Claude Code on the web, DeepSeek-OCR, Anthropic's TPU deal

A thinner week on model launches, but a busy one for the people who build with them.

Agents leave the laptop

Anthropic put Claude Code on the web. You describe a task in the browser, it runs in a cloud sandbox against your GitHub repository, and you can start several at once. The same week it launched Claude for Life Sciences, with connectors to research platforms and skills for tasks like protocol drafting. I care less about the biology than about the pattern. A general model gets a set of connectors and written procedures for one profession, and suddenly it's a vertical product. That's the template I'd use for procurement or accounting.

For my own operation the web version matters because it removes my machine from the loop. A job can run while I'm asleep or on a plane, and the output waits as a pull request for me to review.

Reading pages as pictures

DeepSeek released DeepSeek-OCR, and the interesting claim isn't about OCR at all. The team asked whether a page of text can be turned into an image and stored in far fewer vision tokens than the text itself needs, then decoded back. If that holds up, long documents get cheaper to keep in context. It's a research result and I'd treat it as one, but tender packs are mostly scanned PDFs, so I'm watching it closely.

Alibaba added dense Qwen3-VL models at 32B and 8B. They're much easier to run on a single machine than the 235B version and cover document and video understanding.

Compute

Anthropic announced it will use up to a million Google TPUs, bringing well over a gigawatt of capacity online in 2026, in a deal worth tens of billions of dollars. OpenAI also shipped its ChatGPT Atlas browser this week, though its launch page blocked my link check, so I'm not listing it above.

What I'd watch next

More vertical packages from the big labs. If every lab sells a finance, legal or science bundle, the question for a small builder is where the thin layer of local knowledge still wins.