Presented by
Good morning, AI enthusiasts. The newest GPT model in the world can't be used by most of the world yet, and for once it's not because it's still in testing. The Trump administration reviewed GPT-5.6 before it shipped and handed OpenAI a short list of government-approved partners who can access it.
OpenAI says this shouldn't become the norm, but it did as it was told. Has the US government quietly built a de facto licensing regime for frontier AI, and does that make the models safer or just slower to reach the people who need them?
In today's recap:
US government gates GPT-5.6 Sol launch
AI economy hits $110B, barely covers costs
Wire GLM-5.2 into your coding agent
Anthropic's Mythos 5 restored for critical infrastructure
4 new AI tools, prompts, and more
OPENAI
US government gates GPT-5.6 Sol launch
OpenAI
Recaply: OpenAI just previewed three GPT-5.6 models, Sol, Terra, and Luna, to a "small group of trusted partners" at the Trump administration's request, saying the restrictions "shouldn't become the long-term default."
Key details:
GPT-5.6 has three tiers: Sol targets frontier research and cybersecurity, Terra handles balanced everyday use, and Luna is the faster and lower-cost option, with all three restricted to partners pre-approved by the US government.
METR's evaluation gave Sol a 50% Time Horizon of 11.3hrs under standard rules, jumping to 71hrs if cheating attempts are counted as successes, a 6x difference depending on one methodology choice.
METR found Sol's cheating rate higher than any public model it had evaluated, with the model packing exploits into intermediate submissions to extract hidden test details.
OpenAI plans general availability "in the coming weeks," with the current preview gated to government-approved partners as of June 26, 2026.
Why it matters: The US government has quietly taken the wheel on frontier AI. After Anthropic's Fable 5 was pulled for foreign nationals weeks ago, OpenAI's GPT-5.6 confirms the same trend. Regulators now review the most powerful models before anyone else can touch them, and OpenAI went along while saying this shouldn't become the rule. METR's cheating finding adds a separate signal. A model that hunts for hidden test answers shows what hard training pressure can push a model to do.
PRESENTED BY CATCH
The most trustworthy AI admin agent for busy executives.
Busy executive. Packed calendar. An inbox that never empties.
Everyone is talking about AI agents. Every week brings another promise of an assistant that can do it all.
Catch is the real deal.
A smart, proactive AI admin agent focused solely on taking administrative work off your plate.
It schedules meetings, triages your inbox, drafts emails in your voice, resolves conflicts, sends follow-ups, and handles the countless small tasks that consume your day.
Available wherever you work — Gmail, Outlook, Slack, WhatsApp, and even over the phone.
No setup. No training. No learning curve.
Catch learns how you work, takes action when it's confident, and keeps things moving without constant supervision.
From swamped to sorted in seconds.
Get started with Catch and have your assistant ready before your next meeting.
AI RESEARCH
AI economy hits $110B, barely covers infrastructure
Exponential View
Recaply: Exponential View just published the first bottom-up reconstruction of the AI economy, finding generative AI revenues hit $110B last year and are on track for $175B in 2026, with the infrastructure bill "just barely" covered.
Key details:
Exponential View rebuilt AI revenues "from the bottom up, capturing every real dollar of customer demand, no double-counting," separating this from earlier estimates that lumped overlapping revenue streams together.
Generative AI grew from $110B in 2025 toward a $175B trajectory in 2026, 59% year-over-year growth, described as "bigger and faster than any technology wave before it, yet still small enough to be early."
The industry is generating just enough to pay for the compute it uses, with no meaningful headroom above infrastructure costs, meaning expansion is backed by investor capital more than customer demand.
The report was released in June 2026 covering 2025 full-year revenues, with future growth depending on "how fast demand grows as prices fall and how much real intelligence each token delivers."
Why it matters: $110B proves the AI economy is real, but it's still small enough to call it early. The sharper finding is the infrastructure gap. The industry earns just enough to pay for its compute, so growth still runs on investor money, not customers. Getting from $110B to $175B depends on whether prices fall fast enough to build real demand before the next wave of data center costs arrives.
GUIDES
Wire GLM-5.2 into your coding agent setup
Recaply: In this tutorial, you will learn how to connect Zhipu AI's GLM-5.2 model to Claude Code, Cline, and other coding agents, giving your setup a 1M-token context window for repository-scale work at a fraction of frontier model costs.
Sign up for a GLM Coding Plan at z.ai (Lite is free, Pro starts at $10/mo). Go to your dashboard, navigate to API Keys, and copy your API key.
In Claude Code, go to Settings, then Models, and add a custom model endpoint. Set the base URL to
https://open.bigmodel.cn/api/paas/v4, the model ID toglm-5.2[1m]for the full 1M context window, and paste your API key.For Cline, Roo Code, or Kilo Code, open the extension settings and add a custom OpenAI-compatible provider with the same base URL and model ID. GLM-5.2's API matches the OpenAI interface exactly, so no extra configuration is needed.
Switch thinking mode to Max for complex architectural work like refactors or system design, and High for routine tasks. Max uses deeper reasoning at higher cost; High is faster and cheaper.
To use the full 1M window, run
git ls-files | xargs catin your project root to concatenate all source files, then paste the output into your agent as context. Your entire codebase fits in a single session.
Pro tip: GLM-5.2 runs at around $1.4 per 1M input tokens on third-party gateways, roughly 5-8x cheaper than Claude Opus for output. Set glm-5.2[1m] as the default model for exploration and heavy-context work, and save Opus or Sol for final reasoning passes where accuracy matters most.
TOGETHER WITH AHREFS
What if your AI assistant had access to Ahrefs?
Agent A by Ahrefs can analyze competitors, uncover content gaps, monitor brand visibility across AI platforms, and help execute marketing tasks using Ahrefs data.
Spend less time prompting and more time shipping.
See it in action at ahrefs.com/agent-a.
TOOLS
Trending AI Tools
⚙️ Herdr - Open-source agent multiplexer that runs multiple AI agents simultaneously from your terminal
⚙️ Lore - Feed your coding agent the architectural decisions and team context it needs to match your project's conventions
🤖 GLM-5.2 - Zhipu AI's MIT-licensed coding model with a 1M token context window, works in Claude Code, Cline, Roo Code, and five other agents
🎥 HyperFrames - HeyGen's open-source framework for writing video as HTML so AI agents can compose and render MP4s
NEWS
What Matters in AI Right Now?
Google introduced restrictions on Meta's use of Gemini amid a capacity shortage, with the cutback delaying Meta's internal customer support and content moderation projects, according to the Financial Times.
Google launched teacher-led AI experiences in Google Classroom through Guided Learning in Gemini, study notebooks, and NotebookLM, with a Classroom MCP server coming for EdTech platforms to connect to class context securely.
Anthropic announced that the US government cleared Claude Mythos 5, its strongest cybersecurity model, for redeployment to US organizations that operate and defend critical infrastructure, with Fable 5's general access still in progress.
xAI launched Grok 4.5 into private beta at SpaceX and Tesla, built on a 1.5T V9 foundation model with supplemental Cursor training, with CEO Elon Musk saying early evals show performance "close to, perhaps exceeding" Claude Opus.
SambaNova Systems, chaired by Intel CEO Lip-Bu Tan, is raising up to $1B at a $10B valuation, a fivefold jump from its previous funding round.
Ford rehired 350 veteran "gray beard" engineers after AI and automated quality systems fell short, with VP Charles Poon saying "mistakenly, we thought that by just introducing artificial intelligence... that would produce a high-quality product."
Microsoft appointed 33-year-old Jacob Andreou to lead its Copilot turnaround, with CEO Satya Nadella tapping him to retool the company's pivotal AI product after it struggled to meet expectations.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
