Frontline Hotspot
Frontline Hotspot

Gemini Decisions API: Up to 1,000 Judgments, Zero Output Cost

A November 11, 2026 official blog from Google introduces Gemini Decision 1 and the Decisions API: instead of emitting answer text token by token, the model scores candidate options in a single forward pass, evaluating up to 1,000 yes/no, choice, and rubric-scoring questions at once, with no output-token billing. The model ID is gemini-decision-1, and the input side accepts one multimodal input across text, images, audio, video, and PDFs (up to 256K). This piece checks only what the official page says verbatim and what it does not: no benchmarks, no rankings, no integration tutorial, and any unpublished pricing or availability is marked as not disclosed. As of November 12, 2026, qualifiers such as coming soon and rolling out are kept as written.

Published November 13, 20269 min read
<!-- gemini-decisions-api-scoring-hotspot | hotspot | Gemini Decisions API: Up to 1,000 Judgments, Zero Output Cost -->

A content moderation queue holds thousands of items waiting for a verdict. The usual approach is to have a model emit "compliant" or "non-compliant" text one item at a time, with every generated token billed as output; more judgments mean a longer bill, and latency that piles up with each extra request. The official blog Google published on November 11, 2026 describes the opposite arrangement: it does not write answers at all, it scores candidates, and it finishes a whole batch of judgments in one pass while charging nothing on the output side. For developers used to the idea that generation is what you pay for, that contrast is the right starting point for understanding the product line, and also the place where headline summaries most often go wrong.

As of November 12, 2026, the details below are checked verbatim against that official blog post. The product is called Gemini Decision 1, the interface is the Decisions API, the model ID is gemini-decision-1, and it is available through the Gemini API in Google AI Studio. This piece reads what the official blog says and what it leaves unsaid: qualifying phrases are carried over word for word, anything the page does not give is marked as "not mentioned" or "not published," and no conclusion is filled in on Google's behalf. There are no benchmark scores here, no rankings, and no integration tutorial. The site's Gemini 4 Carbon launch piece and introductory pricing window SOP cover the release cadence and pricing window of a generative model, while this piece covers the boundary conditions of a decision-scoring model; the two are not the same product line. The OpenRouter single-task token measurement from the same batch is a different cost ledger, with a measured scope that does not overlap the billing boundary discussed here.

It Scores Candidates in One Pass Instead of Writing Text

The official positioning is blunt: "Rather than generating answer text token by token ... scores your candidate options in a single forward pass," and, in the same vein, "Evaluate up to 1,000 yes/no, choice, and rubric-scoring questions in a fast, single pass." On the input side the page allows "one multimodal input across text, images, audio, video, and PDFs (up to 256K tokens) along with up to 1,000 independent questions," and it states explicitly that this happens "without token-by-token decoding." Read together, the skeleton is clear: one multimodal input, a batch of independent questions attached to it, and a model that does not decode word by word but scores each question in a single forward pass, returning scores rather than stretches of answer text.

The crux is that generation and decision-making have been split into two separate jobs. A generative model writes answer text token by token, so output length directly sets the output bill: the longer it writes, the more it costs. Gemini Decision 1 scores candidate options or judgment conditions in one shot and produces no answer tokens that could be billed. The phrase single forward pass is the backbone of the whole framing: with no token-by-token generation step, there is no output cost that grows linearly with the length of an answer, and no variable thinking-token component that floats with the request. One more detail is easy to misread. The official text places "up to 1,000 independent questions" and "up to 256K tokens" side by side; the first is the number of independent questions a single request can carry, the second is the context ceiling for the multimodal input. They are numbers on two different axes and must not be merged into one figure or substituted for each other. Describing scoring as "generating an answer" simply erases the distinction the blog is trying hardest to draw.

Zero Output-Token Charges, and Shared Input Billed Once

The pricing line reads "$0.10 per 1M input tokens and $0 for output tokens," restated elsewhere as "zero output tokens billed." More notable than the unit price is how shared input is charged: "Your shared input is billed once per request no matter how many questions you evaluate." The same input is charged once whether it carries one judgment or 1,000, and each question adds only its own instruction and criteria tokens. That makes folding a dozen scattered requests into one request loaded with questions meaningful in billing terms, rather than simply multiplying the question count into the bill. The page also lists, verbatim, what is not charged: "There are no output-token charges, variable thinking-token bills, or separate audio transcription and OCR costs," and it mentions "with an API no-cost tier ... for prototyping." The specific quota of that no-cost tier, however, is not mentioned on the page, so no requests-per-minute or requests-per-day figure should be filled in on its behalf. The site's LLM API pricing comparison lays out the usual input-and-output structure of usage-based pricing, which makes it easier to see where zero output-token charges sit on a price sheet.

The cost advantage has to be read with its qualifiers attached: "nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API." Two premises cannot be dropped. The comparison baseline is Google's own "shared benchmark suite," not an independent third-party evaluation, and the comparison target is "Gemini 3.5 Flash-Lite with the Interactions API," not a universal unit price under any calling convention. Strip the qualifiers and the multiplier becomes a claim with no traceable source and no way to be rechecked. By the same token, the unit price of Gemini 3.5 Flash-Lite itself is not given on this page, so it is not written here either. Turning "not mentioned" into "Google announced there is none" would dress up an information gap as a negative finding; it reads as though the company made an explicit statement, when the page carries no such sentence.

Three Question Types, One confidence score, No Threshold Table

The official page divides submittable judgments into three types, with names reproduced as written: noul (Yes/No) estimates the probability that a condition holds; choice (Multiple Choice) selects among candidates with "Selects from 2 to 255 choices"; and score (Ordered Rubric) assigns a score along "an ordered rubric (starting at 0)." These three words are question types, not model names; reading them as three separate models would send the whole interface in the wrong direction. They describe what you ask the model to answer, not which model does the answering.

Every answer carries a confidence score (from 0 to 1). The official page gives it exactly one stated use: "so your code can route ambiguous cases to a fallback or human reviewer," meaning the caller can hand uncertain samples to a fallback path or a human reviewer instead of pretending every score is equally trustworthy. The 0-to-1 range is in the source text. What is not in the source text is any concrete cutoff: the score at which a case should go to a human, or the score at which it can be accepted directly, is not published anywhere on the page, and there is no threshold table to copy. Any specific number presented as that dividing line comes from somewhere other than this post, from selection experience added elsewhere. This piece stays with what the official blog says and leaves unsaid; it does not design a selection checklist or a confidence gate here, and it does not present selection experience from elsewhere as an official conclusion.

150-250 ms Is Not an Unconditional Promise

The latency figures have to be read in layers. The page gives "150-250 ms end-to-end latency" and a "median latency 50-100 ms of model compute." The second figure carries a qualifier that cannot be omitted: "across multimodal inputs and contexts up to 16K tokens, remaining nearly flat from 1 to 100 questions per request." In other words, the 50-100 ms compute latency, and the observation that it stays nearly flat as the question count rises from 1 to 100, hold only under the conditions of contexts up to 16K tokens and 1 to 100 questions per request. "without token-by-token decoding" and "in a single forward pass" have to be understood under those same conditions; they cannot be generalized into a claim that a 256K tokens input with 1,000 judgments also lands in 50-100 ms. 256K tokens is the input ceiling, 16K tokens is the condition under which latency was measured, and conflating the two is a common misreading around this page. The 150-250 ms end-to-end figure and the 50-100 ms compute figure are likewise two separate measures: the former includes request round-trip time, the latter counts only model compute, so they should not be quoted back and forth as one number.

The internal case studies leave the same kind of room. Gemini Notebook is described with "up to 40% lower latency," a ceiling rather than a fixed 40 percent; Stitch is "in seconds (1x speed)"; CC is "more than 200 options" and "a fraction of a second"; and the Search safety layer is described only as detecting adversarial behavior and content against adversarial evals with high efficacy within the latency envelope, with no concrete number attached. These are Google's self-reported internal examples, not performance guarantees that transfer across scenarios, and they should be written with both the self-reported and the case-specific limits intact. Lifting them out as a general latency conclusion turns someone else's internal measurement into your own expected value.

rolling out and coming soon Are Two Different States

The ecosystem line reads: "Ecosystem integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode alongside the google-genai SDKs." The phrase rolling out describes something being phased in, not something already available everywhere, and not a promise that every listed framework can be called reliably. The Cloud line sits further out: "with support coming soon to Gemini in Google Cloud." The phrase coming soon is a promise not yet delivered. Writing either as a fully shipped or generally available state flattens a progressive state and a future state into a completed one. As of November 12, 2026, what can be confirmed verbatim is simply this: integrations are rolling out, and Cloud support is coming soon. Neither should be read as everything being ready, and the page gives no completion date for either.

The Numbers the Official Page Does Not Give

A fair number of figures circulating around this page do not come from the official blog at all. A common example is an integer produced by converting 256K into an exact value, while the source text gives only "up to 256K tokens." The specific quota of the no-cost tier, tiered rate limits, whether context caching or a batch interface is offered, and scores and placements from unverified benchmarks all lack a verbatim line on this page. Another class of error treats noul, choice, and score as model names, or describes scoring as "generating an answer," which points away from the original wording in a different direction. Such claims may not all be wrong in themselves, but they do not belong to this page and should not be retold as official wording.

Two phrasings need to be kept apart here. "Not mentioned" means the page does not write it; it does not mean "Google announced there is none." Converting the former into the latter packages an information gap as a negative conclusion, and it reads as though the company made an explicit statement when the page carries no such sentence. This piece holds to one consistent rule: anything the official blog does not give stays at "not mentioned" or "not published," with no numbers, no promises, and no comparison prices filled in, and no figures borrowed from elsewhere to make the section look complete. The verification method is plain enough: compare sentence by sentence against the release page, keep every qualifier intact, and leave everything else blank.

Take the four boundaries together. It does not generate text; it scores up to 1,000 judgments in a single forward pass. Output tokens are billed at zero and shared input is billed once per request, so the $0.10 figure applies to the input side alone. The confidence score has a 0-to-1 range and a stated use in routing ambiguous cases to a fallback or human reviewer, but no published threshold. 150-250 ms is the end-to-end figure, while the 50-100 ms compute latency holds only under 16K tokens and 1 to 100 questions. Integrations are rolling out, and Cloud is coming soon. These boundaries matter more than a headline multiplier stripped of its qualifiers, because they determine what scale, what pipeline, and what stage of a project this capability can be placed in, rather than letting one tidy number imply an availability the page never promised.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-13

Related

Frontline Hotspot

DeepSeek App APIs: Five Tools, Fees on Top of Tokens

Only what the DeepSeek November 13, 2026 changelog, guides and pricing page actually say about the full suite of application APIs: web search, deep-thinking search, file input, URL reading and model safety are now exposed on the cloud API, while Search, Fetch, OCR, ASR and TTS remain standalone APIs. Tool calls and context assembly happen server-side. Fees are quoted verbatim as tool usage fees in addition to standard token charges; no price conversion, no cross-vendor comparison, no invented discount end date, and the chat-demo repository and MIT license are described as published.

Nov 16, 20269 min read
Frontline Hotspot

Hy Realtime Lands in Yuanbao, but the Cloud API Is Still Coming

Checks the official wording and channel timing of Tencent Hunyuan Hy Realtime, released on November 13, 2026: voice calls in Yuanbao and Tencent Customer Service are already live, while the Tencent Cloud API remains at the official will-soon stage with no public interface documentation, endpoints, or pricing. It stops at the web-search decision, the six song-request dimensions from the TME Co-Design, and the two entry layers of the Yuanbao voice preset and the partial web demo. Full-duplex interaction and the other two items remain on the roadmap. Internal A/B Test and chart figures are not turned into third-party results.

Nov 15, 20269 min read