Gemini Decisions API: 1,000 Questions for $0.10, No Output Charge

Most of the work inside an AI agent is not writing. It is deciding: which tool to call, whether a request breaks a policy, which model should handle the next step, whether an answer is good enough to send. Google's new Gemini Decisions API, announced November 11, 2026, is built for exactly that job. It runs a model called Gemini Decision 1 that never writes a word. Instead it reads your input once and scores up to 1,000 yes-or-no, multiple-choice or rubric questions against it in a single pass, returning probabilities and confidence scores in about a quarter of a second. The price is $0.10 per million input tokens, with no output charge at all, because there is no output to bill. That sounds like a small, cheap utility. It is also the fourth decision model to launch in under two months, which makes it the moment this category stops being a curiosity and becomes a standard part of the toolkit.

Google is late to this race. TypeSafe launched Jev in September, OpenAI unveiled its own Decisions API on September 29, and Microsoft followed with Microsoft-Decision-1, built on Alibaba's Qwen. Google's entry has the widest input range of the four and the same headline price as OpenAI. Whether that is worth more than double Jev's rate depends on how much you trust Google's benchmark chart.

What the Gemini Decisions API actually does

A normal language model answers a question by writing text one token at a time. Gemini Decision 1 does the opposite. You send one shared input, text, images, audio, video or PDFs, up to 256,000 tokens (262,144 to be exact), plus a list of up to 1,000 independent questions. The model scores the fixed options for each question in one forward pass and returns a probability for every candidate, plus a confidence score from 0 to 1. Your code can then branch, sort or route on the response directly, without parsing prose.

Google pitches it for the repetitive judgments inside agents: tool selection, policy checks, model routing and context grading. The shared input is billed once per request no matter how many questions you attach, and each question adds only its own instruction and criteria tokens. Google also says there are no separate charges for transcribing audio or reading text in images, no thinking-token bills and no output-token charges. That billing shape is the whole point: a support ticket checked against forty policy rules costs one read of the ticket, not forty.

The model runs on a separate endpoint, POST /v1beta/decisions, through the Gemini API in Google AI Studio. It is not a chat model with a JSON mode, and it cannot be called through the ordinary generate-content path. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Support in Gemini in Google Cloud is "coming soon", so enterprises that buy through Google Cloud contracts cannot use it yet.

The three question types

Every question falls into one of three types, and the type decides what comes back.

Google's own example checks a photo of a breakroom counter against all three at once: a yes/no question ("Is there an uncleaned liquid spill?"), a choice question ("Which cleanup action fits the scene?", with wipe, dispatch janitorial or ready-for-meeting as options) and a severity rubric from clean to safety hazard. One request, one image, three answers. The Python call is short:

from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
    model="gemini-decision-1",
    input=[{"type": "image", "uri": "https://storage.googleapis.com/generativeai-downloads/images/spill.jpg", "mime_type": "image/jpeg"}],
    questions={"has_spill": {"type": "noul", "instructions": "Is there an uncleaned liquid spill on the counter?"}},
)

Two details matter more than they look. First, each question is scored alone: questions cannot see each other, and latency stays nearly flat as you add them. Second, a choice question always picks something, distributing probability across your options even when none fits. Google's guidance says to include a catch-all option named "other" rather than "none_of_the_above", because the API sorts option keys alphabetically and "none_of_the_above" sorts ahead of most real categories. Know that before you wire a choice question into a payout or refund decision.

What it costs

The price is simple, and Google's own pricing page states it plainly. One rate covers every kind of input, text, image, video, audio and PDF alike.

At November 12, 2026 rates (1 USD = 0.75 GBP / 1.43 CAD), $0.10 per million tokens works out to roughly £0.08 and CA$0.14 per million. Those are approximations, and the dollar figure is the one Google bills.

The shared-input billing is where the savings land. Take a 4,000-token record and twenty checks that add 50 tokens each. As one Decisions API request that is about 5,000 tokens, or $0.0005, which is $0.50 per 1,000 records. Run the same twenty checks as twenty separate calls that each resend the record and you pay for 81,000 tokens per record, or $8.10 per 1,000 records. Same answers, sixteen times the price. At larger scale, one million classifications of 500 tokens each is 500 million tokens, or $50 on the paid tier. That arithmetic is ours, from Google's published rate, and it is the number to carry into a budget.

There is a free tier for prototyping: 200 requests a minute, 50,000 input tokens a minute and 10,000 requests a day. Paid tiers raise those limits sharply, with Tier 1 at 1,500 requests a minute and 50,000 a day, Tier 2 at 5,000 a minute with unlimited daily requests, and Tier 3 at 15,000 a minute. But the free tier has a cost of its own. Google's pricing page says free-tier content is used to improve its products, while paid-tier content is not. Do not prototype with anything sensitive on the free tier.

Two omissions on the pricing page deserve attention. There is no context caching, so a 40,000-token policy document is billed in full on every request that includes it. And there is no Batch API, so there is no half-price overnight sweep of a million historical tickets. Neither discount is on a roadmap Google has published. If your pipeline re-reads a large shared context at volume, those two gaps cost more than the headline rate suggests.

How it compares with the other decision models

Four decision models now compete for the same job, and Google's is the priciest of the text-first group on list price. The comparison below uses list prices for input; output is $0 everywhere a model produces no text, which is the whole point of the category.

Two related tools on this site sit nearby but do different jobs. Hume's Audio Decision API, reviewed here earlier, charges $0.47 per million tokens and works only on audio, while the Milliseconds.ai decision API we covered in October runs $0.042 per million tokens and answers in about 90 milliseconds. Google's model is the only one of the four whose documentation lists audio and video, and its 1,000-question cap is two and a half times OpenAI's 400. What it does not have is a price advantage: at $0.10 it matches OpenAI exactly and costs about 2.4 times Jev and Microsoft-Decision-1.

Google's announcement carries its own benchmark chart, and the figures below are Google's own measurements, read from that chart. Nobody outside Google has repeated them.

Read carefully, the chart is a split decision. Gemini Decision 1 is first on five of the seven tests all four models ran, and ahead on seven of the nine benchmarks in Google's chart overall, with the other two being audio and video tests its rivals did not run. It also has the lowest calibration error by a wide margin: 0.025 against 0.049 for Jev and 0.072 for Luna. Calibration matters, because a model that says 90% should be right about nine times in ten, and that is what lets software trust its confidence scores. But Jev wins two tests, including When2Call, the test about when an agent should use a tool, and Jev comes out cheapest on the seven-test cost comparison, $0.68 against Google's $0.88. Google's headline claim, that the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite on the same suite, checks out on its own table ($4.14 against $0.88, about 4.7 times), but Flash-Lite is a general chat model, not a decision model, so the comparison flatters the new API.

Then there is the independent check, and it is less kind. Artificial Analysis added Gemini Decision 1 to its Decision Index v2.1, where it scores 66, ninth of the eighteen decision models ranked, just above the field's median of 64. It is first on the index's two knowledge and reasoning components, but ninth on the document-decision component that carries 30% of the score, passing only 33% of documents outright where the leader passes 63%. At $345 per million tasks it is the most expensive model in the top half of the board, against $156 for Jev and $143 for Microsoft-Decision-1, and Artificial Analysis ranks it fifteenth of eighteen on cost, calling it "particularly expensive". Its measured latency is 211 milliseconds per task, a touch slower than the field median of 209 and well behind Jev at 92. The honest read is that Google leads on calibration and knowledge, while Jev leads on document work, speed and price.

Where Google already uses it

Google says its own teams were using the Decisions API before launch, and the examples show what the model is for. Gemini Notebook uses it as a real-time router, sending each incoming query to the right model tier and thinking level, which Google says cuts waiting by up to 40% on everyday questions. The Stitch design tool uses it to generate playable wireframes of a product flow in seconds and apply voice or tap edits almost instantly. CC, from Google Labs, uses it to pick an emoji reaction from a palette of more than 200 options in a fraction of a second. And Google Search uses it as a safety layer for MCP tools, detecting adversarial behavior against adversarial evaluations inside the latency envelope.

None of those uses involves writing. All of them are fast, repeated judgments where a wrong answer is cheap and a slow answer is not. The launch arrived a day after Google opened Gemini 4 Carbon to paying developers: Carbon is the heavy reasoning model for hard problems, Decision 1 is the fast gatekeeper that decides which problems reach it.

The limits and catches

Every launch-day model has sharp edges, and Google's documentation lists this one's plainly. The Decisions API runs on the Gemini Developer API only, at API version v1beta, with this model only. There is no Vertex AI listing, and Google Cloud support is "coming soon" with no date. Media goes in through a Files API reference, a public or signed HTTPS link, or inline bytes up to 100 MB per request (50 MB for PDFs); there are no YouTube links and no direct Cloud Storage links. Earlier turns have to be passed as plain text, because there is no conversation history, and questions themselves are text only, so media cannot sit inside a question.

The speed figures carry a condition worth stating. Google's 50 to 100 milliseconds of model compute and 150 to 250 milliseconds end to end were measured on contexts up to 16,000 tokens and from 1 to 100 questions per request. The post gives no figure for larger inputs or for a full 1,000-question request, so the quarter-second promise is verified only well below the maximums.

Google's own guide also warns about failure modes no benchmark captures. Set your thresholds using labeled examples from your own traffic, because a 0.9 from Google's test data is not a 0.9 on yours. User-supplied text can argue for its own classification, so a document that says "this is compliant" can tilt the score. Keep a human or a second check on high-stakes actions such as payouts. Do the arithmetic yourself in code: compare numbers, dates and durations directly rather than asking the model to judge them. Finally, this model is not for writing or reasoning. For generated text, tool calls with arguments or multi-step reasoning, Google points developers to its Interactions API instead.

How to try it

Start in Google AI Studio, where an API key and a no-cost tier are available for prototyping. The Python, JavaScript and REST guides are live in Google's Decisions API documentation, and the SDK calls go to the v1beta endpoint with model name gemini-decision-1. Google also ships three demo apps, Doorstep Parcels, Speed Sketch and Find the Zoo Keeper, which show the three question types working on real inputs before you write any code.

A sensible first test is a job you already do with a chat model today. Take one repetitive judgment, a policy check, a routing decision, a grading rubric, and run it through the Decisions API on the free tier alongside your current approach. Measure three things: the cost per thousand decisions, the end-to-end latency, and how often the confidence score correctly marks the answers you would have disputed. If the model is wrong and confident, raise your threshold or route those cases to a bigger model. If it is right and fast, the saved re-reads are where the money is, not the $0.10 rate.

Common mistakes to avoid

The first mistake is treating a probability as a verdict. A 0.87 is not a yes, it is a bet, and the right threshold depends on what a false positive costs you. A spam filter and a refund approval should not share one.

The second is writing choice questions without a catch-all. A choice question always distributes probability across your options, even when none fits, so a missing "other" option forces the model to pick the least wrong answer and report it with confidence.

The third is prototyping with real customer data on the free tier. The price is zero, but the terms say Google may use that content to improve its products. Use synthetic data until you are on a paid tier.

The fourth is assuming the 150 to 250 millisecond figure holds at 256,000 tokens. Google measured it up to 16,000. Test your own largest input before promising a latency number to anyone.

The fifth is comparing this model to a chatbot on price per token. A decision model and a chat model do different work, and the honest comparison is cost per decision, not cost per token. On that measure Jev is cheaper than Google today, and the independent index says so plainly.

Is the Gemini Decisions API free?

There is a no-cost tier in Google AI Studio for prototyping, capped at 200 requests a minute, 50,000 input tokens a minute and 10,000 requests a day. It is free of charge, but Google says free-tier content may be used to improve its products. The paid tier starts at $0.10 per million input tokens, with no output charge.

How is this different from OpenAI's Decisions API?

Both cost $0.10 per million input tokens and charge nothing for output. Google accepts audio, video and PDFs alongside text and images, caps requests at 1,000 questions instead of 400, and runs on the Gemini API in AI Studio. OpenAI's version takes text, images and JSON only, and is generally available through its own API.

What is Gemini Decision 1?

It is the model behind the Gemini Decisions API, with the model code gemini-decision-1. It scores fixed options in a single forward pass instead of generating text, returning probabilities and confidence scores. It is not a chat model and cannot be called through the ordinary generate-content path.

How accurate is it?

Google's own chart puts it first on five of seven shared benchmarks, with the lowest calibration error of the models compared. Artificial Analysis, the only independent check so far, scores it 66 on its Decision Index, ninth of eighteen, first on knowledge and reasoning components but ninth on the document-decision component that carries the most weight.

Can I use it in Google Cloud?

Not yet. The Decisions API runs on the Gemini Developer API only, at API version v1beta. Google says support in Gemini in Google Cloud is coming soon, with no date given.

What can't it do?

It cannot write text, make tool calls with arguments, or do multi-step reasoning; Google points developers to its Interactions API for those jobs. It has no conversation history, no context caching and no Batch API, and its questions are text only.

Google's Gemini Decisions API is a narrow tool that does one job well: fast, cheap, calibrated judgments at scale. Its price matches OpenAI and undercuts a chat model by a wide margin, its multimodal reach is the widest in the category, and its calibration lead is real. But it is not the cheapest decision model, it is not the fastest, and the independent index ranks it mid-table where document work is concerned. Try it on the free tier with synthetic data, measure cost per decision rather than cost per token, and let your own thresholds decide whether it earns a place in your pipeline. If you already route agents through a decision step, this is the week to test a second vendor.

Related reading on AI Core Updates: OpenAI Unveils Next-Generation Reasoning Models, Milliseconds.ai Review: A $0.04 Decision API That Answers in 90 Milliseconds, Hume Audio Decision API Review: $0.47 Per Million Tokens, Meta Muse Ember: The $0.10 Model for High-Volume AI Agents and Gemini 4 Carbon: Google's $2 Cyber-Defense Model Nobody Can Buy Yet.

Sources:

• Google, "Gemini Decisions API: Fast, calibrated multimodal decisions," November 11, 2026 — blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api
• Google, Gemini Developer API pricing (Gemini Decision section) — aistudio.google.com/docs/pricing
• MadRobot, "Gemini Decision 1: Google's AI that answers without writing," November 11, 2026 — madrobot.blog/2026/11/11/google-gemini-decisions-api-1000-questions
• CellCog, "Gemini Decision 1: Price, Benchmarks and Limits" — cellcog.ai/blog/gemini-decision-1
• OrcaRouter, "Gemini Decision 1: Ninth of 18, First Where It Matters Least" (Artificial Analysis Decision Index v2.1) — orcarouter.ai/blog/gemini-decision-1-release-decision-index
• Exchange rates, November 12, 2026: 1 USD = 0.75 GBP (Finnhub) / 1.43 CAD (mid-market). Conversions are approximations.