Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

Gemini Decision 1: a Google model that picks answers instead of writing them

Curated by Alfredo Borrás · Plymouth AI Lab

Google launches a decisions-only Gemini model that scores up to 1,000 questions in one pass for $0.10 per million input tokens; Arena voters grow harsher on models that never changed; a study finds small, interchangeable neuron groups behind answer-flipping 'circuits'; researchers quantize to 4 bits one layer at a time; and two agent tools to try.

The story today

Gemini Decision 1: a Google model that picks answers instead of writing them

Google released the Gemini Decisions API on 11 November, backed by a new model called gemini-decision-1. Instead of generating text, it takes one input (text, images, audio, video or PDFs, up to 256K tokens) and up to 1,000 separate questions about it, and scores the candidate answers in a single pass. Three question types are supported: yes/no, multiple choice with 2 to 255 options, and scoring against an ordered rubric. Every answer comes back with probabilities and a confidence score between 0 and 1.

Pricing is $0.10 per million input tokens with no charge for output, and the shared input is billed once per request however many questions you ask. Google quotes 150 to 250 ms end-to-end latency, staying nearly flat from 1 to 100 questions, and says it is nearly five times cheaper across its benchmark suite than Gemini 3.5 Flash-Lite. Google uses it internally for routing queries between model tiers in Gemini Notebook (up to 40% lower latency, it says), for picking emoji reactions, and as a safety check on MCP tools in Search.

Caveats: the benchmark and calibration claims are Google's own, and the post shows no independent comparison. The model only chooses among options you define, so it replaces classifiers and routers, not a general model.

Why it matters: If an agent of yours makes a cheap generative call just to choose a tool, route a request or grade something, try replacing it with a Decisions call and use the confidence score to send unclear cases to a human or a bigger model.

Sources: Google: Gemini Decisions API, fast, calibrated multimodal decisions

Highlights

  1. Chatbot Arena voters called unchanged models 'both bad' more often every month

    Researchers at Stevens Institute of Technology and UMass Boston studied 1,260,967 blind Chatbot Arena votes from May 2023 to August 2024, limited to model versions that could not change (dated snapshots and fixed open weights). For the same pair of models, the share of 'both are bad' votes rose by 1.71 percentage points a month from an average of 17.1%; in fourteen frequently seen pairs it rose in every one, from 15.5% to 23.3%. Which model wins did not drift. The authors read this as people raising their standard as the field improves, so a product that stands still is judged as getting worse. That matters for leaderboards and for human feedback used in training.

    evalshumansSource: arXiv: AI Isn't Getting Dumber, We're Getting Pickier

  2. Neuron 'circuits' that flip a model's answer are often small and interchangeable

    A single-author paper tests every one of the 4,096 subsets of 12 selected MLP neurons in small Qwen3, Llama-3.2 and Gemma-3 models. On a capital-city task, removing the neurons cut the answer margin by 27%, but swapping in their values from other questions barely changed anything, so the author finds no sign they carry question-specific content. On two-hop riddles, patching the neurons pushed the answer to the intermediate entity in 8 or 9 of about 10 chains per model, while random neurons never did. Often a few neurons were enough, and separate subsets could each flip the answer alone. Effects held for rephrasings but were much weaker on other chains. The lesson: showing that a neuron group changes an answer is not the same as finding the mechanism.

    interpretabilityresearchSource: arXiv: Are Causal Neuron Circuits Unique and Transferable?

  3. MACaQ quantizes weights and activations to 4 bits one layer at a time

    A six-author paper presents MACaQ, a post-training method for MXFP4 4-bit weights and activations that treats each linear layer on its own, so only one layer needs to be in memory while quantizing. It keeps a small transform running at inference time, which the authors measure at 2.3% end-to-end overhead over plain MXFP4 on Qwen3-32B in vLLM. They report lower perplexity than DuQuant++ and MR-GPTQ, matching their best Llama-3-8B result in 11 times less quantization time, and quantizing a 122B-parameter model in 64 minutes with 32.5 GB of GPU memory. These are the authors' own measurements.

    modelsquantisationSource: arXiv: MACaQ: Fast, Model Agnostic Calibrated Quantization

Concept of the day

Calibration

A model is calibrated when its confidence matches how often it is right: of all the answers it gives with 80% confidence, about 80% should be correct. Calibration matters most when software acts on the number, for example sending low-confidence cases to a person, because an overconfident model sends too few of them and an underconfident one wastes reviewers' time. It is usually checked by grouping answers by stated confidence and comparing each group's accuracy.

Source: Google: Gemini Decisions API, fast, calibrated multimodal decisions

Models released

ModelMakerSizeNotes

Try this

  1. n8n Nod: an open-source desktop agent that asks before it acts

    n8n Labs released Nod as a research preview on 10 November: a free, Apache 2.0 app for Mac and Windows that runs in its own local sandbox with a browser, uses the AI provider you choose, and works across Gmail, Calendar, Drive, Slack and plain websites. A separate check confirms you asked before it sends, pays, posts or deletes; it asks again once a task passes $5, and you can set a monthly spending cap. n8n reports 93.2% on the LongMemEval memory benchmark. It stops when your laptop sleeps and has one owner per install.

    agentstry-thisSource: n8n Blog: Meet Nod

  2. Atuin Threads: search and resume agent sessions across tools and machines

    Atuin 18.24 adds Threads, which keeps a synced, end-to-end encrypted history of coding-agent sessions. Run atuin resume to search past sessions and continue one anywhere, including in a different tool, such as starting in Claude Code and finishing in Codex. It supports Claude Code, Codex, Pi, Omp and OpenCode, capture is opt-in, and atuin mcp install lets your agents search earlier sessions themselves.

    harnessestry-thisSource: Atuin blog: Atuin Threads

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI dispatch and route optimization

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call