OpenAI has begun a limited preview of GPT-5.6, its most capable model family yet, in three tiers: Sol (the flagship, with notable gains in coding, biology and cybersecurity), Terra (a balanced everyday model priced about half of the previous generation), and Luna (a fast, low-cost option). The twist is the rollout: instead of a public launch, the models went to roughly 20 organizations whose access was approved by the U.S. government. This follows a June 2 executive order, “Promoting Advanced Artificial Intelligence Innovation and Security,” under which leading AI developers voluntarily give the government up to 30 days of early access to their most powerful models before public release. It is the first time the U.S. has preemptively restricted the launch of an American AI model — officials reportedly judged GPT-5.6 to have “Mythos-like” capability (a reference to Anthropic’s own restricted frontier model). Pricing per million tokens: Sol $5 in / $30 out, Terra $2.50 / $15, Luna $1 / $6. General availability is expected “in the coming weeks.”
Ahead of the GPT-5.6 preview, the Trump administration asked OpenAI to hold the release to a small set of government-vetted partners, citing security concerns. The request — reported by Axios and CNN — is the practical debut of the new executive order’s pre-release review process, and it sets a precedent: a U.S. company restricting its own product launch at the government’s request. For communicators, this is the story to watch, because it reframes “model launches” as events that now sit partly inside a national-security review, much as export controls already do. The same week, the U.S. tightened export controls on Anthropic’s Fable 5 and Mythos 5 models, underscoring that the gating now spans multiple labs.
HP Inc. announced it is scaling up its strategic “Frontier” partnership with OpenAI, deploying the company’s AI across customer experiences, software development and internal enterprise operations. It is a useful marker of how the largest hardware makers are now standardizing on a single frontier lab as an operating layer, not just a feature — and a reminder that enterprise AI adoption stories are increasingly framed as multi-year partnerships rather than one-off product news.
The flip side of Washington’s gatekeeping is a growing argument that it hands China an opening. Rather than retaliate directly, Chinese labs are doubling down on open-weight models that anyone can download and self-host — outside the reach of U.S. export controls. Industry estimates put Chinese open-weight models at roughly 61% of tokens flowing through the OpenRouter marketplace by May 2026, and Alibaba’s Qwen family passed one billion cumulative downloads on Hugging Face, overtaking Meta’s Llama as the most-downloaded open family. The strategic question for the months ahead: does restricting closed U.S. models simply accelerate global adoption of open Chinese ones?
Perplexity’s Comet browser adds a “Personal Computer” mode on MacComet’s agent can now edit local files, browse locally and take voice commands, and is now powered by Claude Sonnet 4.6 (Opus 4.6 for Max users), with an iOS pre-order open. See also · Workflows.
Tech journalist and investor Om Malik has diedAzeem Azhar opens this week’s Exponential View with a tribute to Malik — GigaOm founder, journalist and photographer — calling him “one of tech’s truest voices.” See also · Substack.
Exponential View · Jun 28
Workforce & Society
08
“Office workers are spending way too much on AI, too”Pivot to AI flags how individual employees are quietly expensing AI subscriptions and tokens — a budgeting and governance blind spot for ops and IT teams. See also · Substack.
Pivot to AI · Jun 28
09
Fifty years of Moore’s Law “wasn’t fast enough for AI”Azeem Azhar argues that AI’s compute appetite has outrun even the historic doubling of chip performance — the gap now filled by data centers, custom silicon and (literally) plans for orbital compute.
The June 2 executive order behind this week’s model gating“Promoting Advanced AI Innovation and Security” establishes a voluntary process for frontier developers to give the government up to 30 days of pre-release model access. See also · Top Stories.
White House · Jun 2
12
Anthropic’s distillation complaint frames the China-security backdropAnthropic’s June letter to the Senate Banking Committee — alleging operators tied to Alibaba’s Qwen lab ran ~25,000 fraudulent accounts and 28.8M Claude exchanges — is part of the security argument now shaping release policy.
Axios · Jun
From Your Inbox
Substack Highlights.
A dedicated Substack reader connection wasn’t available for this run, so these highlights are drawn from the Substack/newsletter feeds in your Inoreader “AI” folder. Where a newsletter delivered only a preview, the summary reflects the published teaser plus the underlying stories it points to.
Leads with the GPT-5.6 trio and the “kiiinda” caveat — the models exist, but most people can’t touch them yet because of the government-approved limited preview.
Also flags Anthropic’s restricted Mythos model, an AI startup called General Intuition, Google AI Studio’s new app-building, and a push toward better AI benchmarks.
Framing for non-technical readers: the news is less “a new model” and more “a new gatekeeping regime around who gets the most powerful models first.”
Opens with a personal tribute to the late Om Malik, the GigaOm founder and tech journalist who died last week.
Core argument: AI’s demand for compute has outpaced even Moore’s Law — the half-century trend of steadily doubling chip performance — which is why the industry is now chasing custom silicon, vast data centers and even space-based computing.
A useful, big-picture lens for explaining to non-technical stakeholders why “the chip shortage” and energy/compute costs keep surfacing in AI coverage.
A practitioner’s shortlist cutting through tool overload — the kind of curated, “what survives daily use” rundown that’s directly useful for your own stack decisions.
Worth a skim for any tools that overlap with your toolkit (assistants, note and automation apps) before adding anything new.
Argues that individual employees, not just companies, are over-spending on AI subscriptions and usage — often invisibly, on personal cards or expense accounts.
A governance prompt for any comms or ops leader: who is paying for which AI tools across the team, and is that spend tracked?
Another roundup centered on the same defining story — OpenAI’s new models and the government’s pre-release restriction — confirming how thoroughly it dominated the AI conversation this week.
See also · Top Stories
From Your AI Feeds
Inoreader AI Folder.
Eight articles landed in your “AI” folder in the last 24 hours. The set is dominated by one story — OpenAI’s GPT-5.6 launch and Washington’s role in it — with a few useful outliers on tools, spending and AI in the humanities.
HP is scaling its OpenAI “Frontier” partnership to embed AI across customer experience, software development and enterprise operations — a marker of how big hardware vendors are standardizing on a frontier lab.
The Neuron’s daily wrap leads on the GPT-5.6 Sol/Terra/Luna preview and its government-gated rollout, plus quick hits on Mythos, General Intuition, Google AI Studio and AI benchmarks.
A weekly synthesis of the frontier-model race: OpenAI’s launch, Anthropic’s expanding Claude family, and China’s open-source response to U.S. gatekeeping.
A second roundup on the same headline story, with side notes on Mythos, General Intuition and Google AI Studio — useful for cross-checking how the week is being framed.
A curated, practitioner’s shortlist of AI tools that have earned a permanent place in one user’s daily workflow — a quick gut-check against tool sprawl.
A detailed walk-through of using Google’s Gemini to identify and contextualize a page from a 1698 Amsterdam dictionary — a concrete example of AI applied to archives, libraries and historical research.
A Portuguese translation (by Rodrigo Ghedin for Manual do Usuário) of a Pivot to AI piece arguing that individual employees are quietly over-spending on AI tools — a governance and budgeting blind spot.
Azeem Azhar’s flagship essay on why AI’s compute demands have outstripped Moore’s Law, opening with a personal remembrance of the late Om Malik.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
Anthropic makes enterprise MCP connections “zero-touch” — and tags Claude in Slack
Anthropic moved its Enterprise-Managed Authorization for MCP (the standard that lets Claude securely connect to outside tools and data) to stable. In plain terms: an admin sets up a connector once, and everyone on the team gets access automatically on first login — no fiddly per-person sign-ins. Asana, Atlassian, Canva, Figma, Granola, Linear and Supabase are supported at launch, with Slack “coming soon.” Anthropic also shipped “MCP tunnels” (a research preview that lets Claude reach private, behind-the-firewall tools) and the ability for Team/Enterprise users to tag Claude directly inside a Slack conversation and hand it a task. For a global comms team, this is the cleanest path yet to wiring Claude into the tools you already run.
RelevantClaude · Slack · Notion · Figma · team coordination
Obsidian + n8n: auto-distill your AI chats into structured notes
A widely shared workflow uses Obsidian’s “Post Webhook” plugin to fire an n8n automation the instant you save an AI-conversation note. n8n then extracts the key insights, action items and takeaways and writes them back into your vault as clean, structured notes — no copy-paste. A companion template even turns Obsidian notes into natural-sounding audio (via OpenAI) and publishes them as a private podcast feed. For someone who already lives in Obsidian and consolidates reading, this is a practical way to make your vault self-organizing.
Perplexity’s “Personal Computer” brings agentic browsing to your Mac
Perplexity’s upgraded Mac desktop app now includes a “Personal Computer” mode that lets its Comet agent edit local files, browse locally and respond to voice — effectively an assistant that can act on your own machine, not just answer questions. The Comet agent now runs on Claude Sonnet 4.6 by default (Opus 4.6 for Max), and an iOS version is open for pre-order. Worth a careful look for research and drafting tasks, with the usual caution about letting any agent touch local files.
Google AI Studio turns plain-language prompts into working Android apps
Google AI Studio now accepts plain-English instructions and generates a working Android app prototype — no coding required — with one-click export to Google’s “Antigravity” developer environment. For a non-coder who likes to prototype ideas (an internal tool, a simple utility for the team), it lowers the barrier from “learn to build” to “describe what you want.”
The most recent number a reporter is likely to raise: Tencent has signaled it will more than double AI-related spending to north of $5 billion over the next year, underscoring the capital intensity of staying competitive on frontier models and infrastructure. There were no fresh direct Tencent headlines in the past 24 hours, so this remains the key line on the record for media enquiries about Tencent’s AI investment posture.
Two product-level markers worth knowing: WeChat AI entered an internal testing phase in June, letting users invoke and operate mini-programs through natural-language dialogue, and Tencent Cloud documentation lists the older Hunyuan (HY) 2.0 model as discontinued from June 26 — a routine generational hand-off as newer Hunyuan models take over. Tencent’s Yuanbao assistant app has surpassed 100 million monthly active users since its Spring Festival surge.
This week’s U.S. model-gating story (GPT-5.6, plus export controls on Anthropic’s Fable 5 and Mythos 5) is sharpening a broader narrative: Chinese labs leaning into open-weight models that sit outside U.S. export controls. While the headline names are Alibaba’s Qwen and DeepSeek, Tencent’s open Hunyuan releases are part of the same trend — a useful framing if reporters ask how Tencent fits into the U.S.–China AI dynamic.