Watch on YouTube · subscribe for the daily brief

Meta ran Claude Code until it could build its own. Then it cut the user count in half. Your biggest customer is always your next competitor.

Meta had about 60,000 employees using Claude Code earlier this year. It now has about 30,000. The other half moved to coding agents Meta built itself.

That comes from The Information, which reported Monday that Meta and Microsoft are both steering staff off Anthropic's tools. Microsoft had planned to spend at least $1 billion on Anthropic this year for internal use. It has cut that figure by about a third.

These were two of Anthropic's biggest customers. They learned on Claude, then built their own.

What actually moved

Meta's replacements have names. MetaCode, an internal-only tool, has passed 30,000 users. Muse Code, the external Claude Code competitor Meta started testing in August, has over 6,000 internal users. It runs on Meta's own models.

Microsoft's story started in May, when it began pulling Claude Code licenses and moving engineers to GitHub Copilot CLI. Cost was the reported reason. The new cut goes further. It reportedly covers Claude Code, Claude models inside Copilot, and Mythos.

Now the other side of the ledger. Microsoft's investment of up to $5 billion in Anthropic is untouched. So is Anthropic's $30 billion commitment to buy Azure compute. And customer spending on Claude through Azure and Bedrock keeps growing. Microsoft still sells Claude. It just doesn't want to pay for it at home.

The Decoder framed it as Anthropic turning from partner into competitor. That's fair in both directions. Anthropic sells a coding agent that competes with Copilot. Meta sells an agent that competes with Claude. Everyone in this story is somebody's supplier and somebody's rival.

Why this lands on your desk

The coding agent turned out to be the easy part to copy. A company with its own model and enough engineers can rebuild the harness in months, and these two did. Two weeks ago, AWS showed that the wrapper around a model can change its cost about fourfold. Now two of the largest buyers decided the wrapper was cheaper to own.

Most teams can't build their own Muse Code. But you can avoid depending on one vendor's shape. Keep your prompts, tools and evals portable. Make the model a setting, not an architecture. Databricks shipped that idea for catalogs today, with register and unregister APIs so your tables aren't locked to one catalog.

And watch the meter. Cohere launched North 2 with token spending caps built in. Iterate.ai says its Lifeboat runs up to six times more agent sessions per GPU. The New Stack found one MCP server burning 18,000 tokens before it did anything. Microsoft's complaint was cost. Yours will be too.

OpenAI picked its defaults

OpenAI said Monday it will watermark ChatGPT and Codex text in the European Union over the coming weeks. API customers anywhere can opt in starting today. It is off by default. OpenAI says it is not making text watermarking a global default at launch.

Anthropic made the opposite call two months ago and applied its watermark worldwide. OpenAI's own figures say its detector catches about 80 percent of its text at 200 tokens, under ideal conditions. And OpenAI's tests suggest editing can remove it.

The same day, OpenAI said it is putting more ads in ChatGPT, including visual ads while you generate images. A company that defaults to no watermark and more ads is telling you what it optimizes for. Anthropic, meanwhile, is reportedly becoming America's biggest corporate political donor ahead of its IPO. Nobody here is a neutral utility.

Builder's bench

• The Wikipedia operator says OpenAI's rogue agent bots may be linked to a May outage. That extends last week's story. Researchers are also tracking a Chinese AI agent fleet.
• Ars Technica calls agent-to-agent MCP the riskiest protocol you've never heard of.
• Open weights kept coming. Reflection's Beam targets Chinese models at lower compute cost. Aleph Alpha's Kolibri makes the case for European sovereignty. Reka's Rho-1 handles text, images, video and robot control in one model. A safety check running on a laptop nearly matched a 35-billion-parameter model.
• The New Stack says developers are quietly hoping OpenAI misses its launch this month. Dynatrace's $915 million Arize deal, announced in August, bets agents are just another app to monitor.
• On the data stack, NetApp's Novus targets metadata bottlenecks, and NetApp is pushing sovereign storage in Europe. Starburst wrote on the data foundation for production AI. DuckDB can now import data through Java table functions. MIT Technology Review covered connecting agents to enterprise knowledge and predictive analytics for agents.
• On practice, InfoQ published a talk on reusable evaluation frameworks and a platform engineering playbook for production LLMs. Akka tested spec-driven delivery across 65 open source projects. XCOR traces outages in minutes but still pages engineers. LWN updated its Sashiko patch-review system. Simon Willison quoted Felix Rieseberg.
• On security, Denmark's population registry breach hit 8.8 million people. South Korea is probing bank breaches it suspects were AI-powered. AI glasses face their first major government crackdown. An open-source tool deletes 12 gigabytes of Apple Intelligence data on macOS.
• On startups, Ghost, led by a 19-year-old founder, raised $11 million for agent hardware. Equs launched a private personal AI platform. HackerRank showed an AI interviewer, one startup issues AI-written acne prescriptions, and Safeworld wants to prove robots are safe. For physical AI, the bottleneck is now whether factories trust the robots.
• And the mood. Most Americans want AI slowed or stopped, a new poll finds. Sam Altman says some bad things will happen but AI is worth it. An OpenAI spokesperson told a reporter to move on after a question to Altman about a user's suicide. An MIT report finds AI eroding office hours and trust on campus. And the fight over AI taking over mathematics continues.

Through-line: the model is the product you rent. The agent around it is the product your biggest customer will eventually build.

One thing to do today: list every place your code names a specific model or vendor SDK. Pick one and put it behind a setting you can change without a deploy.

Sources covered

1 story · 68 sources (63 news, 4 company, 1 voices)

1. Quoting Felix Rieseberg
   🎙️Simon Willison↗📰SiliconANGLE↗📰SiliconANGLE↗📰SiliconANGLE↗📰BleepingComputer↗📰Ars Technica↗📰SiliconANGLE↗📰The New Stack↗📰The New Stack↗📰TechCrunch↗📰The Verge↗🏢Databricks↗📰TechCrunch↗📰The New Stack↗📰The Verge↗📰Ars Technica↗📰The New Stack↗📰The Verge↗📰The Decoder↗📰The New Stack↗📰The Decoder↗📰The Verge↗📰SiliconANGLE↗📰The Decoder↗📰TechCrunch↗📰The Verge↗📰The Verge↗📰TechCrunch↗📰The Decoder↗📰The Decoder↗📰MIT Technology Review↗📰BleepingComputer↗📰TechCrunch↗📰LWN.net↗📰The New Stack↗📰TechCrunch↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰BleepingComputer↗📰The Decoder↗📰TechCrunch↗🏢OpenAI↗📰InfoQ↗📰The Verge↗📰Ars Technica↗📰MIT Technology Review↗📰The Verge↗📰SiliconANGLE↗📰SiliconANGLE↗📰SiliconANGLE↗📰SiliconANGLE↗📰SiliconANGLE↗📰The Decoder↗🏢Starburst↗📰TechCrunch↗📰The Decoder↗📰InfoQ↗📰InfoQ↗📰Hacker News↗📰BleepingComputer↗📰The Verge↗📰BleepingComputer↗📰Hacker News↗📰MIT Technology Review↗📰Hacker News↗📰MIT Technology Review↗🏢DuckDB↗