Presented by
Good morning, AI enthusiasts. The White House just gave Anthropic 90 minutes to pull Fable 5 from the market, and Anthropic complied, cutting off all customer access overnight.
It's the kind of regulatory intervention Anthropic has openly advocated for years. Did they expect to be the first company it was used against?
In today's recap:
White House export controls shut down Fable 5 overnight
Beijing orders Meta to unwind $2B Manus deal
Build multi-agent workflows with Databricks Omnigent
German court holds Google liable for AI overviews
4 new AI tools, prompts, and more
ANTHROPIC
White House shuts down Fable 5 overnight
Anthropic
Recaply: Anthropic just pulled Fable 5 and Mythos 5 for all customers after the White House issued an export control directive blocking foreign access, citing a jailbreak discovered in the model.
Key details:
The directive, citing national security authority, was triggered after Amazon CEO Andy Jassy flagged concerns to senior White House officials, who then ran the findings past the NSA before acting.
Fable 5 launched publicly just days before the directive arrived, with Chinese AI rival Zhipu surging 33% on markets as investors bet on the gap the suspension creates.
The White House gave Anthropic only 90 minutes to comply with no specific threat details shared, according to a person close to the company.
The directive arrived June 12 at 5:21pm ET, with all other Claude models unaffected. Anthropic says it's working to restore access "as soon as possible."
Why it matters: This isn't just a model going offline. It's the first time the US government has used export controls to shut down a deployed AI product. That sets a precedent for how fast this power can land. Anthropic called the directive disproportionate, but it spent years lobbying for exactly this kind of government authority over unsafe AI. The company just found out what that framework looks like in practice. A 90-minute notice with no technical disclosure.
PRESENTED BY NEO
Watch every match the way it was meant to sound.
This summer, 48 nations play. For a lot of fans in the US, the match doesn't feel right in the wrong language. The commentary, the energy, the way goals sound when your language is calling them.
Norton Neo is a free browser with a free built-in VPN. Stream privately in the language you want. No subscription, no sign-up, no credit card. Built-in VPN, anti-fingerprinting, and ad blocking, all backed by Norton security.
Free to download. Free to use.
Fast. Safe. Intelligent. That's Neo..
META
Beijing forces Meta to unwind $2B Manus deal
Getty Images
Recaply: Meta just moved to unwind its $2B deal with Chinese AI agent startup Manus after Beijing authorities issued a divestiture order, cutting the company off from Meta's internal systems and tools.
Key details:
Beijing's divestiture order forced Meta to sever Manus from its internal infrastructure, with both companies cooperating on compliance roughly two months after the order arrived.
The original deal was valued at $2B, and Manus co-founders are now separately exploring a $1B fundraise, suggesting the startup sees a path to independence without Meta's backing.
Investors including Benchmark and Tencent are cooperating with the divestiture, while China has separately expanded travel restrictions on AI researchers and now requires government sign-off for foreign investment in top AI firms.
Beijing's directive arrived roughly two months ago. The full unwind is now in motion, and Manus's independent fundraising effort signals it is moving on without Meta's platform access.
Why it matters: The China-US AI split has been growing for years, but Beijing forcing Meta to unwind an AI deal marks a new kind of direct move. It isn't just a trade dispute. Meta's $2B bet on Manus was about talent and capability, and losing it removes a key piece from its AI agent roadmap. Cross-border AI deals are now political calls as much as business ones, and every lab with cross-border ties is watching.
GUIDES
Build and share agent workflows with Databricks Omnigent
Recaply: In this tutorial, you will learn how to set up Omnigent above your coding agents. You'll add cost policies, share live sessions, and swap agents in one line, all without changing your existing agent setup.
Step-by-step:
Install Omnigent by following the quickstart at omnigent.ai/quickstart/install. Clone the repo with
git clone https://github.com/omnigent-ai/omnigentand runnpm install, then connect your agent by settingagent: claude-code(orcodex,pi) in youromnigent.yamlconfig file.Launch a session with
omnigent start, which wraps your chosen agent in a sandboxed environment with a uniform API and exposes it across web, mobile, terminal, and native Mac OS interfaces at the same time.Add a cost policy by specifying
cost_budget: 100in youromnigent.yamlfile, which pauses the agent and asks for your confirmation after every $100 spent to prevent runaway costs on long-horizon tasks.Share your live session with a teammate by running
omnigent share, which generates a URL they can open in a browser to view the agent's progress, leave comments on specific files, and send commands in real time with no local setup needed.Swap agents or combine harnesses by updating the
agentfield mid-task, for example switching from Claude Code to Codex, or defining a multi-agent workflow in YAML where different subagents use different harnesses and models for different parts of the job.
Pro tip: Run omnigent start --sandbox when working with untrusted tasks or sharing sessions with external clients. The OS sandbox intercepts network requests at the proxy layer, so agents never see your raw credentials even on approved external calls.
TOGETHER WITH WISPR FLOW
Your best prompts are the ones you'd never bother typing.
The detailed ones. The ones with examples and edge cases. Wispr Flow lets you speak them instead — clean, structured, ready to paste into any AI tool. Free on Mac, Windows, and iPhone.
German court holds Google liable for AI overviews
Recaply: Google just received a landmark ruling from a German court holding it directly liable for false claims in its AI Overviews, with the Munich Regional Court classifying AI summaries as Google's own statements, not aggregated search results.
Key details:
The Munich court found that AI Overviews generate "independent, new, and substantive statements" by evaluating and combining third-party sources, making Google the direct creator of the content rather than a passive intermediary.
Google's AI Overviews appear in billions of search queries, and even a 91% accuracy rate means millions of wrong answers generated daily, the court noted as part of its reasoning.
Traditional search engine liability protections don't apply to AI-generated summaries because the AI makes claims not found in its underlying sources, with Google ordered to pay 80% of legal costs.
The temporary injunction issued June 9 bars Google from spreading the false publisher claims. Google says it's "carefully reviewing" the ruling and plans to appeal.
Why it matters: Every AI company giving users AI-generated answers just got a new legal problem. A German court made one thing clear. When AI combines sources to write new statements, the company behind it is a publisher, not a search engine. Publishers can be sued for false statements. Search engines usually can't. Google is appealing, and a German ruling won't change US law right away. But it's the first real template for AI liability cases, and courts elsewhere are taking notes.
TOOLS
Trending AI Tools
🧑💻 Kimi K2.7 Code - Open-source 1T-param MoE coding model with 30% lower reasoning-token usage and stronger long-horizon task completion than K2.6
🔊 ZONOS2 - Open-source sparse MoE TTS model with zero-shot voice cloning at studio-quality 44.1 kHz, trained on 6M+ hours of audio under Apache 2.0
🤖 GLM-5.2 - ZAI's multilingual coding model, deeply integrated with ZCode 3.0 for stronger agent task execution and long-context coding workflows
🔗 Omnigent - Databricks' open-source meta-harness for combining, controlling, and sharing AI agents across Claude Code, Codex, Pi, and custom agent SDKs
NEWS
What Matters in AI Right Now?
Moonshot AI released Kimi K2.7 Code, a 1T-parameter open-source coding model built on K2.6, with 21.8% stronger performance on Kimi Code Bench v2 and 30% lower reasoning-token usage. Available now via Kimi API and Kimi Code.
ZAI launched GLM-5.2, now fully available for Coding Plan users with 150% usage quota inside ZCode 3.0, featuring stronger agent task execution, better long-context coding, and a new Goal feature. New users get 5 days free with 5M tokens per day.
Zyphra released ZONOS2, an open-source TTS model with 8B total and 900M active parameters, trained on 6M+ hours of audio for high-fidelity 44.1 kHz voice cloning, available free on Zyphra Cloud under Apache 2.0.
Google Research introduced Gemini-SQL2, a text-to-SQL capability powered by Gemini 3.1 Pro that achieves state-of-the-art results on the BIRD benchmark, translating natural language into execution-verified SQL for use across Google's data services.
Databricks open-sourced Omnigent, a meta-harness that sits above existing coding agents such as Claude Code, Codex, and Pi, letting teams compose multiple agents, enforce contextual security policies, and collaborate on live sessions via URL.
OpenRouter launched Fusion, an API that sends prompts to a panel of frontier models in parallel, synthesizes results via a judge model, and achieved 69% on the DRACO deep research benchmark, outperforming every individual model tested, including Fable 5 at 65.3%.
OpenAI announced the OpenAI Partner Network, investing $150M to support a global ecosystem of systems integrators and consultants, with a goal of training 300,000 certified AI practitioners by end of 2026.
Goldman Sachs projected that AI infrastructure capital spending could reach $1.1T in 2027, up from a $765B baseline estimate for 2026, driven by accelerating demand across data centers, compute, and energy buildout.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
