Good morning, AI enthusiasts. The Trump administration doesn't just have opinions about frontier AI anymore. It now controls who gets access to it, one customer at a time.
OpenAI's GPT 5.6 won't have a public launch. The government is approving customers one by one during a preview period, and two White House offices worked closely with OpenAI's team to set it up. The question isn't whether this is a safety measure or something bigger. It's whether OpenAI is the last company this happens to.
In today's recap:
White House gates OpenAI's GPT 5.6 customer rollout
Claude is beating ChatGPT where it counts: paid subscribers
Automate daily AI briefings with Claude and n8n
$320M bet: video game footage can train real AI agents
4 new AI tools and more
POLICY
White House gates OpenAI's new model release
Source: Getty Images
Recaply: OpenAI just staggered the release of GPT 5.6 at the Trump administration's request, sharing access with a select group of partners only as the government approves customers individually during a preview period ahead of any broader rollout.
Key details:
Instead of a public launch, OpenAI plans to distribute GPT 5.6 only to close partners, with the government approving each customer individually. CEO Sam Altman told staff the administration would be approving access "customer by customer" during the preview.
GPT 5.6 is in the same category as Anthropic's Claude Mythos, a frontier cyber model Anthropic voluntarily restricted to a small set of partners over its ability to identify and exploit software vulnerabilities at speeds no human analyst could match.
The two agencies behind the request were the Office of the National Cyber Director and the Office of Science and Technology Policy; OpenAI's staff reportedly "worked closely" with both agencies on the rollout plan, according to The Information.
Altman told staff a broader public release could follow the preview "a couple of weeks later" if the limited rollout goes smoothly, according to The Information.
Why it matters: For the first two years of this administration, AI policy was mostly commentary. This is different. A sitting government is now directly sequencing which companies get access to a specific frontier model, approving them one at a time. That's not safety review in the abstract. It's access control over commercial AI infrastructure with the executive branch as the gate. Whatever the stated rationale, the precedent is now set. OpenAI probably won't be the last frontier lab to navigate it.
ANTHROPIC
Claude wins where it counts: paying subscribers
Source: Getty Images
Recaply: Anthropic just pulled ahead of ChatGPT among paying consumers, with credit card transaction data showing Claude's paying base growing 75% since January 2026, even as ChatGPT still commands the total market by a wide margin.
Key details:
Indagari analyzed billions of anonymized credit card transactions from 28 million U.S. consumers to track paid subscription trends, measuring actual spending on Claude and ChatGPT rather than app downloads or self-reported users.
Claude's paying consumers grew 75% since January 2026 (through May 10), while DataCamp reports self-directed demand for Claude courses outpaces ChatGPT by 3:1, with demand up 18x in the last 30 days, on a platform with about 20 million users.
The growth continued even after Anthropic's March 2026 refusal to work with the Trump administration on mass surveillance and autonomous weapons, a decision that may have added brand trust among privacy-conscious subscribers, according to TechCrunch.
"Claude" is now the most searched term on DataCamp, outranking even the term "AI," a signal that consumer intent is shifting from the category to the brand.
Why it matters: There has been lots of talk about ChatGPT's unassailable consumer lead. Anthropic doesn't agree, and it has the credit card data to back it up. Paid users are the sharpest signal available. They chose to spend money, evaluated alternatives, and picked a winner. The DataCamp figure is the clearest data point. When self-directed learners search for AI education and choose Claude 3:1 over ChatGPT, that's a deliberate preference, not a default.
AI RESEARCH
Gameplay footage is now serious AI training data
Source: General Intuition
Recaply: General Intuition just secured $320M at a $2.3B valuation to scale AI agents trained on hundreds of millions of hours of video game footage, betting that action-labeled gameplay data gives agents the spatial reasoning that internet text was never able to provide.
Key details:
General Intuition trains a single model on gameplay clips from Medal, the gaming clip platform co-founder Pim de Witte also runs, using action labels recording exactly which buttons players pressed and when, not just what the screen showed. The CEO says most competitors can't replicate this from video alone.
The $320M round brings total disclosed funding to $454M, up from a $134M seed in October 2025, led by Khosla Ventures with participation from Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers at Google DeepMind and MIT.
The same model powering a Fortnite-style game agent also drove a quadruped robot in the company's New York office, adapted to the indoor environment using just 8 minutes of real-world robotics data collected on the street outside.
General Intuition plans to make its API broadly available by end of summer 2026, with the vast majority of the $320M earmarked for compute scaling on a CoreWeave infrastructure deal.
Why it matters: Every frontier lab is hunting the same scarce resource: high-quality, action-dense training data that internet text can't supply. General Intuition's answer is that hundreds of millions of hours of gameplay already contain exactly what's missing, labeled sequences of a human making decisions in a physical environment, frame by frame. If the thesis holds, Medal's archive isn't just a gaming clip platform. It's the world's largest unintentional training dataset for embodied AI. The question investors are now funding is whether a game engine is a close enough proxy for the physical world at scale.
NEWS
Beyond the big stories in AI today
Adobe acquired Topaz Labs, an AI image and video enhancement startup that won a 2025 Emmy for its production technology, with plans to integrate Topaz's Astra upscaling and Wonder retouching models into Firefly and Creative Cloud apps.
Apple raised prices on MacBooks and iPads to pass surging memory and storage costs on to consumers, sending shares down more than 6% on Thursday, the company's worst single-day drop since April 2025.
Google launched a new Google Finance Android app and portfolio tracking feature, including an AI briefing tool that generates custom market updates on a user-defined schedule, from pre-market crypto rundowns to watchlist-specific alerts.
Airwallex raised $320M at an $11B valuation, up 38% from December 2025, and launched T:0, an AI-native autonomous finance automation platform, and Airi, an agentic consumer wallet with delegated payment controls and multi-currency balances.
ON Semiconductor agreed to acquire Synaptics in an all-stock deal valued at approximately $7B, adding edge AI compute and connectivity technology to push from AI data centers into physical AI for automotive, robotics, and AR/VR.
Qualcomm announced the acquisition of AI software startup Modular for nearly $4B, adding a chip-agnostic inference platform that lets developers run AI models across Nvidia, AMD, and other chips without rewriting separate code for each processor.
Micron locked in historically high memory pricing through 2030 via 16 strategic customer agreements, reporting Q3 revenue of $41.5B (a 346% year-over-year jump) and setting Q4 guidance at $50B as AI demand keeps supply structurally constrained.
Anthropic accused Alibaba of running the largest distillation attack on its models to date, alleging operators linked to Alibaba used 25,000 fake accounts to conduct 28.8 million interactions with Claude between April 22 and June 5.
GUIDES
Automate daily AI briefings with Claude and n8n
Recaply: In this tutorial, you will learn how to build an automated daily AI news briefing using n8n and Claude, eliminating the manual effort of scanning multiple sources each morning.
Step-by-step:
Open n8n and create a new workflow, add a Schedule trigger node set to fire at 7:00 AM daily, then connect two HTTP Request nodes pulling from TechCrunch AI's RSS feed (https://techcrunch.com/category/artificial-intelligence/feed/) and the Hacker News top stories API (https://hacker-news.firebaseio.com/v0/topstories.json).
Add a Code node after each HTTP Request to extract only the title, URL, and description from each item, keeping the top 5 from each source, then merge both lists into a single JSON array using a Merge node.
Connect the Merge node to an AI Agent node, select Claude as the model, and use this prompt, followed by the merged array: "Summarize each story in 2-3 sentences. Identify the one story most worth reading today and explain why in one sentence. Stories:"
Add a Slack or email node after the AI Agent node, map the output to the message body, then run the workflow manually once to confirm the briefing arrives with clean formatting and no raw JSON artifacts.
Toggle the workflow to Active, then add a filter node before the AI Agent step that drops any story whose title contains "Sponsored," "Advertisement," or "Partner" to keep the briefing signal-clean.
Pro tip: Add a second Claude call at the end that compares today's sources against yesterday's saved output (store it in a JSON file node) and flags any topic appearing on both days as "trending." You get a low-cost momentum signal built from your own briefing data.
TOOLS
Trending AI Tools
🤖 Claude Tag - Anthropic's team AI for Slack, tag @Claude in any channel to delegate tasks asynchronously and let the whole team collaborate with one shared AI teammate
🎬 OpenArt Director - Vibe-directing tool for building short films, music videos, UGC ads, and trailers from chat prompts in a single interface
🎨 Krea 2 - Krea's open-weights in-house image generation model, part of its AI creative suite covering text-to-image, video, and 3D generation
📝 Mistral OCR 4 - Mistral's document intelligence model with bounding boxes, block classification, confidence scores, and support for 170 languages, deployable in a single container
