Google just shipped an AI model that doesn't write anything. The new Gemini Decisions API, live today in Google AI Studio, answers yes/no, multiple-choice and scoring questions by ranking the options in a single pass, and it charges nothing for output.

It's a small launch with a big idea. Much of what an AI agent does all day is pick a tool, check a rule or route a request, and paying a chatbot to write a sentence for every one of those choices is slow and expensive.

Also tonight: ElevenLabs makes Scribe v3 its default speech-to-text model, Alibaba and Baidu both open previews of new flagship models, and four free tools worth trying.

Gemini Decisions is a new endpoint in the Gemini API built for closed-set choices: yes/no checks, multiple choice with 2 to 255 options, and rubric scores. Instead of generating an answer word by word, the new gemini-decision-1 model scores every candidate option in a single forward pass and returns each answer with a confidence score from 0 to 1. It accepts text, images, audio, video and PDFs up to 256K tokens, and you can ask up to 1,000 questions about the same input in one request.

Google's numbers: 150 to 250 milliseconds end to end, with latency staying nearly flat from 1 to 100 questions per request, and nearly 5x cheaper than Gemini 3.5 Flash-Lite across its shared benchmark suite. Pricing is $0.10 per million input tokens and $0 for output, and the shared input is billed once per request no matter how many questions you ask about it. There's a no-cost tier in Google AI Studio for prototyping, and integrations are rolling out with OpenRouter, LangChain, Vercel, LiteLLM and more.

Our take: this is plumbing, but it's the plumbing every agent builder overpays for. Routing a request to the right model, checking whether a reply breaks a policy, or grading whether retrieved context is relevant are all multiple-choice questions, and today most teams answer them by asking a full chatbot to write "yes." If Google's latency holds up, you can put dozens of these checks in front of every agent step without users noticing. Two caveats: every number here is Google's own, and it only handles questions with a fixed set of answers, so it sits next to a generative model rather than replacing one.

Try it: Playground in AI Studio · Docs and quickstarts

Think about what a coding or support agent does between the moments you see it write something: decide which tool to call, decide whether a result is good enough, decide whether a request is allowed, decide which model should handle the next step. Google is betting those micro-decisions are a separate workload that deserves its own model, and it's pricing them like one, with no output charge at all.

Google is already using it in-house: its Gemini Notebook team reports up to 40% lower latency on everyday queries after moving routing to Decisions. The bigger signal is agent economics. As labs push models that run for hours and call tools thousands of times, the cost of each small judgment adds up, and cheap, specialized "judge" models are turning into their own product category. Expect rivals to answer with similar endpoints. For now it's in the Gemini API and AI Studio, with support for Gemini in Google Cloud listed as coming soon.

🎤 ElevenLabs Ships Scribe v3 Speech-To-Text

ElevenLabs released two new transcription models on November 10. Scribe v3 handles recorded audio across 102 languages, and Scribe v3 Live handles real-time audio and, for the first time, labels who is speaking as the conversation happens. Both are now the default in the ElevenLabs app, and developers can switch over in the API by changing the model ID to scribe_v3 or scribe_v3_live.

On ElevenLabs' own tests, Scribe v3 makes 27% fewer word errors than v2 on the multilingual FLEURS benchmark, and v3 Live cuts word errors by 31% versus v2 Realtime. In real-time English, its speaker-labeling error rate was 3.1%, against 5.9% to 7.3% for the other real-time models it tested. Keyterm prompting now takes up to 1,000 terms for product names and jargon, and a medical-tuned Live version is rolling out for healthcare teams.

Our take: live speaker labels are the upgrade that matters for meeting bots and voice agents, which until now had to work out who said what after the fact. The announcement gives no latency figures or pricing changes, and every result is self-reported, so test it on your own noisy audio before you switch.

Try it: ElevenLabs app · API docs

Two of China's biggest labs opened early versions of their next models on the same day. Alibaba's Qwen4-Max-Preview is the first Qwen4 flagship developers can call: a multimodal model for coding, office work, front-end design and agents, with a 1M-token context window, thinking switched on by default, and compatibility with the OpenAI and Anthropic API formats. The catch: it's API-only through Alibaba's Token Plan, it shipped with no published benchmarks, and it runs frozen for a two-week window that ends November 25, when it retires.

Baidu's ERNIE 5.5 Flash Preview is the more builder-friendly of the two on paper. It's a mixture-of-experts model with 280 billion total parameters and 14 billion active per token, a 1M-token context window, and training aimed at coding, 3D modeling, knowledge work and video creation. The API is live on Baidu's Qianfan platform now, and Baidu says the weights will be open-sourced under Apache 2.0 "soon," with no date given.

Our take: neither is production material yet, and both labs say as much. Qwen4-Max-Preview is a two-week test drive, and ERNIE 5.5 Flash Preview is an intermediate checkpoint. The thing to watch is Baidu's Apache 2.0 promise: an open model with only 14B active parameters and a 1M context would be a cheap self-hosting option if it lands.

Try it: Qwen4-Max-Preview docs · Qwen Token Plan · ERNIE 5.5 Flash API on Qianfan

The other stories worth your attention:

Anthropic set the price range for its IPO and launched its roadshow for a Nasdaq listing under the ticker ANT. See how many shares it's selling and at what price. (Anthropic)

Gemini Robotics 2.5 is rolling out to robot makers, after a pilot on Mercedes-Benz production lines with Apptronik's humanoid. See the success rate it hit packing real car parts. (Google DeepMind)

AlphaGenome 2 reads DNA one cell type at a time, and DeepMind says it catches far more disease-linked variants than the previous best model. Researchers can use it now through the API. (Google DeepMind)

Microsoft researchers backdoored AI agents by editing a handful of neurons, with a trigger as simple as a date change, and the tampered agents still passed safety evals. Here's how few neurons it took and how to defend. (Microsoft Research)

United became the first US airline to block personal AI agents from booking its flights, citing account security and federal safety rules. See why a hazmat checkbox is part of the problem. (Skift)

Google launched the Gemini Decisions API, a model that scores options instead of writing answers, at $0.10 per million input tokens with free output and a no-cost tier in AI Studio. ElevenLabs made Scribe v3 its default transcription model with live speaker labels, Alibaba and Baidu opened previews of Qwen4-Max and ERNIE 5.5 Flash, and Anthropic put a price range on its IPO. The thread: as agents take on more work, the race is moving to the small, fast pieces that make them cheap enough to run all day.

Stay building. 🤖

Our newest video covers Claude Fable 5.2, the hidden Fable 6 hints, Gemini Krypton, GPT-6.2 and Sol.