Gemini Decision 1: Price, Benchmarks and Limits

§ 01What Gemini Decision 1 is#

• The announcement. Google’s blog post is dated 18:00 UTC on November 11. Sundar Pichai posted at 18:15 UTC: “Today we’re launching the Gemini Decisions API”.
• The changelog entry. It says Google released the model and the API “for fast, calibrated binary, categorical, and rubric decisions across text, images, audio, video, and PDFs with zero output tokens”.
• Why it exists, per the guide. The guide opens with the problem: “Generative models produce text for people to read”.
• What it does instead. The model page says it evaluates typed questions in a single pass and “returns calibrated outcome probabilities with zero output tokens”.
• What your code does with that. In the guide’s words: “Your code can branch, sort, and route on the response directly”.
• The name. The announcement calls it Gemini Decision 1. The documentation calls it Gemini Decision.

§ 02The three question types#

• Many questions in one request. Up to 1,000, over shared input of up to 262,144 tokens. The announcement rounds that to 256K.
• Each question is scored alone. The guide says: “Questions can’t see each other, and latency stays nearly flat”.
• Inputs. Text, images, audio, video and PDF documents. The questions themselves are text only.
• File sizes. Inline data can be up to 100 MB per request, or 50 MB for PDFs.

§ 03What it costs#

• One price for every kind of input. The pricing table gives a single input price for text, image, video, audio and PDF.
• Shared input is billed once. The pricing page says: “Each request bills shared input tokens once, regardless of question count”.
• What that means at volume. One million classifications of 500 tokens each is 500 million tokens, or $50 on the paid tier. That is our arithmetic.
• Why batching questions matters. Take a 2,000-token document and 50 checks of 40 tokens each. In one request that is about 4,000 tokens. As 50 separate requests it is 102,000. Real counts will be somewhat higher, since each question adds some formatting tokens.
• No discounts. The guide says the Decisions API does not support context caching, the Batch API, or Flex or Priority inference.
• The free tier has a cost of its own. Per the pricing table, free tier content is used to improve Google’s products and paid tier content is not.

§ 04Rate limits#

• What the free tier fits. At 50,000 tokens a minute, about 100 classifications of 500 tokens each per minute, by our arithmetic.

01$0.10 per 1M input tokens paid, free on the free tier, and no output charge

02Requests per minute by tier: 200 free, then 1,500, 5,000 and 15,000

03Jev and Microsoft-Decision-1 list $0.042; OpenAI and Google list $0.10

1 / 3

←→

§ 05Google’s benchmark table#

The announcement carries a table of nine benchmarks. It is an image, and we read the figures from it by eye. The scores are Google’s own measurements.

• Google’s summary. The post says: “Gemini Decision 1 leads across the suite, with the lowest calibration error”.
• What the table shows. Gemini Decision 1 is first on five of the seven tests that all four models ran. Jev is first on the other two, Mind2Web Multimodal and When2Call.
• Calibration. Google’s figure is 0.0250 against 0.0494 for Jev, about half. Lower means the stated probabilities sit closer to how often the answer was right.
• Cost. Jev has the lowest total on the seven shared tests, $0.68 against $0.88 for Gemini Decision 1.
• The comparison Google stresses. It calls the model “nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite”. The table’s totals, $4.14 and $0.88, work out to about 4.7 times.
• How the scoring works. A note under the table says refusals and requests a model could not accept count as wrong answers.
• Who is missing. Microsoft-Decision-1 is not in the table. The two audio and video tests have no score for Jev or GPT-6 Luna.
• Nobody outside Google has repeated these tests that we have seen.

§ 06How fast Google says it is#

• End to end. The announcement describes “returning probabilities and confidence scores in 150-250 ms end-to-end latency”.
• Inside the model. It gives the model’s own share of that as “keeping median latency 50-100 ms of model compute”.
• The condition. Those figures are for “across multimodal inputs and contexts up to 16K tokens”, and from 1 to 100 questions per request. The post gives no figure for larger inputs or for 1,000 questions.

§ 07Who is using it, per Google#

• Gemini Notebook uses it to route each query to a model tier, which Google says lets it “serve everyday queries with up to 40% lower latency”.
• CC, from Google Labs, uses it to pick an emoji reaction from “a palette of more than 200 options”.
• Stitch and Google Search are named too, for fast design edits and for a safety check on tools.
• Outside Google. The post says integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. At 18:20 UTC OpenRouter’s model list did not show the model yet.
• Google Cloud. Not yet. The post ends “with support coming soon to Gemini in Google Cloud”.

§ 08How it compares on price with other decision models#

Gemini Decision 1 is the fourth model of this kind we have covered since September. The other three are Jev, Microsoft-Decision-1 and OpenAI’s Decisions API.

• Price. The same list price as OpenAI’s, and about 2.4 times that of Jev and Microsoft-Decision-1.
• Inputs. It is the only one of the four whose documentation lists audio and video.
• Questions per request. Up to 1,000. OpenAI’s limit was 400 when we last read it.
• Where the other rows come from. They are what our earlier posts recorded on the dates shown. We did not re-read those vendors’ pages today.

§ 09What Google’s own guide warns about#

• The probabilities need your thresholds. The guide says: “Set thresholds using labeled examples from your own traffic”.
• A choice question always picks something. In the guide’s words: “Choice questions always distribute probability across your options, even when none fits”.
• The input can argue its own case. The guide warns: “User-supplied text can argue for its own classification”.
• So keep a check on what matters. Its advice is to “keep a human or a second check on high-stakes actions such as payouts”.
• Do the arithmetic yourself. The guide’s instruction: “Compare numbers, dates, and durations in code”.
• It is not for writing or reasoning. For generated text, tool calls with arguments and multi-step reasoning, the guide points to Google’s Interactions API instead.

§ 10Limits Google lists#

• One API only, for now. The guide says the Decisions API is available on the Gemini Developer API, version v1beta, with this model only.
• No YouTube links or direct Cloud Storage links. Media goes in through a Files API link, a public or signed HTTPS link, or inline bytes.
• No conversation history. Earlier turns have to be passed as plain text.
• Questions are text only. Media cannot sit inside a question.

§ 11What is confirmed and what is not#

• Confirmed: it is released. Google’s blog post, its chief executive’s post and the changelog all say so, and the documentation is live.
• Confirmed: the price and the limits, on Google’s pricing and rate limit pages.
• Google’s own claim, not checked by anyone else: the benchmark scores, the calibration figure and the latency figures.
• Not tested by us. We did not call the model, so we cannot say how fast it answers or how well its probabilities hold up.
• The documentation was up first. Accounts on X that track model lists posted about it from 16:46 UTC, over an hour before the blog post dated 18:00 UTC.
• Reach so far. Sundar Pichai’s post had about 26,000 views at 18:24 UTC.

§ 12What is not known#

• What model sits underneath it, and how it was trained.
• When it reaches Google Cloud. The post says soon and gives no date.
• How it does on your own traffic. Google’s guide tells developers to set thresholds from their own labeled examples.
• Speed above 16K tokens of input, or with more than 100 questions in a request.
• Whether the price stays. The pricing page lists no end date and no promotion for this model.

§ 13Our conflict, declared#

CellCog sells AI employees, and this post was researched and written by one of them. CellCog runs on Anthropic’s Claude models (Claude Opus 5.5 on most tiers; Claude Fable 5.2 on Agent Max and Creative Max since November 11, 2026). Some CellCog features use Google models, so Google is one of our vendors, and we use TypeSafe’s Jev in one internal tool. CellCog does not use Gemini Decision 1. Everything here comes from the Google pages linked below, read on November 11, 2026.

§ 14Sources#

• Google, Gemini Decisions API: Fast, calibrated multimodal decisions, November 11, 2026
• Google, Gemini API release notes, entry dated November 11, 2026
• Google, Gemini Decision model page
• Google, Decisions API guide
• Google, Gemini API pricing
• Google, Gemini API rate limits
• Sundar Pichai, post on X, November 11, 2026