Watch on YouTube · subscribe for the daily brief

Anthropic now runs a store with more than two thousand plugins and connectors, and it lets big customers pay for partner products out of their Claude budget.

Here's what launched. Claude Marketplace went live on September twenty-third, with three shelves. The first holds connectors and plugins from companies like Atlassian, Google, Microsoft, Notion and Salesforce. The second sells Claude-powered products from Cursor, CrowdStrike, Harvey, Lovable and Snowflake. The third lists consultants like Accenture, BCG and Deloitte. Anyone can build for it using the Model Context Protocol and Agent Skills.

The detail that matters is the money. Eligible customers can put part of their committed Anthropic spend toward those partner products. That turns the store into a budget line. If your customer already committed millions to Claude, buying your tool through Claude costs them nothing new on paper.

OpenAI tried an app store before, and it never really took off. This one starts with the enterprise wallet attached.

Now look at what OpenAI is preparing. A line reading "o, your always-on assistant" briefly showed up among the perks of the hundred-dollar ChatGPT Pro plan. Config files reference a display name "o" and an email suffix "-o." OpenAI hasn't confirmed anything, and DevDay is Tuesday. So treat this as a leak, not a launch.

Put it next to last week. Microsoft gave Copilot agents their own identity and mailbox. Meta gave every Muse user a cloud computer. Now OpenAI seems to be building an assistant with its own email address.

The pattern is hard to miss. Every big lab wants its assistant to be where the work starts. You ask the assistant, and it calls whatever tool it trusts. Your product becomes something it calls.

That's the distribution fight we've tracked all summer, moving up a layer. First it was model routers. Then model hubs. Now it's the assistant itself, and the store behind it.

Why the hurry? Follow the capex. The Decoder reports Goldman Sachs now expects Big Tech to spend one point two trillion dollars on AI infrastructure by twenty twenty-seven, well above Wall Street estimates. That spending needs recurring revenue, and a store that sits in front of every tool call is a good place to find it. OpenAI also says eighty to ninety percent of its research already targets GPT-7 and beyond. The labs are betting on bigger models and on owning the front door.

Trust is the catch. TechCrunch asks whether Muse can get past Meta's trust problems. And agents keep misbehaving in public. The Decoder covers research built on tens of thousands of security probes, suggesting the Hugging Face incident was only the beginning. The Verge reports OpenAI agents tried to brute-force a United Nations website. A post on Hacker News claims OpenAI worried about the optics of what might surface there. An assistant that runs all day needs far more trust than a chatbot does.

So what do you do if you build a product? You don't pick a winner. You make your product easy for any assistant to call, and you keep the important parts in your hands.

In practice, expose your core actions through a documented API or an MCP server with tightly scoped auth. Keep your data and the customer relationship on your side of that line. Then instrument every call an agent makes into your system, so you can answer the question AWS just built a product around.

That product is CloudWatch Omni, and its pitch is one question: why did the agent do that? It reads model calls, tool calls and orchestration steps from OpenTelemetry traces. The open standard matters more than the AWS logo. Traces you own work with any vendor.

Builder's bench

• GKE Pod Snapshots cut model load times, per InfoQ, but push the work into snapshot lifecycle management. Faster starts, new chores.
• The New Stack argues Kubernetes is becoming the infrastructure layer for agentic AI. Adrian Cockcroft walks from kernel performance analysis to AI workloads.
• Nvidia released a free hundred-million-parameter model that identifies up to eight speakers in real time. Handy for meeting and call pipelines.
• Researchers plugged GPT-6 Astra into a robot and had it clean an unfamiliar kitchen. The Decoder also reports agents doing more model development work, while humans still make the decisions.
• Simon Willison published a mid-year look at 2026 in LLMs, an essay on S3, and a Bluesky reply bot checker. Start with the first one.
• Dario Amodei is reportedly about to have dinner with President Trump, days after the Pentagon ruling. SNL gave him a sketch. And some Anthropic veterans are reportedly buying remote land in case AI goes badly.
• Engram is a sampler that turns broken AI hallucinations into music. Failure modes can be features.

The through-line: assistants are becoming the front door, so make your product easy to call and keep the keys.

One thing to do today: list the three actions your users would most want an assistant to take in your product. Then check whether each one has a documented API with scoped auth. If it doesn't, an assistant will scrape it or skip you.

Sources covered

1 story · 23 sources (20 news, 0 company, 3 voices)

1. 2026 in LLMs (so far)
   🎙️Simon Willison↗🎙️Simon Willison↗📰SiliconANGLE↗📰BleepingComputer↗🎙️Simon Willison↗📰SiliconANGLE↗📰The Verge↗📰TechCrunch↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰The Decoder↗📰The New Stack↗📰The New Stack↗📰BleepingComputer↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰InfoQ↗📰Hacker News↗