Watch on YouTube · subscribe for the daily brief

Meta released Muse Ember on Monday, its smallest and cheapest model, built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long documents.

The numbers

• The claim: more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at every reasoning setting; Ember "matches Haiku's best score at about a third of the cost", Meta says.
• Price: $0.10 per million input tokens, $0.50 per million output, on the Meta Model API and Muse Code.
• Artificial Analysis index: Ember 35.8 for about $0.14 a task; Haiku 5.5's best setting 36.2 for $0.41 (Meta's chart, Nov 9). Meta notes Ember's costs are estimates at proposed pricing.
• The twist: on real coding tasks (FrontierCode 2.0-AA), Haiku wins, 34.9% to 27.2%.
• The design: a pair. "Muse Wonder plans, Muse Ember handles the volume," per TechCrunch. Ember reads text, images, audio and video, up to 1 million tokens.

For builders: test Ember as the subagent on your own tasks, not the leaderboard, and keep coding work on whichever model wins your tests. A third of the cost only matters if the error rate holds on your data.

Security

• One man, twelve agents: a lone attacker in Taiwan used AI agents to pull 119 dispatch records covering seven Taipei districts, per Acronis via The Hacker News. Commercial models mostly refused, so the attack agents ran on "abliterated" DeepSeek, GLM and Qwen models with safety training removed. One scanner invented CVEs; what the agents found was mostly data already exposed, like a medical data port open since 2018.
• Qubole's trusted role: attackers broke into several Qubole customers' AWS accounts from October 14 to 26 using credentials for the account Qubole uses to reach them, Wiz says. Most roles were unused but still trusted Qubole. If you ever connected it, check CloudTrail from October 1.
• Fireworks AI breach: Notion says fewer than 1% of workspaces may be affected; some Notion AI chats were exposed without user or workspace attached.

Money and power

• Anthropic buys turbines: $3.4 billion of Caterpillar natural gas generators, up to 1.7 GW, per the Financial Times via TechCrunch. The company has preferred to rent.
• Credit stress: CoreWeave's 2032 bond yield climbed above 16%.
• Nvidia: the stock slid more than 4% on Monday and at least four law firms announced investigations into whether Nvidia misled investors. Short seller Culper Research had flagged Huang's ties to the Megaspeed executive on May 13.

Models and data

• Gemini 4 Argon: Google's model for robots; in Google's examples it guides a Boston Dynamics Spot after practice in a MuJoCo twin.
• Pinterest: replaced journey clustering with a fine-tuned 4B Qwen3 model; email click-through up 1.1%, push opens up 1.3%.
• Apple sells AI by the pack: Logic Pro offers more Apple Intelligence usage for $49.99, for features that run on OpenAI or Gemini models.
• Gemini and your inbox: Google's help page now says human reviewers may see Gmail and Drive content you connect to Gemini. Our full story is on the channel.

The through-line

Every agent reaches somewhere: a cheap model into your budget, a trusted role into your cloud, an assistant into your inbox. Know what each one can touch.

One thing to do today: list every outside company with a trusted IAM role in your cloud accounts, and remove the ones you no longer use.

Sources covered

1 story · 120 sources (67 news, 48 company, 5 voices)

1. llm-anthropic 0.31
   🎙️Simon Willison↗🎙️Simon Willison↗🏢LangChain↗🏢Vercel AI↗📰SiliconANGLE↗🏢Google Research↗📰TechCrunch↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Databricks↗📰The Verge↗📰SiliconANGLE↗🏢AWS AI/ML↗🏢Diraq↗🏢Databricks↗🏢OpenAI↗📰BleepingComputer↗🏢LangChain↗🏢LangChain↗🏢LangChain↗📰SiliconANGLE↗📰The New Stack↗📰The Verge↗📰TechCrunch↗📰RIA Novosti↗🏢Databricks↗📰TechCrunch↗📰BleepingComputer↗🏢ClickHouse↗🏢Meta AI↗🏢ClickHouse↗🏢ClickHouse↗🏢Google Cloud AI↗🏢ClickHouse↗🏢ClickHouse↗📰The Decoder↗🏢Snowflake↗🏢Fivetran↗🎙️Damien Benveniste↗🏢IBM Research↗📰TechCrunch↗🏢Hugging Face↗📰Hacker News↗📰SiliconANGLE↗📰The Decoder↗📰TechCrunch↗📰The Hackers News↗📰AWS↗📰TechCrunch↗📰Ars Technica↗📰MIT Technology Review↗📰TechCrunch↗📰The Verge↗📰The New Stack↗🏢NVIDIA↗📰The Verge↗📰Hacker News↗📰The Decoder↗📰The New Stack↗📰Hacker News↗📰The New Stack↗🏢MongoDB↗🏢MongoDB↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰The New Stack↗🏢OpenAI↗🏢Google Cloud AI↗📰The Verge↗🏢Hugging Face↗📰Ars Technica↗📰SiliconANGLE↗📰MIT Technology Review↗📰TechCrunch↗📰The Decoder↗🏢Diraq↗📰BleepingComputer↗📰TechCrunch↗📰The Verge↗📰The Decoder↗📰InfoQ↗📰The Verge↗📰The Verge↗📰Ars Technica↗📰BleepingComputer↗📰LWN.net↗📰The New Stack↗📰LWN.net↗📰InfoQ↗📰The Verge↗📰The New Stack↗📰SiliconANGLE↗📰TechCrunch↗🏢Berkeley AI Research (BAIR)↗🏢NVIDIA↗📰MIT Technology Review↗🏢OpenAI↗📰The Decoder↗📰The Decoder↗📰InfoQ↗📰InfoQ↗📰The Hackers News↗🏢NVIDIA↗🏢NVIDIA↗🏢NVIDIA↗🏢Vercel AI↗🏢NVIDIA↗🎙️Simon Willison↗🎙️Simon Willison↗📰InfoQ↗📰TechCrunch↗🏢Temporal↗🏢Temporal↗🏢Elastic↗🏢Anthropic Research↗🏢Vercel AI↗🏢Vercel AI↗🏢Vercel AI↗