Summary:

Gemini Decision 1 brings typed multimodal decisions to Google AI Studio. A four-model scorecard separates task accuracy, calibration and cost.

Google has launched Gemini Decision 1, a model for choosing among predefined options rather than writing an answer. Its new Decisions API accepts text, images, audio, video and PDFs, making it relevant to application routing and structured checks.

What is available now

The November 11 launch is available through the Gemini Developer API and Google AI Studio. The model ID is gemini-decision-1. Google lists a no-cost prototyping tier and pricing of $0.10 per million input tokens, with no output-token charge. Google Cloud support is described as coming soon.

A request can contain up to 1,000 independent questions and 262,144 tokens across shared input and questions. Questions do not see one another's answers. The API supports yes/no probabilities, a choice among 2–255 options, or a score against an ordered rubric. It returns confidence information alongside those results.

How it compares in Google's tests

The scorecard below reproduces the numbers in Google's November 11 launch table in a new layout. It compares Gemini Decision 1 through Decisions API, TypeSafe Jev 1.14, GPT-6 Luna through OpenAI Decisions API, and Gemini 3.5 Flash-Lite through Interactions API at its minimal setting.

Google-reported comparison across nine task benchmarks, calibration error and cost for Gemini Decision 1, TypeSafe Jev 1.14, GPT-6 Luna and Gemini 3.5 Flash-Lite. Each row's best numeric result is bold.

Higher is better for the task scores; lower is better for calibration error and cost. Most task rows report top-1 accuracy, while Amazon ESCI reports Macro-F1. Google counts refusals, safety blocks and requests a model cannot accept as incorrect, using the full benchmark as denominator. Calibration and cost cover seven shared evaluations; the cost figures are suite totals, not per-request prices. Missing audio/video results are left unranked.

Gemini leads the reported JevBench result, but Jev leads Mind2Web Multimodal and When2Call and has the lowest suite cost. That makes the task mix important when selecting a model. These are Google's developer-run results with the named endpoint configurations, not a universal ranking across deployments.

Build a routing decision, not an open-ended answer

A useful first pilot is a support queue selector: provide a customer message and a fixed set of queues, then compare the selected queue against previously reviewed cases. Keep a separate fallback for ambiguous or out-of-scope requests rather than forcing every case into one of the business queues.

Define each question narrowly. If one result should change which question runs next, make that dependency explicit in application code or a later request. Parallel questions in a single call cannot perform that branching themselves.

Measure end-to-end latency on representative inputs. Google's 150–250 ms latency claim covers its reported workload, including contexts up to 16K tokens, and should not be projected onto the maximum context window.

Limits worth checking before deployment

The documentation specifies the v1beta API and Google Gen AI SDK 3.2 or later. Question instructions are text-only. Direct YouTube URLs, conversation-history blocks and tool-call blocks are not supported; multi-turn context must be serialized as text. Context caching, Batch API, and Flex or Priority inference are also unavailable for this endpoint, so input repeated across separate requests is billed again.

For a routing pilot, track mistaken routes, fallback frequency, latency and actual input cost together. Test the confidence threshold against reviewed examples before allowing the result to trigger consequential actions.

Sources