Headlines

1. OpenAI Launches GPT-5.5-Turbo with Agentic Voice and Tool Use
OpenAI unveiled GPT-5.5-Turbo, a streamlined flagship optimized for real-time agentic workflows, featuring sub-200ms voice response and native multi-tool orchestration. Early benchmarks show a 34% cost reduction per token compared to GPT-5.2.
Source: The Verge

2. EU and US Sign Landmark AI Safety Pact for Frontier Models
Regulators from Brussels and Washington agreed on a joint framework requiring mandatory adversarial stress tests and third-party audits before deployment of any model above 10^26 FLOPS. The pact also establishes a bilateral rapid-response mechanism for AI incidents.
Source: Financial Times

Model Releases

3. Google DeepMind Unveils Gemini 3 Pro with 10M-Token Context
Gemini 3 Pro introduces a 10-million-token context window and natively generated video as output, achieving state-of-the-art scores on MMMU-Pro and LongBench-10M. The model is available immediately in Vertex AI, with a consumer rollout through Gemini app later this week.
Source: Google DeepMind blog

4. Mistral AI Debuts Mistral Large 3, a 400B MoE for Hybrid Deployment
Mistral Large 3 packs a 400-billion-parameter mixture-of-experts architecture with a 1-million-token context, and includes a quantized version that runs on a single H100 node. It claims 92% of GPT-5.5-Turbo’s performance at roughly one-third the inference cost.
Source: TechCrunch

Industry

5. Anthropic and Microsoft Move Claude onto Azure Government Cloud for Defense Work
Anthropic finalized a deal to deploy Claude-based reasoning assistants within Microsoft’s Azure Government environments for logistic planning and code analysis, drawing criticism from AI safety researchers who fear military applications. The companies said all deployments will follow “human-on-the-loop” approval procedures.
Source: Bloomberg

6. Nvidia Announces Blackwell Ultra B300 with 288GB HBM3e Memory
Nvidia’s Blackwell Ultra B300 delivers 1.5x training throughput over the previous generation and 288GB of HBM3e, with hyperscalers reportedly placing orders exceeding $30 billion. The GPU also adds native support for encrypted federated learning workloads.
Source: Reuters

Research Highlights

7. “Inference-Time Compute Saturates in Reasoning Models”
Stanford and Google DeepMind researchers found that extending chain-of-thought scaling beyond a certain threshold yields diminishing returns on math and logic benchmarks, suggesting a need for new architecture-level reasoning methods. The paper includes a predictive saturation curve that benchmarks against OpenAI and Anthropic reasoning models.
Source: arXiv

8. DroughtGPT: Generative Diffusion Model for Regional Climate Forecasting
IBM Research and the European Centre for Medium-Range Weather Forecasts released a diffusion-based generative model that predicts regional precipitation and drought conditions up to 20 days ahead. It outperforms traditional numerical weather models by 18% on F1 score for extreme events across sub-Saharan Africa.
Source: Nature Climate Change

Compiled from multiple media reports.