Headlines

1. OpenAI Unveils GPT-Next, Leaping to Frontier-Level Multimodal Reasoning
OpenAI introduced GPT-Next on September 24, claiming major gains in long-horizon planning, live web tool use, and native video understanding. The model reportedly scores 85% on the new FrontierBench-2026, surpassing previous state-of-the-art by 11 points.
Source: Reuters

2. U.S. and EU Strike Landmark AI Governance Deal Ahead of Global Summit
Negotiators finalized a joint framework for frontier model oversight, including binding high-compute incident reporting and mandatory CAPA-driven red-teaming. The agreement, announced September 25, is expected to shape the UN AI Summit in October.
Source: Politico EU

Model Releases

3. Anthropic Launches Claude 4.5 Opus with 2M-Token Context and Adaptive Memory
Claude 4.5 Opus arrived on September 23, featuring persistent across-session memory, encrypted tool orchestration, and a 40% cost-per-token reduction. Early benchmarks show it rivaling GPT-Next on professional exams and agentic coding tasks.
Source: The Information

4. Meta Releases Llama 4 Ultra, Its Largest Open-Weight Offering
Meta open-sourced Llama 4 Ultra on September 26, a 1.2T-parameter sparse mixture-of-experts model with a permissive license and full fine-tuning tooling. Independent eval runways show it narrowing the gap with proprietary models on math and code generation.
Source: TechCrunch

Industry

5. Microsoft and OpenAI Ship “Copilot Agent Mesh” for Cross-Tool Enterprise Workflows
The enterprise agent platform, generally available September 25, lets Copilot orchestrate workers across Microsoft 365, Salesforce, and SAP. More than 40% of Fortune 500 companies joined the pilot, citing agent-to-agent API interoperability.
Source: Bloomberg

6. Nvidia Begins Mass Production of B300 “Blackwell Ultra” Accelerators
Nvidia confirmed September 24 that the B300, featuring 288GB HBM3e and FP4 tensor cores, is shipping to major cloud providers. At the same time, Dell announced a liquid-cooled 8-GPU rack doubling inference throughput for frontier-class models.
Source: The Verge

Research Highlights

7. “Test-Time Scaling is a Poor Universal Law for Long-Horizon Reasoning”
A joint Stanford–Anthropic study, posted September 24 on arXiv, shows diminishing returns for chain-of-thought compute beyond a task-specific threshold. The authors propose a temperature-verified search policy that yields up to 23% accuracy gains on PhD-level science probes.
Source: arXiv

8. DeepMind’s AlphaProof 2 Solves Two Unsolved Problems in Combinatorics
In a Nature paper published September 25, DeepMind described AlphaProof 2, which provides constructive proofs for a 1979 Ramsey theory conjecture and a new bound on the Erdős–Straus primitive set problem. The system autonomously generated the key lemmas and human-verifiable formal proofs.
Source: Nature

Compiled from multiple media reports.