The US shut down Anthropic's frontier models. Talks are still ongoing.

subject: 🏛️ The US shut down Anthropic's frontier models. Talks are still ongoing.

Good morning 👋 Six days ago the US government ordered Anthropic to pull its two most capable models offline for every customer on earth. They're still offline. Talks are ongoing in Washington.

In today's issue:

• 🔭 The US shut down Anthropic's Fable 5 and Mythos 5
• 🧠 Models: AgentCore knowledge expansion, Amazon Quick autonomous agents, HF agentic evals
• 🔬 Research: AMIE in Nature, near-autonomous drug discovery, enterprise data agents
• 🛠️ Tools: benchmark your agent before you ship, plus 5 trending
• 🔗 Quick links

🔭 THE ONE THING

🏛️ The US shut down Anthropic's frontier models. Talks are still ongoing.

Three days after Anthropic launched Fable 5 and Mythos 5 on June 9, the Commerce Dept's Bureau of Industry and Security issued a directive: suspend all access for any foreign national, whether inside or outside the US. Anthropic couldn't distinguish US from non-US users at runtime and disabled both models for every customer globally, without warning. The stated trigger was another company claiming to have jailbroken Mythos; Anthropic disputes the severity, calling it a "narrow, non-universal jailbreak" rather than a systemic exploit. Senior technical staff have been in Washington meeting Commerce officials since June 15, and as of yesterday no deal exists and no restoration date has been set. If your production stack runs on Fable 5 or Mythos 5, you're still locked out. Anthropic statement

🧠 MODELS & RELEASES

• 🤖 Amazon Bedrock AgentCore launched new capabilities today: agents can now tap live web search, paid knowledge APIs, and org-specific knowledge bases, plus a continuous learning loop that refines performance from production feedback. A direct path from capable in testing to getting smarter in prod. link

• 📊 Amazon Quick added autonomous agents to its business intelligence platform. They work continuously, surface insights across connected data sources, and push the most important actions to the top. The target is the manual "go pull this report" loop that still eats hours in most orgs. link

• 🧪 Hugging Face published "Is it agentic enough?" An open evaluation methodology for scoring open models against your actual tool stack, not synthetic benchmarks. A model that aces standard evals may still fail badly at calling your specific tools in your specific context. link

🔬 RESEARCH HIGHLIGHTS

• Google AMIE in Nature. Google's medical AI system matched primary care physicians in managing complex, multi-condition diseases over extended patient consultations, per a peer-reviewed study published in Nature. The comparison covered real-world chronic disease management scenarios across a range of conditions. This is not a toy benchmark. link

• Near-autonomous AI chemist. OpenAI and Molecule.one used GPT-5.4 to systematically improve Chan-Lam coupling, a drug-synthesis reaction where yields have historically been too low for practical pharmaceutical use. The AI proposed experimental directions; Molecule.one ran synthesis experiments; human chemists validated representative results and confirmed yield improvements across a majority of compounds tested. Full cycle: about 2.5 months. link

• Data Intelligence Agents (arXiv). A system of three agents (Data Interpreter, Schema Creator, Query Agent) that automates the handoff between data owners, engineers, and analysts in enterprise pipelines. The central claim: the biggest bottleneck in production data integration is repeated, lossy translation between roles, and coding agents can close most of it. link

🛠️ TOOLS & HOW-TO

Benchmark your agent against real tools before you ship

1. Pull 20 to 30 real tasks from your production logs: actual user intents, not invented examples

2. Define your tool schema clearly with expected inputs, outputs, and failure modes for each call

3. Run at least two open models side-by-side using the HuggingFace "Is it agentic enough?" framework

4. Score on task completion rate and tool-call accuracy, not just response quality

5. For every failure: model reasoning problem or poorly documented tool? Fix the right thing first

Prompt: Here are [N] tasks where my agent failed. Each includes the user request, the tool it called, and the actual vs. expected output. For each failure, determine whether it's a model reasoning issue or a tool specification problem. For model reasoning failures, suggest a better chain of thought. For tool specification failures, rewrite the tool description.

5 trending right now

• 🤖 MiMo-Code (Xiaomi): open framework for co-evolving model training and agentic deployment in production; 9,700+ GitHub stars. link
• 📊 "Is it agentic enough?" (Hugging Face): open benchmark harness to score open models against your actual tool stack, not toy tasks. link
• ⚡ P-EAGLE (AWS SageMaker): parallel speculative decoding that cuts inference latency on existing SageMaker endpoints without retraining. link
• 🧬 LifeSciBench (OpenAI): expert-authored benchmark across 6 life science task categories; useful for stress-testing any model on domain-specific scientific reasoning. link
• 🌐 SageMaker Async Inference inline payloads: send payloads directly in the request body, no S3 pre-upload required. Simpler and faster async workflows. link

🔗 QUICK LINKS

• Google DeepMind published an AI Control Roadmap for securing AI agents in internal systems, combining traditional safeguards with real-time behavioral monitoring and kill-switch mechanisms. Worth reading if you're deploying agents with production system access. link
• NVIDIA at VivaTech Paris: a year after France announced its AI ambitions, the factories are live. Agents running in production, open frontier models deployed, startups shipping. link
• Coherent broke ground on an expanded Texas facility that makes the lasers, optical components, and compound semiconductors that interconnect AI systems at scale. The physical infrastructure AI actually runs on. link

See you tomorrow.

Pradeep Perugu