Gemini Decisions API, A2UI 1.0, and More
Google shipped something genuinely different on Wednesday: the Gemini Decisions API, powered by a model called Gemini Decision 1. It doesn't write text at all. You hand it one input (text, images, audio, video, or PDFs) and up to 1,000 yes/no, multiple-choice, or rubric questions, and it scores all of them in a single pass, returning calibrated probabilities in roughly 150 to 250 ms.
Pricing is $0.10 per million input tokens, and output is free. Google is pitching it at agent routing, policy checks, and grading, which is exactly the glue work most of us currently do by prompting a full chat model and parsing its answer. Artificial Analysis gave it a 66 on its Decision Index, and Simon Willison wrote up how it compares with Jev and OpenAI's decision model. If you run classifiers or guardrails inside an agent loop, this is worth an afternoon. (Google blog · Sundar's post · Artificial Analysis · Simon Willison)
What Else Happened
A2UI 1.0, Google's open protocol for generative UI, is now stable. Agents describe what to show, and your app renders it with your own registered design-system components in React, Angular, Lit, Flutter, SwiftUI, or Compose. Version 1.0 adds two-way function calls, accessibility metadata, and agent SDKs for Python, TypeScript, and Dart. Atlassian, Bloomberg, and ByteDance are already building on it, and it works inside MCP Apps in Claude and Gemini. For anyone maintaining a design system, this is the one to read closely, because it keeps agents on your components instead of improvised markup. (Google Developers Blog · a2ui.org)
DeepMind's AlphaProof Code formally verified Signal's protocol down to the compiled code. Agents powered by Gemini Carbon wrote about 6.5 million lines of Lean proofs in four weeks, covering PQXDH and the Triple Ratchet in the compiled libsignal code, and humans only had to review about 1% of it. Along the way it caught two historical bugs that had been re-introduced. The proofs are public on GitHub. (DeepMind blog · DeepMind post · Lean proofs)
Alibaba opened a two-week Qwen4-Max preview. It's the next flagship multimodal model, with a 1M-token context, aimed at coding, frontend design, and agentic work, and it speaks both the OpenAI and Anthropic APIs. It's Token Plan only, with no benchmarks or weights yet, and the preview is retired on Nov 25. (Alibaba Cloud docs · early demos on X)
AWS released AWS Emulator, a local copy of 200+ AWS services written entirely by AI agents. It's free, runs on your laptop or in CI with the same APIs as the cloud, and needs no account. AWS says none of the code was human-written or human-reviewed, which is either the most impressive or the most nerve-racking detail of the week, depending on your on-call rotation. (AWS announcement)
Artificial Analysis launched AA-Robotics v0.1, and Claude Fable 5.2 leads it. Fable 5.2 scored 75, ahead of GPT-6 Astra at 69, and was the first model to score on the measure-box task using video alone. Gemini 4 Carbon broke a gripper during its 7.5-hour run. (Fable leads · Gemini broke an arm)
Runway made Praxis-1 open weights under Apache 2.0. It's a world action model that turns video pretraining into control for real robots. Runway also added Seedance 2.5 in 4K (Krea has it too), and showed Runway inside Claude with Fable 5.2 taking a flute from concept to a Blender prototype. (Runway · Seedance on Runway · Seedance on Krea · flute demo)
Gemini Managed Agents now accept your own sandbox. You can bring self-hosted environments for agents on the Gemini API. (Google AI Studio)
Perplexity red-teamed its computer-agent sandbox. It asked 12 models to escape the VM and bypass network policy, and published the research. (Perplexity)
DeepMind announced AlphaGenome 2. The new model predicts how DNA variants affect specific cell types, to help pinpoint disease drivers. (DeepMind thread)
levelsio turned Interior AI into video-to-3D. Upload a walkthrough video, photos, or a floor plan, and you get a walkable 3D room where you can empty it, move furniture, and repaint anything. He built it in about three hours on a headless Blender VPS. (demo · first working version)
Mariko Mori used Claude to write the code for a generative animation in her Tokyo retrospective, "All That Shines." A nice reminder that these tools land in galleries too, not just in repos. (Claude post)
Design Craft
A few posts worth your time if you care about how things feel. Emil Kowalski shared an app-polish prompt that has agents test at different widths, with touch emulation and throttled 3G, and capture before and after screenshots. Rauno Freiberg made the case for why one novel idea beats ten prompted side projects, and Guillermo Rauch posted a reminder that great products still take extraordinary human focus.
What I'm Reading
- Matt Pocock's Home Repo pattern – a setup for running agents across many repos, with a prompt to get started.
- Simon Willison on the Gemini Decisions API – a practical comparison against the other decision models.
- @bigbrovar's solar planner – instead of open-sourcing his own install, he built a tool that matches your load and budget to a full solar design.
Accounts Worth Following
- @mattpocockuk – Matt Pocock, the TypeScript teacher, now sharing practical agent workflows and open agent skills.
- @laraavci_ – Lara Avci, a Codex ambassador whose Codex-built sites and games got 30K+ visits in four days, now building Tech Grails.
- @bigbrovar – builds useful, real-world tools in public.
That's it for today. 👋