Google has launched a new Gemini Decisions API that answers up to 1,000 closed questions in a single pass, scoring options instead of writing replies, in a bid to speed up AI agents.

Key Points

  • Google launched the Gemini Decisions API on Wednesday, 11 November 2026, powered by a new model called Gemini Decision 1. It is available through the Gemini API in Google AI Studio, with a no-cost tier for prototyping.
  • One request can carry a shared input of up to 256,000 tokens — text, images, audio, video or PDFs — and up to 1,000 independent questions of three kinds: yes-or-no checks, a choice among 2 to 255 options, or a score on an ordered scale.
  • The model scores every candidate option in a single forward pass and returns probabilities plus a confidence score from 0 to 1 for each answer, in 150 to 250 milliseconds end to end.
  • It costs $0.10 per million input tokens, with no charge for output tokens, and the shared input is billed once per request no matter how many questions are asked.
  • Google says the model leads its own benchmark suite and is nearly five times cheaper on that suite than Gemini 3.5 Flash-Lite — company claims, with no numeric scores published and no outside test results yet.

What happened

Google announced the service on its developer blog on Wednesday, in a post by DeepMind product manager Min Choi and principal scientist Felix Yu. The pitch is deliberately narrow: most AI models write answers word by word, but Gemini Decision 1 does not write at all. It takes the options a developer supplies, scores them all at once, and hands back the winner with a probability for every candidate.

That design is aimed at the small, repeated choices inside software. A developer can ask whether a photo shows a spill on a counter, which cleanup action fits the scene, and how severe the mess is — all in one request, against the same photo and note. The same pattern covers tool selection, policy checks, routing a query to the right model, and grading context, Google said. A call recording can be paired with account details, or a scanned insurance claim with photos of the damage, in a single shared input.

Every answer comes with a confidence score, and that is the point of the tool. Code can act on high-confidence answers straight away and send doubtful cases to a fallback system or a human reviewer, instead of treating every machine answer the same way. The three question types have their own names in the API: noul for yes-or-no, choice for picking among supplied alternatives, and score for grading against an ordered scale that starts at zero.

Developers can start now with an API key in Google AI Studio, where a free tier covers prototyping. Python, JavaScript and REST guides are live, along with three demo apps — Doorstep Parcels, Speed Sketch and Find the Zoo Keeper — and integrations are rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Support in Gemini on Google Cloud is listed as coming soon, and there is no context caching, Batch API or conversation history for this model, according to developer notes on the launch.

Google is already using the system inside its own products. Gemini Notebook uses it as a live router, sending everyday queries to lighter models and saving deeper reasoning for hard ones — a change Google says cuts latency by up to 40%. The rapid canvas experiment in Stitch uses it to build playable wireframes in seconds, a Google Labs tool called CC uses it to pick emoji reactions from a palette of more than 200 options, and Google Search uses it as a safety layer that watches for adversarial behaviour in its tool-calling features.

Why it matters

Decision models are a new niche, and Google is the latest big name to enter it. TypeSafe AI introduced its Jev model on 15 September, OpenAI announced its Decisions API on 29 September and made it generally available on 22 October, and Microsoft released Microsoft-Decision-1 on 9 October. The Economic Times, which covered the launch, noted that Google charges the same $0.10 base rate per million input tokens that OpenAI lists for its own Decisions API.

The logic behind the category is simple. AI agents make hundreds of small choices as they work — which tool to call, whether a document passes a rule, which model should handle the next step — and asking a full chat model to write out an answer for each one is slow and wasteful. A model built only to pick from a fixed menu can do the same job in a fraction of the time and, its makers argue, at a fraction of the price. Google puts the end-to-end time at 150 to 250 milliseconds, with 50 to 100 milliseconds of that spent on the model itself, and says the time stays nearly flat from one question to 100 at contexts of up to 16,000 tokens.

But Google’s headline cost claim needs reading carefully. The company says the Decisions API is nearly five times cheaper than Gemini 3.5 Flash-Lite on its own benchmark suite — a comparison against its own cheaper model, on its own tests, with no scores published. Against specialist rivals, the picture looks different. Latent Space, which runs an independent Decision Index, scores Gemini Decision 1 at 66 with a Brier score — a measure of how well its confidence matches reality — of 0.18, at $345 per million tasks. That is more than double the $156 per million tasks it records for Jev, and its average of 211 milliseconds per task is slower than Jev’s 92 milliseconds. No independent accuracy results for the model exist yet, and Google has published no details of its size or architecture.

That does not make the launch unimportant. It makes the test a practical one: teams will have to run the API on their own data, check whether its confidence scores hold up outside Google’s test suite, and work out whether the flat pricing — one bill for the shared input, however many questions ride on it — beats the alternatives on their workload. A thousand questions on one document is a striking offer, but only if the answers are right often enough to trust without a human in the loop.

What happens next

The first thing to watch is whether developers adopt it beyond the demo apps. Google’s pricing and free tier make it cheap to try, and the integrations with LangChain, Vercel and OpenRouter put it in front of the teams already building agent software. If the confidence scores prove reliable in the wild, decision-only models could become a standard layer inside agent stacks — the quiet machinery that picks, routes and checks while bigger models do the talking.

The second is Google Cloud. The API is live in AI Studio, but the larger enterprise customers who run workloads on Google Cloud are still waiting, and Google has given no date. Until that support arrives, the service is a developer tool rather than a platform play.

The third is the price war. OpenAI and Google now charge the same headline rate, and the specialist firms are faster and, on one independent measure, cheaper per task. The next move may come from any of them — a lower price, a wider question limit, or proof from real deployments that their scores can be trusted. For now, Google has put a price on judgment: ten cents per million tokens in, and the answers themselves free. Whether that is cheap depends entirely on how often those answers are right.

Gemini 4 Carbon, Google’s strongest model, opened to paying developers earlier this week, and the two launches show Google pushing on both ends at once — a bigger brain for the hardest jobs, and a smaller, faster judge for the thousands of small choices in between. The same week, United Airlines barred third-party AI agents from booking flights, a reminder that the power to decide is also the power to spend someone else’s money — which is exactly why the confidence score matters. And the Gemini family keeps spreading into the physical world, with Gemini Robotics 2.5 running a humanoid on live Mercedes-Benz factory lines.

Sources: Google blog, 11 November 2026; The Economic Times, 12 November 2026; SiliconANGLE, 11 November 2026; ANI, 12 November 2026; Latent Space, 11 November 2026. Published 12 November 2026.