🧮 OpenAI's next model solves 10 unsolved math problems 🧮
An unreleased model solves decades-old problems at unprecedented speed.

OpenAI revealed that an internal version of its next major model family solved 10 long-open problems in mathematics and theoretical computer science, some unsolved for nearly 30 years. The results span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics — including proving the existence of non-sofic groups, solving Alain Connes's rigidity conjecture, and clearing three Erdős problems. Every proof has been formally verified in Lean, with chain-of-thought walkthroughs released publicly. In a competitive twist, Anthropic's Levent Alpoge claimed within 24 hours that he reproduced five of the 10 proofs using Anthropic's Fable model with a generic prompt — a capability jump redefining AI's role in frontier mathematical research.

⚡ DeepSeek ships V4 Flash with agentic speed ⚡
The efficient open-weight rollout beats its own bigger sibling on benchmarks.

Chinese AI lab DeepSeek released the production version of V4 Flash, featuring stronger agentic performance and an attached speculative decoding module designed to accelerate inference speed. The model reportedly surpassed the larger V4 Pro Preview on several benchmarks while activating significantly fewer parameters. Competitive pressure on Western labs grows as compute costs drop — DeepSeek continues delivering frontier results through efficiency gains rather than raw scale. V4 Flash is available on Hugging Face as open weights, consistent with DeepSeek's prior distribution approach and reinforcing its position as a leading Chinese AI lab challenging Western counterparts.

🎙️ Microsoft tests MAI Realtime voice model 🎙️
Microsoft's hidden voice model hints at replacing OpenAI's technology.

Microsoft's first native real-time voice model, MAI Realtime, has surfaced as a hidden early-access entry in the company's MAI Playground. The listing describes a bidirectional, full-duplex system capable of listening and speaking simultaneously — a significant architectural step beyond current turn-based voice assistants. Two voices are available, both reported to sound noticeably more natural than Microsoft's existing Copilot voice mode. The model is expected to roll out through Microsoft Foundry and Copilot voice, though no official timeline has been announced. This reduces Microsoft's dependence on OpenAI for voice capabilities and positions it to compete directly with OpenAI's Advanced Voice Mode and Google's Gemini Live.

INTERESTING TO KNOW

🚀 Opus 5 renders Tolkien in 5,500 code lines 🚀

Andrej Karpathy showcased the creative integration capabilities of Anthropic's Opus 5 by asking it to create a Three.js render of the first paragraph of The Lord of the Rings, granting a 1-million-token budget. The model returned approximately 5,500 lines of code that procedurally generated polygon assets and animated them according to the narrative — a capability jump in long-form structured code generation. The result demonstrates Opus 5's ability to orchestrate complex creative and engineering tasks end-to-end when given large context windows.

📱 Gemini Desktop adds image and video generation tabs 📱

Google is rolling out dedicated image and video generation tabs alongside a new camera attachment feature for the Gemini desktop application. The upgrades add competitive pressure on ChatGPT's multimodal desktop experience, signaling Google's strategy to make Gemini a one-stop creative and productivity hub rather than a text-only chatbot. The camera feature suggests tighter integration with device hardware for real-time visual input, positioning Gemini Desktop to compete more directly with integrated AI tools that combine generation and analysis in a single interface.