1. Moonshot's Kimi K3 Pushes Open Weights to the Frontier
Chinese startup Moonshot AI launched Kimi K3 on July 16 — a 2.8-trillion-parameter Mixture-of-Experts model with a 1M-token context window, native multimodal input and reasoning always on. The API is live at $3 per million input tokens and $15 per million output, with full open weights dated for July 27. Moonshot concedes K3 still trails Claude Fable 5 and GPT-5.6 Sol overall, but it debuted at #1 on the Frontend Code Arena at 1679 Elo — a 17-place jump from Kimi K2.6.
2. Xi Unveils WAICO, a China-Led Global AI Body
At the World AI Conference in Shanghai on July 17, Xi Jinping announced the World Artificial Intelligence Cooperation Organization — an intergovernmental body headquartered in Shanghai with 29 founding member countries including Indonesia, Brazil, Malaysia, South Africa, Russia and Pakistan. Xi said AI development "should not be a solo performance by a single country," a thinly veiled jab at US dominance. China pledged 5,000 AI training slots for developing countries over five years.
3. Meta Pulls Instagram's Muse Image After Consent Backlash
Meta yanked Muse Image from Instagram on Friday, admitting it "missed the mark." The tool, live only days, let anyone generate AI images from any public account's photos by @-mentioning it — with every public account opted in by default and no notice when photos were used. SAG-AFTRA told members to opt out, calling anything short of clear opt-in "unacceptable." CAA piled on. Muse Image remains available via WhatsApp and the Meta AI app.
4. Apple Sues OpenAI Over Trade Secret Theft
Apple filed suit in Northern California on July 10, alleging OpenAI ran a systematic scheme to lift its hardware IP. The complaint says OpenAI hardware chief Tang Tan — a former Apple VP — directed job candidates still employed at Apple to bring "actual parts" to interviews for show-and-tell, and that OpenAI coached departing staff on evading Apple's exit security. Over 400 former Apple employees now work at OpenAI. The two partnered on ChatGPT-in-iOS back in 2024.
5. Publishers and Scott Turow File Class Action Against Google
Hachette, Cengage and Elsevier, joined by novelist Scott Turow, sued Google on July 10 for willful copyright infringement in training its Gemini models. The complaint alleges Google secretly repurposed millions of works submitted to Google Books for strictly limited uses, and downloaded unauthorized scrapes of the web including known pirate sources and paywalled content. Plaintiffs seek statutory damages plus disgorgement of Google's profits on behalf of a proposed class of authors and publishers.
6. OpenAI Ships GPT-Live, Ending Turn-Based Voice
GPT-Live rolled out July 8 as the new engine behind ChatGPT Voice. Its full-duplex architecture listens and speaks simultaneously, so it can drop in a "mhmm," interject mid-sentence, or simply stay quiet while you think. Under the hood it is two layers: a continuous interaction layer running the conversation, and a delegation layer that hands hard problems to GPT-5.5 in the background. GPT-Live-1 is default for Go, Plus and Pro; free accounts get the mini. API access is waitlisted.
7. Gemini 3.5 Pro Slips a Third Deadline
Google's rebuilt Gemini 3.5 Pro missed its widely reported July 17 target. Third-party reporting says DeepMind scrapped the original model after finding structural failures in recursive tool-calling and SVG generation, then failed reliability bars on the rebuild — frequent hallucinations, and benchmarks short of GPT-5.6. The public Gemini API still lists only gemini-3.5-flash and gemini-3.1-pro-preview. Worth flagging: every specific claim here, including the date, comes from leaks and unnamed sources, not from Google.
8. Global Venture Funding Hits a Record $510B in H1, Mostly AI
Crunchbase puts first-half 2026 global startup investment at $510 billion — more than the $440 billion raised in all of 2025. OpenAI and Anthropic alone took $217 billion of it, 43% of every startup dollar. The tail is still thick: defence-AI firm Helsing raised $1.8 billion on July 13 with JPMorgan, Lightspeed and Iconiq leading, and Together AI closed $800 million at an $8.3 billion post-money valuation.
// KEY TAKEAWAYS
Two themes dominated the week. First, the open-weight gap is closing fast — Kimi K3 took the top frontend-coding slot from a US frontier model, and the UK AI Security Institute now puts the open-versus-closed distance at four to seven months, down from six to ten. Second, the legal and political bill is coming due: Apple v OpenAI, publishers v Google, Meta's consent retreat on Muse, and Xi's WAICO all landed within nine days of each other. Meanwhile Google quietly missed a third Gemini 3.5 Pro deadline, a reminder that shipping frontier models reliably is still harder than announcing them.