Claude wrote so much code that Anthropic had to replace Git

The usual reason to use somebody else's software is that you have other things to do. You could build your own system for tracking changes to code, but Git exists, it works, and presumably your business is something other than inventing Git. AI makes that calculation more interesting. It can make the replacement cheaper to build. It can also produce enough work to make you need the replacement.

Anthropic ran into both sides of this. Its daily merged code changes went from about 2,500 to more than 30,000 between May and September. Git is designed to let people keep complete repositories and work offline. Anthropic had a great many agents that were online already. It was paying to preserve a freedom its busiest users had little use for.

So seven engineers, with Claude, built Scion and a collection of related tools, and migrated in five months. They also acquired responsibility for their own code review, search, and CI. That is quite a lot of other things to do. Anthropic says ordinary Git users should keep using Git.

We expect more teams to build the part that makes their work different and buy the parts that don't. Anthropic needed a new source-control system. A booking app still needs someone else's payment processor.

Top Stories#

Meta released Muse Spark Open, an open-weight model. After releasing Muse Wonder earlier in the week, Meta released Muse Spark Open on October 15: a multimodal model with 816 billion parameters, 42 billion active at a time. Meta demonstrated it running on two M5 Ultra Mac Studios using Apple's MLX framework. Two very well-equipped desktops are still quite a commitment, but developers can download and customize the model rather than depend entirely on a hosted API. The weights come under Meta's Muse Model License, with limited commercial terms.

Cloudflare's decision models now take audio and video. Clef-omni extends the company's open-weight models beyond text and images. Decision models score possible answers instead of composing a free-form response: whether a recording contains a particular sound, for example, or whether a video shows a fault. Cloudflare also announced a faster Clef and cheaper Clef-flash. These releases broaden the tasks that can use a specialized decision model, without asking a general-purpose assistant to explain every choice.

The Jev copycats are getting good by fine tuning other open models. In Thursday's community-maintained Jev Decision Index, the top two entries are d3, built on Qwen3.8-27B, and linne-oak, built on Gemma 4 31B. They score 66.3 and 65.6, compared with Jev 1.14's 61.5. Both are full fine-tunes: they update the existing model's weights rather than train a foundation model from scratch. They are still specialized decision models, but the specialization starts with something already on the shelf. The ranking measures decision quality; it doesn't establish which deployment is cheapest or fastest.

Anthropic is letting outside evaluators look around inside. Six employees from METR, a nonprofit that evaluates frontier AI, now have access roughly comparable to Anthropic's internal risk assessment teams. They can examine models, safeguards, and practices, then publish their findings without Anthropic vetoing the report. Defined categories of confidential information can still be redacted, and METR can publicly dispute those redactions. Anthropic compares the arrangement to resident bank examiners: outsiders need to see what the institution knows about its own risks. The arrangement is voluntary, but it gives independent evaluation considerably more to work with than a model's answers alone.

Top Links#

• AI in Practice 2026 (theoryvc.com): Theory Ventures' survey of 455 technical builders finds teams connecting models to their own systems while customizing the models less; owning the application doesn't require owning every layer.
• Agent mode in DuckDB (duckdb.org): A new CLI mode makes truncation explicit and output compact; its experiment cut result tokens substantially without reducing total cost or runtime, a useful reminder to measure the whole job.
• A 74B model in 24GB (nvidia.com): NVIDIA's new Nemotron 3.5 Bolt uses aggressive quantization to bring a large model to local hardware; the useful comparison is how well it does your work within the memory you have.
• Databricks versus an iPhone (fivetran.com): Fivetran's CEO ran analytical queries on his phone with DuckDB and compared them with cloud clusters; the phone did surprisingly well, provided it could keep cool.