🗞️ AI Daily Briefing — 2026-04-17

🔥 Top Story

Anthropic ships Claude Opus 4.7 — but keeps Mythos locked away. Anthropic released Claude Opus 4.7 yesterday, delivering a 13% lift on coding benchmarks, 3x more production tasks resolved, high-resolution vision support (up to 3.75 megapixels), and a new tokenizer. It’s the best publicly available Claude model yet — but Anthropic openly concedes it trails their unreleased Mythos Preview, which scores 93.9% on SWE-bench Verified and remains behind a 50-company firewall. Pricing stays at $5/$25 per million tokens. The messaging is unusual: shipping a strong model while admitting you have a better one you won’t release. (Anthropic) (CNBC) (Axios) (AI Business)

🚀 Model & Research News

  • Perplexity launches Personal Computer for Mac: A full desktop AI agent for Max subscribers ($200/mo) that integrates with iMessage, Mail, Calendar, and local files. Hit both Command keys to summon it. Can run 24/7 on a Mac mini. Actions are sandboxed, auditable, and reversible. The most aggressive consumer AI agent play yet. (MacRumors) (9to5Mac)
  • Canva AI 2.0 goes agentic: Canva unveiled a complete platform overhaul powered by proprietary models — Proteus (style transfer), Lucid Origin (image gen), and I2V (image-to-video) — claiming up to 7x faster and 30x cheaper than frontier alternatives. The platform is now a conversational agent that can design, automate, and publish from natural language prompts. Research preview rolling out to the first 1M users. (TechCrunch) (Creative Bloq)
  • Google ships AI Mode in Chrome with split-screen browsing: AI Mode now opens webpages side-by-side with the AI assistant. You can add browser tabs (PDFs, lecture notes, product reviews) as context for queries. Rolling out on desktop, Android, and iOS in the US. The browser-as-AI-workspace vision is getting real. (Google Blog) (Droid Life)
  • Antioch raises $8.5M to be “the Cursor for physical AI”: The simulation startup lets robot builders spin up digital twins of their hardware with simulated sensors, enabling edge-case testing and reinforcement learning before real-world deployment. Seed round at $60M valuation, led by A* and Category Ventures. (TechCrunch) (SiliconANGLE)

🛠️ Tools & Developer Updates

  • dbt Labs State of Analytics Engineering 2026: 72% of data teams now use AI-assisted coding, but only 24% have AI-assisted pipeline management. The gap between AI speed and data governance is widening. Also announced: a dbt VS Code extension powered by dbt’s MCP server integrated with Google’s Antigravity IDE. (dbt Labs) (BigDATAwire)
  • NVIDIA National Robotics Week updates: New Isaac GR00T open models let robots understand natural language instructions and perform complex multistep tasks. New Cosmos world models help systems learn more efficiently and generalize across environments. (NVIDIA Blog)
  • Google Chrome AI Skills: Users can now save and reuse favorite AI workflows as “Skills” across the web, announced April 14. (TechCrunch)

💰 Funding & Business

  • Aehr Test Systems gets $41M record order: Lead hyperscale AI customer placed a follow-on production order for package-level burn-in of custom AI processor ASICs. Second-half bookings now exceed $92M. Signals continued massive AI chip buildout. (Aehr)
  • Stellantis × Microsoft AI partnership: Stellantis is accelerating its AI-led strategy with Microsoft to enhance customer experiences and digital transformation across its vehicle portfolio. (Microsoft)
  • 80% of enterprise workers rejecting AI tools: Fortune reports that roughly 8 in 10 enterprise workers are either avoiding or actively resisting the AI tools their employers are spending record sums deploying. A growing tension between top-down AI mandates and ground-level adoption. (Fortune)

🐦 Notable from the Timeline

  • The New Yorker investigation into Sam Altman continues to dominate conversation — 100+ interviews, 70 pages of secret Ilya Sutskever memos alleging “a consistent pattern of lying.” OpenAI responded by announcing a Safety Fellowship programme. (Semafor)
  • OpenAI’s TBPN acquisition still generating debate — Slate called it “sleazy,” CNN framed it as “buying influence,” while Ben Thompson at Stratechery offered a more nuanced take on the token tsunami. The deal was reportedly in the low hundreds of millions. (Stratechery)
  • Gartner: Organizations with successful AI initiatives invest up to 4x more in data and analytics foundations than those struggling. Reminder that AI success is a data infrastructure problem first. (Gartner)

📊 Benchmark Watch

LMArena standings (as of April 14): Claude Opus 4.6 Thinking still #1 overall at 1504 Elo. Opus 4.6 dominates coding at 1549 Elo. Grok 4.20 Beta1 has climbed to #4 globally (1491 Elo), surpassing GPT-5.4. The Intelligence Index ceiling of 57.18 has held since February — neither Gemini 3.1 Pro nor GPT-5.4 has broken through. 5.8M+ votes across 339 models. Expect Opus 4.7 to shake things up in the coming days. (LMArena) (LLM Stats)

🎙️ Podcast Highlights

  • TBPN is now under OpenAI ownership as of April 2. Co-hosts John Coogan and Jordi Hays continue daily broadcasts with “editorial independence.” The acquisition itself became the story — Hard Fork devoted a segment to whether corporate-owned tech media can stay credible. (NPR)

🔗 Worth Reading