OpenAI Cuts Pricing in Half With GPT-6 Sol and Luna

A federal appeals court handed GitHub, Microsoft, and OpenAI a massive win, ruling that Copilot creates new code rather than scrubbing copyright tags off existing repos. The DMCA challenge is officially dismantled.

Alibaba Cloud unveiled the V900 AI chip, linking up to 500,000 accelerators into single supernodes while expanding into Finland and Turkey. Scaling custom silicon globally is becoming the standard cloud playbook.

Samsung fired up 2nm prototype silicon in Texas to fabricate Tesla's AI5 processor. Running the exact same neural architecture across autonomous vehicles and Optimus humanoids simplifies compute dramatically.

Google DeepMind introduced server-side memory for Private AI Compute, syncing cross-device context through hardware vaults where Google holds zero decryption keys. Persistent memory without persistent surveillance.

DeepSeek open-sourced DeepSeek-V4-Flash-Vision-Exp, a 305-billion-parameter multimodal model built specifically for software agents navigating interfaces and parsing visual documents under an MIT license.

Researchers built EAGLE, an AI model that spots esophageal cancer on standard, noncontrast chest CT scans with 98.5% specificity across 80,000 patients. Repurposing routine scans could transform early diagnosis.

SpaceXAI announced Grok 4.7, expanding base weights and reinforcement learning for multi-hour coding and document drafting. Keeping prices flat while claiming double the speed puts heavy pressure on rivals.

Alphabet subsidiary Intrinsic open-sourced Intrinsic Core, providing industrial robotics developers with real-time control, digital twin simulation, and ROS compatibility so teams stop reinventing kinematics.

Researchers published CHOREO, a training-free framework that chains nearly 3,000 disparate humanoid skills with a 95.4% success rate in simulation, proving motion trajectories beat constant fine-tuning.

Tsinghua researchers introduced Cache-to-Cache, letting models swap raw KV cache states directly instead of translating thoughts into tokens. Bypassing text cuts latency up to 14x and sharpens accuracy.

Tongyi Lab launched Qwen3.7-Max, an agent foundation model that logged 35 hours and 1,158 tool calls without crashing. Decoupling runtimes from validators keeps automated enterprise loops from falling over.

China's Supreme People's Court published dispute guidance on artificial intelligence, directing judges to stick to existing civil and patent law rather than inventing special doctrines for algorithms.

JetBrains announced JetBrains Air, consolidating cloud agents, shared context, and budget governance into one platform. Coordinating rogue codebots across enterprise teams is the new developer tool frontier.

Light Origins launched Light-O1, an embodied foundation model trained on 100,000 hours of human action footage. Scaling up video tokens significantly cut prediction errors across different humanoid platforms.

Google updated the Antigravity SDK to support local offline swarms using Gemma 4. Developers can now run on-device builders coordinated by a cloud model, balancing operational privacy with heavy reasoning.