Morning Edition LIVE
Vol. I · No. 1
Est.
MMXXVI

The A.I. Beat

Dispatches from the frontier of machine intelligence
Three
Dollars
← Front page Tools & Releases November 12, 2026 · 5 min read
Tools & Releases

Google Ships Gemini Decisions API for Fast Classification Tasks

Plus a security patch for Simon Willison's LLM tool and Huawei open-sources its quantization library.
Google Ships Gemini Decisions API for Fast Classification Tasks

Google released the Gemini Decisions API yesterday, a new endpoint optimized for classification and decision-making tasks. It’s powered by a model called Gemini Decision, which Google claims is faster and better calibrated for multimodal decision tasks than their general-purpose models.

The API is designed for use cases where you need quick yes/no answers, categorization, or routing decisions. Think content moderation, support ticket triage, or anything involving structured choices rather than long-form generation. Google says the model is “calibrated,” which in this context means its confidence scores should actually map to real-world accuracy (if it says 80% confident, it should be right 80% of the time).

This is Google catching up to what OpenAI has been doing with specialized models for specific tasks. It’s not revolutionary, but if you’re building agent systems or classification pipelines on Google’s infrastructure, it’s worth testing. Pricing and exact availability details weren’t included in the announcement.

Who should care: Anyone building decision systems or classifiers with Gemini already. The speed gains could matter for high-volume applications.

Who can skip: If you’re not on Google’s AI platform or your use case needs reasoning depth over speed.

Security fix in LLM 0.37.2

Simon Willison’s LLM command-line tool got an emergency security patch yesterday. Version 0.37.2 fixes two related issues where your OpenAI API key could accidentally get sent to third-party providers when using OpenAI-compatible endpoints.

The vulnerability affected the extra-openai-models.yaml configuration mechanism (now managed via the llm openai endpoint add command added in 0.37). If you’d configured a third-party endpoint but hadn’t stored the key it referenced, or if certain response conditions were met, LLM would fall back to sending your OpenAI key instead.

Daniel de Araujo reported both issues (tracked as GHSA-v9h7-jj4g-f862 and GHSA-r639-rr7r-x72w). If you use LLM with multiple providers, update immediately.

Huawei open-sources MACaQ quantizer

Huawei’s Compute Systems Lab released MACaQ, the quantization library they use internally for running models locally. Quantization compresses models to use less memory and run faster by reducing the precision of weights (typically from 16-bit to 4-bit or 8-bit).

MACaQ is Huawei’s take on the problem, designed for their hardware but usable elsewhere. The README doesn’t include detailed benchmarks yet, so it’s unclear how it stacks up against established tools like llama.cpp’s quantization or Hugging Face’s Quanto.

Worth watching if you’re deep in the quantization space or working with Huawei hardware. For everyone else, this is more interesting as a signal that major vendors are open-sourcing their inference tooling than as something to drop into production today.

Also shipped

OpenAI published a customer story about Holiday Extras using Codex. It’s a case study showing improved booking conversion and internal tooling, but light on technical specifics. If you’re selling AI tools internally, it’s a decent reference for how companies talk about these deployments.

Simon Willison also added tag-based browsing to his collection of AI-generated tools. The site now automatically analyzes each tool’s code using tree-sitter to extract which JavaScript APIs, libraries, and web services it uses. It’s a neat example of using AI to organize AI-generated code, and the tag clouds make it easier to find tools using specific APIs.

developer tools tools