AI News — Wednesday, September 23, 2026

10
OpenAI Introduces GPT-6 Sol and Luna, Enhancing Frontier AI Capabilities

OpenAI has officially launched GPT-6 Sol and Luna, marking a significant advancement in their frontier AI models with new capabilities.

OpenAI Blogproduct
9
Snorkel AI's Valuation Triples to $3.5B Amid Surging Demand for AI Training Data

Snorkel AI has seen its valuation triple to $3.5 billion, reflecting the booming market demand for high-quality AI training data.

TechCrunchindustry
9
Qualcomm Unveils New AI-Focused Smartphone Chips

Qualcomm has released two new smartphone chips, placing a strong emphasis on enhanced AI processing capabilities for mobile devices.

TechCrunchproduct
8
OmniEdu: Open Foundation Models Revolutionize Learning and Teaching

OmniEdu introduces open foundation models designed to transform educational practices for both learning and teaching.

Hugging Faceopen-source
8
WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

WorldCrafter proposes a novel video world model that achieves consistent generation through implicit 3D-aware memory.

Hugging Faceresearch
7
GameHorizon Suite: Multi-Horizon Data and Evaluation for Advanced Gameplay AI

GameHorizon Suite provides comprehensive multi-horizon data and evaluation tools to advance AI research in gameplay.

Hugging Faceresearch
7
DeepSeek Elastic Compute (DSec): Sandbox Infrastructure for Scalable Agentic Training

DeepSeek Elastic Compute (DSec) offers a sandbox infrastructure designed for effective and scalable agentic AI training.

Lobste.rsindustry
7
Better Prompt Caching for GPT-6 Improves Efficiency

OpenAI announces improvements to prompt caching for GPT-6, aiming to enhance the efficiency and speed of the model.

OpenAI Blogproduct
7
Parallel Cuts Research Time and Cost in Half with GPT-6 Astra

GPT-6 Astra is enabling Parallel Cuts to halve their research time and costs, demonstrating significant efficiency gains in scientific workflows.

OpenAI Blogproduct
6
Realtime-Venus: A Full-Duplex Interaction System with Asynchronous Delegation

Realtime-Venus introduces a novel full-duplex interaction system that utilizes asynchronous delegation for more natural and efficient AI communication.

Hugging Faceresearch
6
Document Retrieval-Aware Chunking (D-RAC): Universal Ingestion of Enterprise Documents

D-RAC offers a universal method for ingesting enterprise documents into retrieval systems through PDF normalization and multimodal markdown conversion.

Hugging Faceresearch
6
Grounded Action Model: 3D Grounding as a Foundation for Robotics

This research proposes a Grounded Action Model that uses 3D grounding as a foundational element for advanced robotics applications.

Hugging Faceresearch
5
VideoGen-Agent: Reinforcing Video Generation Agents for Improved Performance

VideoGen-Agent explores methods to reinforce video generation agents, leading to enhanced performance and more realistic video outputs.

Hugging Faceresearch
5
RULER: Instance-aware Rubric Rewards for SVG Generation

RULER introduces instance-aware rubric rewards to improve the quality and control of SVG generation processes.

Hugging Faceresearch
5
onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents

onPanda presents an efficient method for annotating on-policy alignment data for Large Language Models and agents using token-level correction.

Hugging Faceresearch