🗞️ AI Daily Briefing — 2026-04-28

🔥 Top Story

China blocks Meta’s $2 billion acquisition of Manus. Beijing’s National Development and Reform Commission ordered the deal’s cancellation, invoking export control and technology transfer laws. Meta acquired the Chinese-founded AI agent startup in December 2025 and had already integrated its tech internally — unwinding the deal will be messy. Two Manus co-founders have been banned from leaving China. This is Xi testing extraterritorial reach over AI assets in a deepening US-China tech war, and it sets a precedent for every cross-border AI acquisition going forward.

🚀 Model & Research News

  • OpenAI ships GPT-5.5: Released April 23, GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users. OpenAI calls it their “smartest model yet” — matches GPT-5.4 latency while using fewer tokens per task. Available in the API as GPT-5.5 and GPT-5.5 Pro.

  • DeepSeek drops V4 in preview: V4-Pro (1.6T params, 49B active) and V4-Flash (284B params, 13B active), both with 1M context. Open source, open weights. V4-Pro beats all open models on math and coding and trails only Gemini 3.1 Pro on world knowledge. DeepSeek also slashed API cache-hit pricing to 1/10th.

  • Claude Opus 4.7 dominates LMArena: Sitting at 1504 Elo — 30+ points ahead of the field. With 5.9M votes across 347 models, it’s the clearest frontrunner the arena has seen.

  • Anthropic unveils Claude Mythos Preview: A new model with advanced cybersecurity capabilities, released alongside the Opus 4.7 wave.

  • AI Scientist-v2 paper accepted at major conference: The first fully AI-generated paper accepted at a top venue — an agentic tree-search system that autonomously proposes hypotheses, runs experiments, and writes papers. A milestone for automated scientific discovery.

🛠️ Tools & Developer Updates

  • SpaceX swoops in on Cursor with $60B deal: Cursor was raising $2B at a $50B valuation (a16z, Thrive, Nvidia) when SpaceX offered a $10B collaboration fee and a path to $60B acquisition. Prediction markets say the option gets exercised. The AI coding tool built by four MIT dropouts just became the most valuable developer tool company ever.

  • LMDeploy vulnerability exploited within 13 hours: CVE-2026-33626, a server-side request forgery flaw in the popular open-source LLM deployment toolkit, came under active exploitation almost immediately after disclosure. Patch now if you’re running it.

  • Anthropic working on Bugcrawl for Claude Code: A new feature that scans repositories for bugs using 10 parallel agents — likely targeting Teams and Enterprise plans.

💰 Funding & Business

🐦 Notable from the Timeline

  • Musk v. Altman goes to trial this week. Nine jurors will deliver an advisory verdict. Ilya Sutskever, Mira Murati, and Satya Nadella are all expected to testify. The New Yorker’s pre-trial investigation, based on leaked Sutskever memos and 200+ pages of Dario Amodei’s notes, has been brutal for Altman’s credibility.

  • Alexandr Wang’s Meta era is real. Muse Spark, Meta’s first proprietary closed-weight model, is now live in Meta AI and rolling out across Facebook, Instagram, WhatsApp, and Ray-Ban glasses. The open-source Llama era is officially over as a primary strategy.

  • ICLR 2026 wraps today in Rio de Janeiro. Google’s TurboQuant paper on KV-cache compression was a highlight. The conference has been a showcase for efficiency-focused research this year.

  • Q1 2026 venture funding hit $300B globally, with AI driving the bulk of the surge. Crunchbase calls it the biggest quarter for startup investment in history.

📊 Benchmark Watch

Claude Opus 4.7 holds the #1 spot on LMArena at 1504 Elo with a 30+ point gap over rivals. The arena now has nearly 6M human preference votes across 347 models. GPT-5.5 benchmarks are still rolling in but early numbers suggest it’s competitive on coding and research tasks without displacing Opus 4.7 at the top. DeepSeek V4-Pro is the new open-source benchmark king, topping all open models on math and code while remaining dramatically cheaper to run.

🎙️ Podcast Highlights

  • TBPN is now OpenAI-owned as of April 2, raising questions about editorial independence. Slate published a piece calling it “the most profitable trend in podcasting” and “a big ugly problem for the rest of us.”

  • All-In and Lex Fridman are likely covering the Musk v. Altman trial this week — expect fireworks as Sutskever and Murati take the stand.

🔗 Worth Reading