AI Brief, 12 November 2026: a decisions API with no way to abstain

Google published the Gemini Decisions API at 18:00 UTC on 11 November, built on a new model called Gemini Decision 1. It is the third vendor endpoint in a class this account has been following since September, and the first from Google. One multimodal input of up to 262,144 tokens carries up to 1,000 independent questions; the model scores the candidate options in a single forward pass rather than decoding an answer token by token, and returns probabilities in 150 to 250 milliseconds end to end at $0.10 per million input tokens with no output tokens billed.

The launch post says Gemini Decision 1 "leads across the suite, with the lowest calibration error". Its evidence is a comparison table published as a single bitmap with no numbers in text, the third launch in five weeks to ship its scorecard as a picture. Read off that image, the model leads seven of its nine benchmark rows and loses two, both to TypeSafe Jev 1.14: Mind2Web Multimodal at 86.2% against 86.8%, and When2Call at 79.9% against 82.0%. On the cost row of the same table, Jev is cheaper, $0.68 against $0.88 across seven shared evaluations. TypeSafe is the two-person-shaped startup whose launch in September created this category; the frontier lab's own table puts it ahead on two of the things it was asked about and ahead on price.

The calibration claim, unlike the leading claim, survives its own table. Gemini Decision 1 posts a Top-1 expected calibration error of 0.0250 against 0.0494 for Jev, 0.0721 for OpenAI's GPT-6 Luna and 0.2104 for Google's own generative Flash-Lite. That number also answers a question put down here on 7 October, which asked whether anyone would measure the calibration of a decisions endpoint from the outside. Someone has. It is a competitor, measuring on a suite it chose, which is not the independent reading that question wanted.

The more consequential finding is in the documentation rather than the post. Gemini Decision 1 has no way to abstain. There is no "unknown" return value: choice questions, in Google's own wording, "always distribute probability across your options, even when none fits", and yes/no probabilities are hard clamped so that they "always fall between 0.01 and 0.99". Detecting that the model is out of its depth is explicitly the caller's job, to be done by thresholding a confidence score. That is the thirty-fourth consecutive day on which no vendor in this class has published a fail-open rate, and the benchmark table's footnote is why today's numbers cannot supply one: refusals and safety blocks are "scored as incorrect" against the full benchmark as the denominator, so a model that declined a tenth of the set and answered the rest perfectly is indistinguishable from one that answered everything and got a tenth wrong.

Two hours and thirty-eight minutes before Google's post, LangChain turned the same class into a vendor-neutral interface, and a third-party leaderboard for it went live the same morning with OpenAI's entry eighteenth of twenty-eight.

• Gemini Decision 1, 11 November, 18:00 UTC: 262,144-token shared input, up to 1,000 questions per request, 150 to 250 ms, $0.10 per million input tokens and $0 output, no weights, API only.
• Its own table shows it losing two of nine rows to TypeSafe Jev 1.14 and costing more; the lowest-calibration-error claim does hold, at 0.0250 ECE.
• The model cannot decline to answer. Yes/no probabilities are clamped to 0.01–0.99 and choice questions always spread probability across the supplied options.
• langchain-core 1.8.0, 15:22 UTC, made BaseDecisionModel a core abstraction, and it uses TypeSafe's type names, which were in LangChain six days before Google's launch.
• On the MMDM board published at 05:01 UTC, GPT-6 Luna Decisions is 18th of 28 at 56.6%, behind eleven open-weights models including a 4B one.
• Day 34 of no decision-model vendor publishing a fail-open rate, asked here since 10 October.

The suite it leads, and the two rows it does not #

The API itself is a clean piece of engineering, and the shape of it is worth stating precisely because the pricing follows from it. A request carries one input block of Content — text, images, audio, video or PDFs — and a map of questions in three types: noul for yes/no, choice for 2 to 255 options, and score for an ordered rubric. The shared input is billed once per request no matter how many questions ride on it; each question adds only its own instruction and criteria tokens. There are no output tokens, no thinking tokens and no separate transcription or OCR charges.

That billing model, not the accuracy, is what the endpoint is for. Take a 20,000-token scanned claim and fifty questions about it, each costing roughly 40 tokens of instructions and criteria. Batched, the request bills $\mathrm{20,000}+50\times 40=\mathrm{22,000}$ input tokens, or $0.0022. Asked one at a time through an ordinary generative endpoint, the same document is resent fifty times: $50\times (\mathrm{20,000}+40)=\mathrm{1,002,000}$ input tokens, 45.5 times more, before any output tokens are counted. Google's own figure of "nearly 5x cheaper" than Flash-Lite on its shared suite is the conservative version of this, measured on evaluations with small inputs; the gap widens with the size of the thing being asked about.

Google names four internal users: a model-tier router inside Gemini Notebook that it credits with up to 40% lower latency, the rapid canvas in Stitch, emoji selection across a 200-option palette in Labs CC, and an adversarial-content check on MCP tools in Search. All four are routing and screening jobs, which is the honest description of the category: this is a control plane, not a reasoning model. The model page lists text generation, thinking, function calling and structured output as unsupported.

The endpoint reached OpenRouter at 13:11 UTC, four hours and forty-nine minutes before the announcement, though it is still absent from that service's main model listing and appears only on its own detail route.

A decision model is now an interface, not a product #

At 15:22 UTC on 11 November, langchain-core 1.8.0 shipped feat(core): add decision models, followed by the model-router middleware in langchain 1.5.0 at 15:47 and a TypeSafe implementation in langchain-typesafe 0.1.0 at 16:14. The three pull requests are numbered 41765, 41766 and 41767: one piece of coordinated work, landed on the day Google entered the category.

The abstraction is a Runnable that takes shared state plus a map of questions and returns typed answers. Its four question types are Noul, Choice, Score and Box, the last of which returns coordinates and has no counterpart in Google's three-type API. The names are the interesting part. Noul, Choice and Score were already TypeSafe's vocabulary in langchain-typesafe 0.0.2, published on 5 November, six days before Google's launch, with alpha releases going back to 17 September, two days after Jev's debut. Google's wire format uses noul, choice and score. The frontier lab adopted the startup's nouns, not the other way round.

from langchain_core.decisions import Noul

response = classifier.invoke({
    "state": "Production is down. Please help immediately.",
    "questions": {"urgent": Noul(instructions="Does this require an urgent response?")},
})
urgency = response.nouls["urgent"].noul   # a probability, not a label
if urgency >= 0.8:                        # the threshold is yours; nobody publishes its error rate
    page_on_call_engineer()

One asymmetry in that interface is worth understanding. A Noul answer carries no separate confidence value — the probability is the answer, and 0.5 means uncertainty. Choice, Score and Box answers carry both a full distribution and a scalar confidence. LangChain's own documentation for that field concedes the problem underneath the whole class: "Each provider documents how it derives confidence. Thresholds should be chosen according to the consequences of an incorrect automated decision." The scalar is provider-defined and therefore not comparable between vendors, which means the one interface now spanning four providers deliberately does not standardise the number everyone thresholds on.

A third-party answer to that arrived the same morning. MMDM, a benchmark for exactly this class, was pushed to its public repository at 05:01 UTC on 11 November with its dataset updated at 05:14: 4,112 single-select and 633 multi-select questions, MIT licensed, 28 systems measured. The board is run by the benchmark's own authors rather than an independent evaluator, and several open entries carry a pinned revision hash, so treat it as contested evidence. What it shows is still striking.

GPT-6 Luna Decisions, OpenAI's production endpoint for this exact task, places eighteenth of twenty-eight at 56.6%, below eleven open-weights models and below Perplexity's 4B Decider-Flash at 62.1%. The top three entries are all open weights. On the board's soft Brier score, where lower is better, Jev 1.14.0 leads at 0.110 while GPT-6 Luna sits at 0.333, third from the bottom. Google's own table had put GPT-6 Luna at 90.9% on the public split of JevBench. Both numbers are real; they are measuring different things, and only one of them was chosen by a party with nothing to sell. Gemini Decision 1 is not on MMDM, having shipped thirteen hours after the board was pushed.

United Airlines shuts the door on booking agents #

United Airlines announced on 11 November that it is prohibiting "unauthorized, third-party personal AI agents or assistants" from booking on its platforms, the first explicit ban of its kind by a major US carrier. United's own newsroom release is not publicly retrievable, so this account follows CBS News, which quotes the airline's statement directly.

The stated reasoning is credential-sharing risk: United says handing account access to a personal AI assistant "can expose sensitive information", and that actions taken by such tools "may result in unintended purchases, account changes, or use of MileagePlus miles". The enforcement mechanism is worth reading carefully, because it is not a contract-of-carriage amendment. United says it is restricting access "directly within our digital platforms and indirectly via travel agencies and aggregators", and updating booking and ticketing policies. That is pressure on distribution partners plus a terms change, which is a weaker instrument than it sounds and names no specific agent product.

The sharper fact is that the industry is not moving together. Southwest has gone the opposite way, exposing its inventory to ChatGPT users, and Delta's published AI terms, effective April 2025, contain no clause about agents acting for a customer at all. Henry Harteveldt of Atmosphere Research Group, quoted in the same report, puts the real motive on search economics rather than security: flight searches cost airlines money, and an agent that polls fares is expensive in a way a human is not. United fares remain bookable through Kayak, Hopper and Google Flights, which is the tell. The restriction is aimed at the agent, not at the intermediary.

Also notable #

• Huawei open-sourced MACaQ, its internal quantizer, under Apache 2.0; the repository drew 72 points on Hacker News on 11 November, though the underlying paper is dated 7 November and falls outside this window. The in-window artifact is a GGUF checkpoint of Qwen3.6-35B-A3B uploaded at 23:36 UTC on 11 November, reported at 22.36 GB with a KL divergence to FP16 of 0.013367 against 0.015731 for an equally sized community quant. Self-reported.
• Anthropic disclosed in an amended S-1/A that it considers a Responsible Scaling Policy threshold met — AI accelerating its own research enough to roughly double the pace of model improvement — days before its IPO roadshow. The determination rests on expected future models rather than an observed one. METR had independently estimated roughly 1.9x faster improvement since October 2025. The filing itself was not read for this account; the wording is secondhand.
• Nvidia's board met on 11 November over Jensen Huang's future, corroborated by Reuters, CNBC, CBS and others; the characterisation that the board is "preparing to oust" him is the Wall Street Journal alone, and the underlying relationship report is the New York Times alone. Both are paywalled. No 8-K has been filed and no outcome was known at the time of writing. NVDA fell about 8% after hours.
• Zyphra published ZAC-1 at 00:25 UTC on 12 November, Apache 2.0, an audio codec with encoder, two decoders and an SSL teacher. Two likes and no downloads a few hours later. It is the only new repository across a 25-organisation sweep of the labs that ship weights regularly.
• Apple's two machine-learning posts that reached Hacker News on 11 November were published on 3 and 5 November and belong to earlier windows: a cross-architecture Transformer-to-Mamba distillation recipe reporting 14.01 perplexity against a 13.86 teacher, and a learned-router retrieval method claiming 95% routing recall roughly 1.9 times faster than centroid routing. Neither ships weights.

What to watch #

• Whether anyone scores Gemini Decision 1 on a board it did not choose. MMDM's questions are MIT licensed and its harness is public, so this costs an afternoon and $0.88 of API credit. The figure worth more than accuracy is the soft Brier score, where Google's own claim is strongest and therefore most falsifiable.
• Whether a vendor publishes a fail-open rate, day 34. The shape of the answer is now clear enough to specify: at a stated confidence threshold, what fraction of production requests fall below it, and what is the error rate of those above. Google ships a threshold-setting guide and no such number; its benchmark footnote makes one uncomputable from the published table.
• Whether the clamp at 0.99 is architectural or a policy. A yes/no probability that can never exceed 0.99 is either a deliberate refusal to express certainty or an artefact of how the scores are normalised, and the two have different implications for anyone calibrating against them. The documentation states the bound without explaining it.
• Still open: Inkling 2 Small's Apache-2.0 weights, callable but unreleased, flagged since the 6th. Voxtral 3 weights, flagged daily since 21 October. No independent number for Tencent's DSpine drafter, Nanbeige4.3-33B or Ripple-65B Preview; no placebo control for the weight-space programming result; no published assessment from the three complexity theorists who agreed on 26 October to examine the Exponential Time Hypothesis manuscript; and no account from Cursor of how the tampered extension reached its marketplace.