AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Key topics and reactions
Apple Sues OpenAI, Chief Hardware Officer Tang Tan Over Alleged Hardware Trade Secret Theft
On July 10, 2026, Apple sued OpenAI and two former Apple employees in the U.S. District Court for the Northern District of California, alleging OpenAI systematically misappropriated iPhone-related hardware trade secrets to build its hardware business. Defendants include OpenAI Foundation, OpenAI Group PBC, Chief Hardware Officer and former Apple vice president Tang Tan (Tang Yew Tan), former senior system electrical engineer Chang Liu, and OpenAI-acquired hardware subsidiary io Products. Apple is seeking an injunction, destruction and return of stolen materials, product redesign, a jury trial, and damages.
According to the complaint, Chang Liu did not return his company laptop and exploited authentication and network-storage bugs to access Apple's internal network, downloading dozens of confidential hardware files totaling more than 1,000 pages, including circuit board manufacturing and testing details. Tang Tan allegedly emailed Apple supplier information and internal industry summaries to himself before leaving. Apple further claims OpenAI advised departing employees not to reveal their next employer and coached them on avoiding the standard two-week notice period, and that interviews involved a 'show and tell' in which candidates brought physical parts such as batteries, logic boards, SIPs and shields. Apple says OpenAI now employs more than 400 former Apple staff.
The suit reflects a sharp turn in the companies' relationship. After partnering in 2024 to integrate ChatGPT into Apple Intelligence, ties cooled after OpenAI acquired Jony Ive's io Products for roughly $6.4–6.5 billion in 2025 and entered hardware. Tan's team has signed a manufacturing deal with iPhone assembler Luxshare. OpenAI and Tan declined to comment. Apple is reportedly shifting toward Google's Gemini for Siri. (Sources: CNBC, Reuters, WIRED, Axios.)
OpenAI Releases GPT-5.6 and Integrates Codex Into ChatGPT
OpenAI announced GPT-5.6 and integrated Codex into ChatGPT. The desktop app now switches seamlessly between three modes — Quick Chat, Codex and ChatGPT Work. Developers reported completing complex 1,000-line tasks with almost no manual intervention, instant generation of production-grade websites via the Sites feature, and strengthened GitHub integration and Computer Use browser control.
OpenAI says its healthcare-oriented 'Luna' configuration surpasses the highest-reasoning GPT-5.5 setting while cutting costs to roughly one twenty-fifth, and that GPT-5.6 reached a first-time tied top ranking on the Code Arena: Frontend coding evaluation.
Feedback praised the stability of tool use and long-running agent execution, though some noted lingering regressions from the GPT-5.5 era. Users generally described GPT-5.6 as a clear step forward while noting Claude Code still edges it on fine-grained CLI control. GPT-5.6 has also rolled out to Perplexity, Perplexity Computer and Microsoft 365.
xAI Releases Grok 4.5 and Free Grok Build for Agentic Coding
xAI's Grok 4.5 and Grok Build gained traction in agentic coding. Developers reported that a two-line prompt could drive an end-to-end loop from design document to TODO breakdown to implementation, producing space-battle games, racing games and voxel worlds within an hour.
Grok 4.5 recorded the top score on Perplexity's WANDR evaluation with class-leading token efficiency of 1.9M tokens per task, and ranked first on the SWE-Atlas-QnA benchmark. Grok Build was made free for all users.
Users welcomed access to a full agentic harness at no cost, while noting realistic failure rates — one reported seven of eight builds succeeding — and recommending external tools for complex 3D asset generation. Grok 4.5 was also adopted as the orchestrator model in Perplexity Computer.
Anthropic Adds Model-Controlled In-App Browser to Claude Code Desktop
Anthropic added a sandboxed in-app browser to the Claude Code desktop app that Claude itself can control. The model can open documents, designs, production apps, social media and any website, reading content and performing clicks and form entries autonomously, much as it works with a local dev server.
All actions occur within the sandbox, with OS-level restrictions such as Apple Events blocking (adjustable in settings), and users choose whether to persist sessions such as login state. The integration removes prior manual copy-and-paste for external content and connects to external tools and data via MCP (Model Context Protocol), which Anthropic positions as differentiating it from browser-plugin competitors.
Developers welcomed the ability to open URLs and debug directly, saying it makes agentic workflows more practical and lets Claude find and self-verify bugs in running apps. A reported bug causes duplicated viewport display when using resize_window (avoidable via reload or manual resize), and some noted feature gaps between the desktop and terminal versions. (Source: 9to5Mac.)
Category highlights
Google Advances Gemma 4 for On-Device Multimodal and Agentic Use
Google pushed Gemma 4 for fast multimodal and agentic operation on offline devices, reporting a 5x speedup for single-A10G inference in a Hugging Face challenge. GPT-5.6 (Luna/Terra/Sol) advanced on both performance and cost, while Meta's Muse Spark 1.1 reached ninth place on Code Arena, updating the cost-efficiency frontier.
Alibaba Tongyi Unveils Wan-Streamer Real-Time Omnimodal Model
Alibaba Tongyi introduced Wan-Streamer, a real-time bidirectional omnimodal model that synchronizes audio and video output at roughly 550ms latency. It is part of a broader trend toward low-latency interactive generation, alongside models such as Vidu S1, moving real-time omnimodal generation toward practical use.
Creators Compare GPT-5.6 Sol and Fable 5 on Animation Generation
Higgsfield AI published multiple side-by-side comparisons of GPT-5.6 Sol and Fable 5 across styles including samurai, cartoon, action, paper-cut and monster battles using the Seedance 2.0 pipeline. Creators generally rated Fable 5 ahead on aesthetics, logical consistency and detail, positioning Sol as the more cost-efficient option.
TTS Arena Adds Voice Styles as ElevenLabs and Cohere Expand Audio Tools
TTS Arena added Voice Styles with tone, emotion, age and accent control. ElevenLabs supplied an AI voice agent to Greece's Alpha Bank, while Cohere demonstrated local Arabic transcription running on a Mac, extending the shift toward on-device and localized audio processing.
HeyGen, Perplexity Computer and Video Tools Add Features
HeyGen announced a Figma-to-video integration and is supplying avatar technology to Telemundo. Perplexity Computer added the Grok 4.5 orchestrator and credit analysis features. On video, Vidu S1 offers real-time interactive generation, Pika added VFX skills, and Google Flow's Omni Flash added multi-shot support.
Key trends
Kling AI NEXTGEN Awards Held in Seoul as ICML 2026 Opens
The Kling AI NEXTGEN Awards took place in Seoul, drawing more than 12,700 submissions. ICML 2026 held its first day of workshops. Chinese image and video models including Seedream 5.0 Pro (fourth overall on Image Arena), Seedance 2.0, Kling, Wan and Vidu continued to rank highly and see wide adoption in production work.