frontier model providers to submit continuous risk-management logs and independent red-team results every 180 days. Anthropic and Google DeepMind confirmed they will pilot the regime with their next frontier-model releases.
(Source: EURACTIV, Financial Times)

3. DeepSeek R2 Becomes Top Open-Weights Model on SWE-bench and ARC-AGI-2
DeepSeek released R2, a 670-billion-parameter sparse Mixture-of-Experts model, with public weights on Hugging Face. It achieved 87.6% on SWE-bench Verified and 31.4% on ARC-AGI-2, clearing the previous open-weights best by wide margins.
(Source: Hugging Face Blog, MIT Technology Review)

Model Releases

1. Google DeepMind Launches Gemini 3.5 Pro with “Persistent Memory”
Gemini 3.5 Pro rolled out to Vertex AI on Sept 3, adding a 2-million-token context and optional multi-session memory that compresses personal and project context into compact learned keys. DeepMind says the feature boosts longitudinal task completion by 38% in internal project-lifecycle tests.
(Source: Google DeepMind Blog)

2. Mistral AI Ships Codestral 4, a 200B Code Model with Repo-Level Reasoning
Codestral 4 uses a 200B-active-parameter hybrid architecture and supports a 16M-token repository context, achieving 58.2% on the new RepoBench-Lite suite. Mistral says the model will be commercially available on Azure and on-prem from Oct 1.
(Source: Mistral AI Blog)

Industry

1. Microsoft and Nvidia Announce $12B Frontier AI Compute Joint Venture
Microsoft and Nvidia agreed to build three 10-gigawatt liquid-cooled AI campuses in Wyoming, Ontario and Norway, optimized for Nvidia Blackwell Ultra GPUs and Microsoft’s custom Maia accelerators. The first cluster is expected online in Q1 2027.
(Source: Bloomberg, The Information)

2. Anthropic and AWS Expand “Claude for Government” with FedRAMP-High Region
Anthropic unveiled a new sovereign AWS region for U.S. federal agencies, offering Claude Opus 5.2 under FedRAMP High authorization and dedicated inference capacity. The program follows recent G7 guidance on state-aligned AI systems and secure public-sector adoption.
(Source: Nextgov/FCW, Reuters)

Research Highlights

1. “Predictive Vector Quantization for Durable Long-Horizon Memory”
A Stanford, UC Berkeley and DeepMind team introduced PVQ, a memory layer that compresses episodic state into learned discrete keys, reducing memory overhead by 93% and boosting ALFWorld task success from 72% to 84%. The preprint was released on arXiv and accepted to NeurIPS 2026.
(Source: arXiv preprint