1. Moonshot's Kimi K3 undercuts Claude and GPT on some coding tests: Moonshot, the Chinese lab behind the Kimi assistant, released K3, a free frontier-class model you can download and run yourself rather than rent from a US lab. It is live now on kimi.com and the API, with full open weights due July 27. Moonshot's own numbers show it topping GPT-5.6 Sol on browsing and automation tasks and beating Anthropic's Claude on a popular frontend coding leaderboard. Developer Simon Willison found it capable but a token hog, burning past 16,000 output tokens on a single graphics test, and Moonshot suggests 64 or more accelerators for serious use. The skeptical read: The benchmark wins are self-reported vendor tables, and one independent tester found it costly to actually run, so open here does not mean cheap. 2. Apple takes OpenAI's hardware division to court: Apple sued OpenAI, its io hardware unit, and two former Apple staffers, including ex-design chief Tang Tan, claiming they funneled confidential hardware secrets into OpenAI's device push. Apple called the unit rotten to its core; OpenAI denies it. The skeptical read: This is one side's complaint, not a verdict, and OpenAI disputes the claims, so treat the theft allegations as unproven until a court says otherwise. 3. Murati's ex-OpenAI lab gives away its first open model, Inkling: Mira Murati's new lab, staffed with ex-OpenAI researchers, released its first open-weight model, Inkling, a multimodal system with a 1-million-token context window and full weights free to download. The lab says it leaned on other open models, including Moonshot's Kimi, to bootstrap its training data. 4. xAI's Grok coding tool was uploading entire private repos: An independent wire-level teardown found that Grok Build, the coding tool from Elon Musk's xAI, uploaded whole repositories to a Google Cloud bucket regardless of what the agent actually read, contents unredacted. If you ran it on a private codebase, xAI got a copy of files you never told it to open. xAI has since open-sourced the tool. The skeptical read: xAI hasn't confirmed what it does with the stored data or why whole-repo uploads were necessary at all. 5. OpenAI splits its Codex coding tool into three AI models: OpenAI's Codex, its AI coding assistant, now routes work across three new GPT-5.6 models built into a broader ChatGPT superapp push: Sol for hard problems, Terra for everyday builds, Luna for fast tasks. Codex usage reportedly grew 10x in six months to 7 million users, even as some users report the update deleting their files. The skeptical read: OpenAI's explanation for the file-deletion reports downplays the issue as an edge case rather than addressing why an agent should ever delete user files unprompted. | 