AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass

Collected Nov 11, 2026

Google DeepMind has introduced the Gemini Decisions API, an interface aimed at fast, closed-ended decisions such as yes/no answers, multiple choice and ratings on a scale. The model behind it, Gemini Decision 1, handles up to 1,000 independent questions in a single pass and spans text, images, audio, video and PDFs. It scores options directly and returns probabilities together with confidence values, with response times of 150 to 250 milliseconds.

In an example Google described, a single pass over a photo of a kitchen counter combines all three question types: yes/no stain detection at 0.95 confidence, selection of the right cleaning action at 98 percent, and a severity rating of 1.19 out of 2 on a scale.

Pricing is $0.10 per million input tokens with no charge for output tokens. Google says that is roughly five times cheaper than Gemini 3.5 Flash-Lite. The API is available now through Google AI Studio, and integrations with OpenRouter, LangChain, Vercel and other platforms are rolling out.

Google says it already runs the API internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Google Search. For the benchmarks, Gemini Decision 1 reportedly leads the suite in categories including agent routing, reasoning and multimodal understanding, while posting the lowest calibration error on Top-1 ECE.

The launch puts Google into a cluster of recent releases. After the startup Jev introduced its version, OpenAI, Cloudflare and Microsoft quickly followed with similar offerings.

A few technical points help explain what the API is doing. Traditional generative models produce free text, so a developer who needs a yes/no or a pick-one answer typically prompts the model, parses the output and hopes the wording holds. A decision model instead returns scores over a fixed set of options, which makes the result machine-readable and, in principle, easier to log and threshold. Confidence values and calibration metrics matter here: Top-1 ECE measures how closely stated confidence tracks actual accuracy, and for routing or safety decisions a well-calibrated 0.95 is more useful than a confident-sounding sentence.

Batching many questions into one pass is the other design choice. Evaluating the same image against many independent questions at once avoids repeated encoding of the input, which plausibly explains the low millisecond latency and the per-token pricing. This is my own reading of how the pieces fit, not something the announcement states.

The trade-off is that the API is narrow by design. It suits closed-ended classification, routing and rating, not open-ended generation, so teams will likely pair it with a general model. The five-times-cheaper claim is relative to one specific comparison model and depends on input volume, so real savings will vary by workload.

Why it matters: Teams building agent routers, moderation gates, tagging pipelines or visual inspection tools get a cheaper, faster primitive that returns probabilities rather than prose, which should simplify plumbing and thresholding. The crowded field, with Jev, OpenAI, Cloudflare and Microsoft already shipping related offerings, suggests decision-specific models are becoming a standard layer in AI stacks, and Google's willingness to use the API in its own products is a signal of confidence that developers can weigh against their own evaluation.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Google Deepmind's new Gemini Decisions API evaluates up to 1,000 closed-ended questions in a single pass across text, images, audio, and video. Gemini Decision 1 returns results in 150 to 250 milliseconds at roughly five times lower cost than Gemini 3.5 Flash-Lite, according to Google. The article Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass appeared first on The Decoder .