Hello, AI Explorer!
Good Wednesday. This week: Claude Sonnet 5 is now the default -- free for everyone & Only 18% of companies generate value from AI - story
Welcome back to your weekly AI digest: human-curated, zero fluff.
Every Wednesday and Saturday, we bring you:
- The week's most important AI developments (no fluff)
- Hand-picked AI tools actually worth your time
- Expert insights on what these changes mean for you
Let's dive into this week's discoveries! ⚡
🔥 This Week in AI
🔥 This Week in AI |
July 1, 2026 |
| 🤯 | Only 18% of businesses generate real value from AI -- yet 86% are increasing budgets anyway |
| 🧠 | Claude Sonnet 5 is now the default model for free and Pro users -- near Opus performance at Sonnet prices |
| 🇨🇳 | China's Meituan trained a 1.6 trillion-parameter AI model entirely on domestic chips and open-sourced it |
| ⚡ | Etched exits stealth with $800M raised and $1B in contracts for a chip that only runs AI |
| 🧬 | Meta's Brain2Qwerty reads brain activity as text -- 61% accuracy with no surgery required |
| 🍎 | Apple is pushing security patches months early because AI is shrinking the window before exploits get built |
| 📔 | NotebookLM now turns any document into a 60-second vertical video -- rolling out to Google AI subscribers |
| 🔬 | OpenAI's biology benchmark GeneBench-Pro stumps top AI -- best model solves only 31.5% of problems |
ACCENTURE RESEARCH · AI ADOPTION
| 🤯 |
You're More Ahead Than You Think -- The AI Adoption Gap in One Chart |
A visualization is making the rounds that puts AI adoption into sharp perspective. Picture a grid of dots -- each one representing 3.2 million people. The vast majority are gray. Those are people who have never interacted with an AI tool at all. A smaller cluster sits in green -- people using a free chatbot. And then, at the very bottom of the chart, almost invisible against the rest, a tiny group of yellow and red dots. Those are us. People paying for AI tools, running real workflows, and building with coding assistants.
Accenture put numbers to this in a global survey of 1,320 C-level executives: only 18% of businesses currently generate measurable value from AI. Yet 86% plan to increase their AI budgets in 2026. The spending is ahead of the capability. Most businesses are trying to buy their way into results before they have the workflows to back it up.
What the numbers actually mean:
• The gap between AI investment intent and actual value creation is still massive -- and growing
• Using a free chatbot already puts you ahead of most of the world's population
• For builders, consultants, and early adopters -- this gap is an opportunity window that is still very much open
Why it matters:
Inside the AI bubble, everything looks saturated. Outside it, most businesses have no plan and no tools. If you can help close that gap -- by knowing what works and why -- you're not entering a crowded market. You're entering the early part of a long one.
ANTHROPIC · MODEL RELEASE
| 🧠 |
Anthropic Ships Claude Sonnet 5 as the New Default -- Near Opus Performance, Lower Price |
Anthropic released Claude Sonnet 5 on June 30 and made it the default model for both free and Pro accounts the same day. The model can plan, run browsers and terminals autonomously, and tackle the kind of agentic tasks that, until recently, required Opus. Pricing through August 31 sits at $2 per million input tokens and $10 per million output tokens.
That undercuts both GPT-5.5 and Gemini 3.1 Pro at launch. After August 31, pricing moves to $3/$15. On safety, Sonnet 5 refuses more misuse attempts and resists prompt injection attacks better than Sonnet 4.6, which it replaces as the default.
Sonnet 5 at a glance:
• $2 input / $10 output per million tokens through August 31, then $3/$15 -- cheaper than GPT-5.5 and Gemini 3.1 Pro
• Default now for free and Pro accounts -- no plan change required to access it
• Near Claude Opus 4.8 performance on agentic reasoning, tool use, and coding tasks
Why it matters:
The performance gap between Sonnet and Opus just closed significantly. For anyone running AI agents or building on the Claude API, the cost equation for production workloads changed overnight. You're getting roughly what you used to pay Opus prices for -- at Sonnet prices, for at least two more months.
| 🇨🇳 |
China Trains the World's Largest AI Model on Domestic Chips -- and Open-Sources It |
Meituan released LongCat-2.0 on June 30 -- a 1.6 trillion-parameter model the company says is the first of its size trained end-to-end on a 50,000-chip domestic compute cluster, a direct response to US export controls on advanced semiconductors. The model carries a 1 million-token context window, scored 59.5 on SWE-bench Pro (GPT-5.5 sits at 58.6), and Meituan open-sourced the weights so the community can run the benchmarks themselves. The training-on-domestic-chips claim relies on Meituan's own account, but if accurate, it's the most concrete evidence yet that China can build frontier models without US hardware.
WHY IT MATTERS:
DeepSeek used domestic chips only for inference. LongCat-2.0 claims full-stack training on Chinese hardware. That's the difference between running models and building them -- and it moves the chip export control debate into new territory.
| ⚡ |
Etched Exits Stealth with $800M Raised and a Chip Built Only to Run AI |
AI inference chip startup Etched came out of stealth on June 30 with a working chip, $800 million raised across multiple rounds, and more than $1 billion in signed customer contracts. The chip does exactly one thing: runs transformer-based AI models. No GPU flexibility, no general-purpose compute -- just AI inference, fully optimized. At a $5 billion post-money valuation, the company has backing from Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, and several major quant trading firms. Rack-scale inference systems start shipping this summer.
WHY IT MATTERS:
Nvidia's strength is flexibility -- GPUs that happen to be excellent at AI. Purpose-built inference chips like Etched's could deliver far better throughput per watt for the part of the AI stack that actually runs in production and costs the most money. $1B in contracts before shipping is the kind of signal that makes the market pay attention.
⚡ AI Tools of the Week
🎨 Adobe Firefly
Generative Fill now lives inside every copy of Photoshop. Select any area, type what you want there instead, and Firefly generates a match in about 3 seconds. Every output is commercially licensed, so agencies can use it on client work without a second thought.
Try Adobe Firefly →📱 Dr.Fone
Dr.Fone moves WhatsApp chats and 17 other file types between iPhone and Android in minutes, no cable juggling or cloud backup waiting required. Anyone switching phones this year will hit this exact problem at least once.
Try Dr.Fone →🔬 Consensus
Consensus searches 200 million research papers and tells you straight up whether the evidence backs a claim or contradicts it. It's the fastest way to check if "studies show" is actually true before you repeat it.
Try Consensus Free →⚖️ Harvey AI
Harvey runs legal research, contract review, and case strategy against millions of legal documents, then cites the specific precedents it relied on so lawyers can actually verify the work. Firms report saving 10+ hours per case.
Discover Harvey AI →⚡ Quick Hits
🧬 Meta's Brain2Qwerty v2 reads brain activity as text -- 61% word accuracy, no surgery
Using a helmet-mounted MEG scanner and an AI model trained on 22,000 sentences from nine volunteers, the system hits 61% average word accuracy -- up from 8% with previous non-invasive methods. Designed to help people with brain lesions communicate again. Code and dataset are open. Read the Meta blog post →
🍎 Apple is releasing iOS security patches months ahead of schedule because of AI
Apple pushed iOS 26.5.2 early -- fixes that would have waited for the next major cycle. The reason: AI tools now help attackers turn a known vulnerability into a working exploit faster than the old patch timeline allowed. Apple confirmed no active exploitation, but changed the schedule anyway. Read the full story →
📔 NotebookLM now turns any document into a 60-second vertical explainer video
Short Video Overviews launched June 30 for Google AI Ultra and Pro subscribers. Upload any source -- document, note, research paper -- and NotebookLM generates a 60-second narrated vertical video on a single concept, powered by Nano Banana 2 Lite. Free-plan rollout is coming. See the feature →
🔬 OpenAI's biology benchmark GeneBench-Pro stumps every top AI model
129 problems in genomics, quantitative biology, and translational medicine -- designed for real research judgment, not fixed recipes. GPT-5.6 Sol Pro leads at 31.5%. Claude Opus 4.8 scores 16%. Each problem takes a human expert 20-40 hours. The benchmark is open. Read the OpenAI release →
🎬 HeyGen published the full AI video prompt playbook for that viral conference intro
HeyGen reverse-engineered how a jaw-dropping conference opener was made with AI and published the complete recipe -- prompt by prompt. If you create AI video content, this is the kind of breakdown worth saving. Read the playbook →
NVIDIA H200 GPU: Full Specs, Price, and Cloud Rates for 2026
AWS raised H200 rental prices 15% in January, even after NVIDIA's next-gen B200 started shipping, because high-memory Hopper supply still can't keep up.
|
141GB HBM3e memory, nearly double the H100's 80GB |
$28K-$45K Per-GPU price range, PCIe and SXM5 variants |
That's all for this edition! What tools are you most excited to try?
Reply to this email - we read every response and use your feedback to improve future editions.
See you on Saturday,
The AI Tool Discovery Team
Feedback
What'd you think of today's newsletter?
Vote below to let us know how we're doing.