🗞️ AI Daily Briefing — 2026-07-13

🔥 Top Story

Apple sues OpenAI for trade secret theft in blockbuster lawsuit. Filed July 10-11 in Northern California federal court, Apple alleges that OpenAI — including hardware chief Tang Tan (a former Apple VP) — systematically stole Apple’s trade secrets to build its own consumer hardware. The complaint claims Tan directed Apple employees interviewing at OpenAI to bring “actual parts” for “show and tell” sessions, and names over 400 former Apple employees now at OpenAI. This is a dramatic reversal of the Apple–OpenAI partnership that began in 2024, and escalated after OpenAI acquired Jony Ive’s io Products for $6.4B. (CNBC) (TechCrunch) (Fortune)

🚀 Model & Research News

  • Anthropic extends Fable 5 included access again — now through July 19. The metered credits switch ($10/$50 per Mtok) that was supposed to kick in today (July 13) has been pushed to July 20. This is the second extension; the original cutoff was July 7. Claude Code weekly rate limits also got a 50% bump through the same date. Anthropic says the delay is capacity-driven, not permanent. (Explainx) (Fable5.app)

  • Gemini 3.5 Pro confirmed for July 17 launch after full architectural rebuild. Google DeepMind scrapped the Gemini 2.5 Pro base entirely, rebuilding from scratch with a 2M token context window, a “Deep Think Reasoning Layer” (locked behind the $250/month Ultra plan), and autonomous workflow capabilities. The rebuild aims for parity with GPT-5.6 and Fable 5. Meanwhile, four senior DeepMind researchers — including AlphaFold’s John Jumper — have left for Anthropic. (BigGo Finance) (HackerNoon)

  • SpaceXAI launched Grok 4.5 publicly on July 8 — its first flagship model since SpaceX absorbed xAI. Built on a 1.5T-parameter foundation and trained on real Cursor session data, Elon called it “an Opus-class model, but faster.” Priced at $2/$6 per Mtok (input/output), significantly undercutting Opus 4.8’s $5/$25. Not yet available in the EU. (Axios) (TechCrunch)

  • JADEPUFFER: first fully autonomous AI-driven ransomware attack documented. Sysdig disclosed an LLM agent that autonomously performed recon, stole credentials, moved laterally, escalated privileges, and encrypted data — adapting to failures in real time (failed login to working fix in 31 seconds). It exploited a Langflow vulnerability (CVE-2025-3248) and encrypted 1,342 Nacos service configs. Security researchers are calling this the arrival of “agentic threat actors.” (Bleeping Computer) (Sysdig)

  • “Opus 5” leak surfaces in Cursor. A model called “Claude Honeycomb EAP” briefly appeared in the Cursor AI editor for some users before disappearing. Its spec sheet matches Fable 5’s architecture (1M-token context, extra-high-effort mode). Community speculation labels this a potential “Opus 5” or next-tier “Mythos” model. Anthropic has not confirmed anything. (TechTimes)

  • DeepSeek V4 expected mid-July alongside a new peak-valley pricing mechanism. DeepSeek also raised $7B (51B yuan) — the largest single AI fundraise in China’s history — and is developing its own inference chip to reduce reliance on NVIDIA and Huawei. (Bloomberg)

  • Google DeepMind presents 130+ papers at ICML 2026, including Project Astra (universal AI assistant) and AlphaEarth Foundations (planetary geospatial insights from satellite embeddings). (Google Research)

🛠️ Tools & Developer Updates

  • Terry Tao’s blog post on coding agents hit #1 on Hacker News. The Fields Medalist used AI agents to port his 27-year-old Java math applets to modern JavaScript in hours — and the agents found two bugs he’d missed. He’s now considering adding interactive visualizations to future papers, calling the downside risk “acceptable” since they’re supplements, not mission-critical. (terrytao.wordpress.com) (TechTimes)

  • NVIDIA Cosmos 3 and Isaac GR00T announced. Cosmos 3 is described as the first world foundation model unifying synthetic world generation, vision reasoning, and action simulation. GR00T enables robots to understand natural language and perform complex multi-step tasks with vision-language-action reasoning. (NVIDIA Blog)

  • dbt Wizard launched — an AI agent for governed data development that handles the full lifecycle (investigation, building, validation, shipping), grounded in your dbt project’s lineage, tests, contracts, and metric definitions. Live demos scheduled for July 15 and 22. (dbt Labs)

  • PostHog’s AI suite deepens with an AI assistant that queries product data, writes SQL, creates feature flags, and sets up surveys — plus AI session recording analysis that clusters rage clicks, errors, and confusion. All available via MCP for Claude Code and VS Code. (PostHog)

  • Hugging Face ships Transformers v5.13.0 with major new model support, broader export/kernel tooling, and generation/attention/cache fixes. Also launched a hardware filter on the Models page — filter by GPU, CPU, or Apple Silicon to find compatible models. (Releasebot)

💰 Funding & Business

  • Anthropic has overtaken OpenAI in revenue — running at ~$47B annualized vs. OpenAI’s projected $25–33B. Claude Code alone grew from $1B to $2.5B ARR in about two months. Enterprise makes up ~85% of Anthropic’s revenue. (CNBC) (Epoch AI)

  • OpenAI confidentially filed its S-1 with Goldman Sachs and Morgan Stanley leading. Private valuation sits at ~$730B; if it prices near that, it would be the largest tech IPO ever. Reuters reports the timeline may slip to 2027. (CNBC) (Forbes)

  • SK Hynix’s ADR listing raised $28–29B on July 10 — the largest ADR offering in history — giving the memory chipmaker direct US capital market access at peak AI demand.

  • Meta stock surged 7% on July 10 after an internal memo revealed plans to double total computing power by 2027, including a $10B, 1GW data center in Alberta, Canada backed by long-term supply deals with Samsung.

  • SambaNova closes $1B Series F at $11B valuation. Led by General Atlantic with BlackRock, Intel Capital, QIA, T. Rowe Price, Battery Ventures, and Vista Equity participating. JPMorgan Chase selected SambaNova as an inference infrastructure partner. (Bloomberg)

  • Qualcomm in talks to acquire Tenstorrent for $8–10B. Would give Qualcomm real AI chip competitiveness against NVIDIA and AMD. Tenstorrent uses RISC-V and was founded by chip legend Jim Keller. (Yahoo Finance)

  • US HHS will use ChatGPT for fraud detection — analyzing audit reports from all 50 states to hunt for fraud across ~$2.1T in annual Medicare and Medicaid spending. (Medium)

🐦 Notable from the Timeline

  • Alexandr Wang (Meta Chief AI Officer) told a Meta town hall that their next model “Watermelon” is matching GPT-5.5 benchmarks, trained with 10x the compute of Muse Spark. A coding-focused Muse Spark update is imminent.

  • Sam Altman has been framing GPT-5.6 Sol’s government-supervised release as the template for responsible scaling: “build, test, share, iterate.” Sol is reportedly 54% more token-efficient on agentic coding tasks.

  • Anthropic confirmed early-stage discussions with Microsoft to run Claude inference on Microsoft’s custom Maia 200 chips (3nm, inference-optimized) via Azure — a potential strategic shift away from exclusive AWS reliance.

  • Marc Andreessen appointed co-chair of the Fed’s new “Productivity and Jobs” AI task force (July 9). Also declared “AI is already a better doctor than 99.99% of human doctors” on July 12, reigniting the healthcare AI debate. Critics flagged the conflict: a16z manages $90B with $3.4B in AI investments. (WaPo) (Benzinga)

  • Jim Fan declared “Rest in Peace, VLAs” at Sequoia AI Ascent — arguing Vision-Language-Action models are fundamentally brittle and the industry should pivot to video-first World Action Models (WAMs). NVIDIA also open-sourced ASPIRE, a robot skill library where success rates on dual-arm handover jumped from 20% to 92%. (zeus.rocks)

  • François Chollet launched ARC-AGI-3, the first interactive reasoning benchmark evaluating how AI systems explore unknown environments, model them, set goals, and plan autonomously. A new $2M Kaggle competition accompanies it. His startup Ndea (YC W2026, with Zapier co-founder Mike Knoop) raised $43M. (arcprize.org)

  • Superintelligent newsletter (Kim) ran “The Commoditization of Intelligence” on July 4 — Chinese open models now undercut Western flagships by up to 100x while scoring within a few points of them. (getsuperintel.com)

📊 Benchmark Watch

Fable 5 continues to lead every major Arena leaderboard (Text, Code, Agent) with 7.15M+ votes across 369 models. Sonnet 5 (thinking) was added to Code, Text, Search, Vision, and Document leaderboards this week. GPT-5.6 Sol’s full benchmark numbers are still being evaluated — early reports suggest it’s competitive with Fable 5 on coding/reasoning at 2x cheaper pricing. Grok 4.5 entered the rankings claiming “Opus-class” at Sonnet-class prices. (Arena.ai) (Swfte)

🎙️ Podcast Highlights

  • TBPN acquired by OpenAI. Scott Galloway interviewed John Coogan and Jordi Hays (July 9) about turning their daily tech show into one of the fastest-growing media businesses — and OpenAI’s first acquisition. (Prof G Media)

  • Lex Fridman #490: “State of AI in 2026” with Nathan Lambert (Allen AI) and Sebastian Raschka. Covered LLMs, coding agents, scaling laws, China competition, GPUs, and AGI timelines. (lexfridman.com)

  • All-In E279 (July 3): “AI Sovereignty Wars” — covered OpenAI vs Anthropic IPOs, Meta’s new model, Zuck’s price war, and CCP considering export controls on Chinese models. (allin.com)

  • Reid Hoffman (July 6) discussed OpenAI and Anthropic IPOs, arguing the rivalry narrative is overblown and both companies have room to win. (Yahoo)

🏛️ Regulation Watch

  • Illinois signed the AI Safety Measures Act (July 6) — the most comprehensive state-level AI law yet, requiring developers to publish catastrophic risk frameworks and submit to independent third-party safety audits. Modeled after California and New York bills. (WTTW)

  • FTC issued a policy statement on AI accuracy, seeking public comment (due July 31) on laws requiring alteration of “truthful outputs of AI models.” (FTC)

  • EU AI omnibus regulation will ban AI systems generating non-consensual intimate images of real people, effective December 2026.

🔗 Worth Reading

  • “Old and new apps, via modern coding agents” — Terence Tao on using AI agents to port decades-old math visualizations and why the risk profile is acceptable for supplementary work. (terrytao.wordpress.com)

  • “JADEPUFFER: Agentic Ransomware for Automated Database Extortion” — Sysdig’s full technical writeup on the first documented autonomous AI ransomware operation. (Sysdig)

  • “Sam Altman Seeks New World Order for AI” — Fortune on how OpenAI is slowly losing ground to Google and Anthropic while Altman pushes for a government-industry framework. (Fortune)