Google Launches Gemini Decision 1 — a Model That Scores 1,000 Questions in One Pass
Google has launched Gemini Decision 1, a new kind of AI model that does not write answers — it judges them. The model, available through the new Gemini Decisions API, can evaluate up to 1,000 yes/no, choice and rubric questions in a single forward pass, returning a probability and a confidence score for each one instead of generated text. Google announced it on 11 November 2026, and it is aimed squarely at the unglamorous but enormous job of scoring, checking and routing AI output at scale.
A decision model is a model built to decide rather than to generate. Where a chatbot produces a paragraph, a decision model looks at a piece of content — a claim, a document, an image — and answers a set of questions about it: is this true, is this safe, which category does this belong to, how confident am I? Gemini Decision 1 does that for up to a thousand questions at once, in one pass through the model.
What Gemini Decision 1 actually does
The Gemini Decisions API is powered by Gemini Decision 1 (gemini-decision-1) and is available through the Gemini API in Google AI Studio. A developer sends one multimodal input — text, images, audio, video or PDF, up to 256,000 tokens — together with up to 1,000 questions, and gets back probabilities and confidence scores from 0 to 1 for every question. There is no generated text in the response: the output is numbers, not words.
That design matters because a probability is a decision a system can act on: route this message to a human, approve that claim, reject this image, rank these results. Google is pitching the model for evaluation, moderation, classification, ranking and agent verification — the checking layer around generative models.
The latency figures, which are Google's claims, are striking. Google says the API delivers 150–250 milliseconds end-to-end, with 50–100 milliseconds of median model compute, and that latency stays nearly flat from 1 to 100 questions at contexts up to 16,000 tokens. Note the scope of that claim: it covers up to 100 questions at up to 16K tokens. Google does not say the same flat latency holds at the full 1,000-question limit or at the maximum 256K-token context, and this article does not imply that it does...
 |
| A laptop on a wooden desk showing a dark dashboard with blue confidence bars, beside a coffee cup, an open notebook and reading glasses |
Speed, pricing and where you can use it
Pricing is simple and cheap by design: $0.10 per one million input tokens, and $0 for output — because there is no generated text to bill for. Input is billed once per request, regardless of how many questions are attached, which is the detail that makes the economics work: scoring 1,000 questions costs the same input price as scoring one. There is also a no-cost tier in AI Studio for prototyping.
Google claims the model is about five times cheaper than Gemini 3.5 Flash-Lite used with the Interactions API for the same kind of scoring work. That is Google's own comparison, and it is framed against its own smaller model rather than against rivals — worth remembering when the number is quoted.
One limit: the API is available on the Gemini Developer API v1beta only, and Google Cloud support is "coming soon", with no date given. AI Studio developers can use it now; teams running everything through Google Cloud must wait.
Integrations are rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode — the frameworks and platforms developers actually use to build agents and applications. That list is a signal of intent: Google wants this model embedded in the agent stack, not sitting alone in a demo.
Google's benchmarks — and the independent check
Google published its own benchmark table alongside the launch, and on its own numbers Gemini Decision 1 leads. On JevBench it scores 92.2%, against 90.9% for GPT-6 Luna and 90.0% for Jev 1.14. Its calibration error — a measure of how well its confidence scores match reality — is 0.0250, against 0.0494 for Jev. But Google's own table also shows Jev ahead on two benchmarks: Mind2Web Multimodal, where Jev scores 86.8% against 86.2% for Gemini Decision 1, and When2Call, where Jev leads 82.0% to 79.9%. Google's lead is real on the headline benchmark, and it is not a clean sweep.
That is where the independent check comes in, and it tells a different story. The Artificial Analysis Decision Index v2.1 — an independent benchmark of decision models — puts Gemini Decision 1 at 66 (65.98), ninth out of 18 decision models tested. The leader, Drex Max v1.1, scores 72.6. Jev 1.14 sits just ahead of Google's model at 66.23, and the field median is 64. On this independent ranking, Gemini Decision 1 is a solid mid-table performer, not a leader.
| Benchmark | Gemini Decision 1 | Best rival shown |
|---|---|---|
| JevBench (Google's table) | 92.2% | GPT-6 Luna 90.9% |
| Calibration error (Google's table, lower is better) | 0.0250 | Jev 0.0494 |
| Mind2Web Multimodal (Google's table) | 86.2% | Jev 86.8% |
| When2Call (Google's table) | 79.9% | Jev 82.0% |
| Artificial Analysis Decision Index v2.1 (independent) | 66 (65.98), 9th of 18 | Drex Max v1.1, 72.6 |
Neither ranking alone tells the truth: Google's table measures tasks Google chose, on its own terms; the independent index measures a different set across 18 models. A developer choosing a decision model should test both — and above all on their own questions, because a scoring model is only as good as its accuracy on the exact judgements a product needs.
 |
| A developer reads API documentation on a laptop at night in a dark office, headphones around his neck |
Why this launch matters
Decision models are a quiet but fast-growing part of the AI stack. Every agent that acts in the world needs a checker: something that verifies a claim before it is published, scores a response before it is sent, or routes a case to the right human. OpenAI has moved in the same direction with its Decisions API for agents, and the two launches together show the labs converging on the same idea — that the next competitive layer is not just generating better text, but judging it reliably.
Gemini Decision 1 also extends Google's Gemini line in a new direction. Readers can compare it with the flagship systems in our flagship AI model comparison, with Gemini 4 Carbon itself, and with the head-to-head in Gemini 4 Carbon vs Claude Fable 5.2. The decision model is not a successor to those systems; it is a companion — a smaller, cheaper, faster model built to sit beside them and check their work.
The honest assessment is this: Gemini Decision 1 is a useful, well-priced tool for a real job, with strong speed and a genuinely clever billing model. It is not, on the independent evidence available today, the best decision model in the field. For teams that need to score a thousand questions in one pass, that combination — cheap, fast, mid-table accurate — may be exactly the right trade-off. For teams where accuracy is the whole job, the ranking says to test it against Drex and Jev before committing.
 |
| A monitor showing a generic analytics dashboard with revenue, users and conversion figures beside a potted plant |
Key Takeaways
- Google launched the Gemini Decisions API on 11 November 2026, powered by Gemini Decision 1 (gemini-decision-1), via the Gemini API in Google AI Studio.
- It evaluates up to 1,000 yes/no, choice and rubric questions in a single forward pass, returning probabilities and confidence scores (0–1) — no generated text.
- One multimodal input (text, images, audio, video, PDF) up to 256K tokens; Google claims 150–250 ms end-to-end latency, nearly flat for 1–100 questions at up to 16K tokens.
- Pricing: $0.10 per 1M input tokens, $0 output, billed once per request regardless of question count; no-cost tier in AI Studio for prototyping.
- Available on the Gemini Developer API v1beta only; Google Cloud support is "coming soon" with no date. Integrations rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.
- Google's benchmarks lead on JevBench (92.2%) but trail Jev on Mind2Web Multimodal and When2Call; the independent Artificial Analysis Decision Index v2.1 ranks Gemini Decision 1 9th of 18 at 66, behind leader Drex Max v1.1 at 72.6.
Conclusion
Gemini Decision 1 is Google's bet that the next big AI product is not a better writer but a better judge. Scoring up to 1,000 questions in one pass, at $0.10 per million input tokens and in a claimed 150–250 milliseconds, is a genuinely useful capability for anyone building agents, moderation or evaluation pipelines. The launch is real, the pricing is clear, and availability is limited for now to the Developer API while Google Cloud support waits. The open question is accuracy: Google's own benchmarks show a lead on one test and losses on two others, and the independent index places the model ninth of eighteen. The right response is to test it — on your own questions, at your own scale, against the rivals it will actually compete with.
Comments
Post a Comment