🗞️ AI Daily Briefing — 2026-04-16

🔥 Top Story

Anthropic preparing Claude Opus 4.7 and an AI design tool — possibly this week. Per The Information, Anthropic is readying its next flagship model alongside a design tool that lets users build websites, presentations, and landing pages from natural language prompts. Anthropic has partnered with Figma on a “Code to Canvas” feature that converts Claude Code output into editable Figma designs. Adobe, Wix, and Figma shares each dropped >2% on the news. If the model ships this week, it would extend Anthropic’s lead on LMArena while opening an entirely new product surface against Canva, Adobe, and Wix. (Dataconomy) (TechBriefly) (CNBC / Figma)

🚀 Model & Research News

  • Novo Nordisk × OpenAI strategic partnership: Announced April 14 — OpenAI will help Novo Nordisk accelerate drug discovery, optimize manufacturing and supply chain, and upskill its global workforce on AI. Pilot programs launch immediately with full integration targeted by end of 2026. The biggest pharma-AI deal yet by scope. (CNBC) (Euronews)
  • Meta Muse Spark climbing the ranks: Meta’s first model under Alexandr Wang (code-named Avocado) now scores 4th on the Artificial Analysis Intelligence Index, behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Notably, it’s closed-source — a full break from Meta’s open-weight Llama tradition. Rolling out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta glasses. (Axios) (TechCrunch)
  • MiniMax M2.7 — open-source self-evolving agent model: Features a three-component system (short-term memory, self-feedback, self-optimization) and hits 87.4% GPQA with a 41.9 coding index. The best open price/performance story on the board right now. (buildfastwithai)
  • AI Scientist-v2 gets peer-reviewed paper accepted: Sakana AI’s agentic tree-search system autonomously wrote a paper that cleared ICLR workshop peer review — the first fully AI-generated paper to do so. Uses a VLM feedback loop for figure refinement. Published in Nature. (arXiv:2504.08066)

🛠️ Tools & Developer Updates

  • Anthropic × Figma “Code to Canvas”: The new feature converts AI-generated code (from Claude Code and similar tools) into fully editable Figma designs. This closes the loop between AI coding agents and design teams — build a working interface with prompts, then hand it to design for refinement. (CNBC)
  • Parasail raises $32M for “AI Supercloud”: Deploys and scales AI agents in minutes across distributed infrastructure. Co-led by Touring Capital and Kindred Ventures, backed by Samsung NEXT. One to watch if you’re building agent infra. (Tech Startups)
  • dbt goes AI-ready at Google Cloud Next: dbt Catalog now visualizes what agents and LLMs are building, the Semantic Layer gives AI access to well-defined metrics, and Advanced CI lets you preview every data change before it ships — even when AI is making the changes. (dbt Labs blog)

💰 Funding & Business

  • Q1 2026 venture funding hit $300B — an all-time record. AI captured $242B (80% of total). Four of the five largest venture rounds ever closed in Q1: OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B). These four alone absorbed 65% of all global venture investment. (Crunchbase)
  • Anthropic valued at up to $380B after its $30B Series G in February. VCs reportedly offering up to $800B in the next round, driven by Opus 4.7 and Mythos hype. The company now has 1,000+ enterprise customers spending $1M+/yr. (The Decoder)
  • Gizmo — $22M Series A: AI-powered studying and personalized learning loops. (Tech Startups)
  • U.S. data center buildout hitting walls: More than half of planned builds delayed or canceled due to supply shortages, power constraints, and Chinese import reliance. ASML raised its 2026 outlook as customers scramble to secure capacity. (Tech Startups)

🐦 Notable from the Timeline

  • @sama: The New Yorker investigation fallout continues to ripple. Ilya Sutskever’s internal memos alleging a “pattern of lying” are being re-read through the lens of Mythos vs GPT-5.4-Cyber — two companies, two bets on how dangerous AI should reach the market, and questions about who’s trustworthy enough to make that call. (Semafor)
  • @fchollet: ARC-AGI-3 leaderboard still shows frontier models in low single digits vs humans near 100%. If you needed a cold shower after the “AGI is here” hype cycle, this is it. (ARC Prize)
  • @pmarca: Continuing to push federal preemption as state AI bills pile up — Maine, Maryland, Nebraska all introduced new legislation this week. The anti-AI violence against executives is only strengthening the “let Washington handle it” argument.
  • EU AI Act countdown: Starting August 2, 2026, every AI system used in hiring, task allocation, and performance monitoring becomes “high-risk” under EU law. Mandatory risk assessments, bias testing, human oversight, and continuous monitoring. Fines up to €15M or 3% of global turnover. U.S. employers are covered if AI outputs touch EU workers. (Asanify)

📊 Benchmark Watch

  • LMArena Text: Claude Opus 4.6 Thinking holds #1 (~1504 Elo). Gemini 3.1 Pro ~1493, Grok-4.20 Beta1 ~1491. The top 5 are within ~20 Elo of each other — effectively a five-way tie. (Arena.ai)
  • Coding: Claude Opus 4.6 leads SWE-bench Verified at 80.8%. MiniMax-M2.5 at 80.2% remains the best cost-performance story.
  • Muse Spark: 4th on Artificial Analysis Intelligence Index (score 52), behind Gemini 3.1 Pro, GPT-5.4, and Opus 4.6. Impressive for Meta’s first closed-source play.
  • The unreleased ceiling: Mythos Preview’s leaked 93.9% SWE-bench / 94.6% GPQA scores remain the high-water mark — just not shippable.

🎙️ Podcast Highlights

  • TBPN (now OpenAI-owned): Coogan & Hays continuing daily coverage — this week focused on GPT-5.4-Cyber rollout and the Altman attacks. The editorial-independence question remains the live wire post-acquisition. OpenAI paid “low hundreds of millions” per the FT. (Slate) (CNBC)
  • All-In: Expect heavy coverage this weekend of Mythos vs GPT-5.4-Cyber and the state AI law pileup. Sacks and Friedberg have been pre-loaded on both threads all week.

🔗 Worth Reading