DevDay apps, Gemini drives a browser, 7M-parameter reasoning

What shipped

Google DeepMind

Gemini 2.5 Computer Use model

A Gemini model that clicks, types and scrolls through a real browser through the API.

OpenAI

Apps SDK

Lets developers build interactive apps that run inside ChatGPT, announced at DevDay.

Google DeepMind

CodeMender

An agent that finds and patches security bugs, with human review before anything is merged.

inclusionAI (Ant Group)

Ling-1T

A one-trillion-parameter open model with roughly 50B active per token.

Samsung SAIL Montreal

Less is More: Recursive Reasoning with Tiny Networks

A 7M-parameter recursive network that does well on Sudoku, mazes and ARC-AGI puzzles against far larger models.

Anthropic

A small number of samples can poison LLMs of any size

Found that about 250 poisoned documents were enough to backdoor models regardless of size.

OpenAI's developer day landed this week, though the item I found most useful for my own work came from Google.

Models that use a computer

Google DeepMind released the Gemini 2.5 Computer Use model through the API. It looks at a screenshot, then clicks, types and scrolls to finish a task in a browser. The practical case in my world is every government and supplier portal that has no API. Etimad, vendor registration forms, customs systems. An agent that can operate the interface a human operates is the only integration those systems will ever have. It's early and it will make mistakes, so anything involving submission stays behind a human approval.

At DevDay OpenAI introduced the Apps SDK, which lets developers build interactive apps that live inside ChatGPT, along with an agent-building toolkit and general availability for Codex. I'm linking the Apps SDK docs because the launch posts blocked my link check. The strategic point is that ChatGPT is trying to become the front door, the place a user starts, and other software lives behind it. If you sell to businesses, that will matter for how people find you.

DeepMind also introduced CodeMender, an agent that finds and fixes security vulnerabilities in code and sends patches for human review. Security patching is dull work with real consequences, and a good candidate for agents.

Big and tiny

Ant Group's inclusionAI opened Ling-1T, a trillion-parameter model with about 50B active parameters. At the other end, a Samsung researcher posted a paper on tiny recursive models: a network with 7M parameters that solves hard puzzles by looping on its own answer. I wouldn't read it as the end of large models. I read it as evidence that architecture still has room to surprise people.

Security

Anthropic, with the UK AI Security Institute and the Alan Turing Institute, showed that roughly 250 poisoned documents can plant a backdoor in a language model, and the count didn't grow with model size. That overturned an assumption many of us had made, that scale dilutes bad data.

What I'd watch next

A smaller, cheaper Claude. Sonnet 4.5 is out, and Haiku has been quiet for months.