GPT-5 Cracks Quantum Physics as Gemini 3 "Deep Think" Smashes GPT-5.1 on AGI Benchmarks

AI Research & Scientific Contributions

GPT-5 Contributes Key Insight to Peer-Reviewed Physics Paper: GPT-5 generated a critical insight for a paper accepted in Physics Letters B, demonstrating AI's potential in scientific research. The paper, co-authored by researchers, used GPT-5 in a Generator-Verifier setup to derive new operator integrability conditions in quantum physics.

• GPT-5 generated the key insight for a paper accepted to Physics Letters B, a serious and reputable peer-reviewed journal
  • Tweet by Mark Chen
  • Tweet by Steve Hsu
  • arXiv Paper
  • ScienceDirect Publication
  • Google Drive Supplementary Material

AI Model Releases & Benchmarks

Gemini 3 "Deep Think" Outperforms GPT-5.1 on ARC-AGI-2: Google’s Gemini 3 "Deep Think" mode scored 45.1% on ARC-AGI-2, more than doubling GPT-5.1’s performance. The mode leverages System 2 search/RL techniques to enhance reasoning in complex domains like math and science.

• Gemini 3 "Deep Think" benchmarks released: Hits 45.1% on ARC-AGI-2 more than doubling GPT-5.1
• Gemini 3 Deep Think now available

Claude Opus 4.5 Trial Access Rolled Out: Perplexity AI has granted surprise trial access to Claude Opus 4.5 for select users, allowing them to test its advanced capabilities under undisclosed usage limits.

• Surprise Trial Access to Claude Opus 4.5 - Anyone Else See This?

AI Infrastructure & Hardware

OpenAI’s $4.6B GPU Supercluster in Australia: OpenAI and NextDC are building a 550MW hyperscale GPU supercluster in Australia by 2027, aimed at training next-gen AI models with data sovereignty and low latency for the APAC region.

• BREAKING: OpenAI to build massive $4.6 Billion "GPU Supercluster" in Australia (550MW Hyperscale Campus by 2027)
• BREAKING: OpenAI and NextDC to build massive $4.6 Billion "GPU Supercluster" in Australia (Sovereign AI)
  • Forbes Coverage

NVIDIA Achieves 10x MoE AI Performance Boost with GB200 ‘Blackwell’: NVIDIA’s GB200 NVL72 servers deliver a 10x performance leap in Mixture-of-Experts (MoE) AI workloads, enabled by co-design breakthroughs for faster training and inference.

• NVIDIA Shatters MoE AI Performance Records With a Massive 10x Leap on GB200 ‘Blackwell’ NVL72 Servers
  • WCCFTech Article

AI Model Optimizations & Frameworks

Qwen Next CUDA Optimizations Merged into llama.cpp: Speed optimizations for Qwen Next have been integrated into llama.cpp, improving performance for CUDA-based inference with faster response times.

• speed optimizations for Qwen Next on CUDA have been merged into llama.cpp
  • GitHub PR #17584

AI Product Updates & Features

Mistral AI Introduces Custom Instructions per Project: Mistral AI now allows users to set custom instructions per project, enhancing flexibility for tailored AI interactions in specific workflows.

• Custom Instructions per Project 🎄
  • Demo Video

Mistral AI Teases Upcoming Coding Advancements: Mistral AI hinted at new coding-related features to be announced in the coming days, suggesting improvements in AI-assisted programming capabilities.

• More on coding in a few days... 👀
  • Mistral AI Tweet

Mistral AI Hosts Webinar on Production-Grade AI Systems: Mistral AI is organizing a webinar covering best practices for designing, building, and optimizing production-ready AI systems, including real-world use cases and a live demo.

• Building Production-Grade AI Systems