Claude Haiku 4.5, Skills, Veo 3.1, and nanochat

What shipped

Anthropic

Claude Haiku 4.5

A small model that Anthropic says matches Sonnet 4 on coding at about a third of the cost and twice the speed.

Anthropic

Equipping agents with Agent Skills

Skills are folders of instructions and scripts that Claude loads only when a task needs them.

Google

Veo 3.1

Richer native audio, reference images and scene extension arrive in the Flow filmmaking tool and the Gemini API.

Andrej Karpathy

nanochat

A small codebase that trains and serves a ChatGPT-style model from scratch for around a hundred dollars of compute.

Baidu PaddlePaddle

PaddleOCR-VL

A compact vision-language model for document parsing with tables, formulas and many languages.

Google DeepMind and Yale

Cell2Sentence-Scale and a new cancer hypothesis

A Gemma-based model proposed a drug combination that lab experiments then partly confirmed.

Anthropic spent the week shipping the pieces of a working agent stack, and I found the smaller ones more interesting than the model.

Haiku 4.5 and Skills

Claude Haiku 4.5 arrived on Wednesday. Anthropic says it reaches roughly the coding level of Sonnet 4 at about a third of the cost and twice the speed. The reason that matters is arithmetic. If you run many agents all day, the cheap model does the bulk of the work and the expensive one is called only for judgment. I tier my own agent work that way, and a better small model shifts the line.

The next day came Agent Skills. A skill is a folder holding instructions, scripts and resources, and Claude reads it only when the task calls for it. That keeps the main context lean while letting you give the agent deep, specific know-how. I've been doing something like this by hand for months, with written procedures for each recurring job, so seeing it become a product feature is validating. It also means a company's operating knowledge can be written once and reused across tools.

Video and building from scratch

Google released Veo 3.1 with better native audio, support for reference images, and the ability to extend scenes, available in Flow and the Gemini API. For short-form content work the reference-image support is the practical gain, since it keeps characters consistent between shots.

Andrej Karpathy published nanochat, a compact codebase that goes from raw text to a working chat model for about a hundred dollars of compute. I don't plan to train a model, but I learn the most about a tool by looking at the smallest honest version of it.

Documents and science

Baidu's PaddleOCR-VL is a small document-parsing model for tables, formulas and many languages. Anyone drowning in scanned PDFs should test it. Google and Yale published work where a Gemma-based model proposed a cancer-related drug combination, and lab tests partly supported it. That is a real example of a model producing a testable hypothesis.

What I'd watch next

OpenAI is rumored to be building its own browser. If it ships, the contest moves from chatbots to who owns the page you're looking at.