Gemini Decision 1 Review: Multimodal Decision API Capabilities and Pricing

Gemini Decision 1 launched on November 11, 2026, through Google AI Studio and the Gemini Developer API. The model handles multimodal inputs and returns calibrated probabilities for closed-set questions without generating output tokens. It targets tasks such as routing, policy checks, and rubric scoring rather than open-ended text production.

Gemini Decision 1 Review: Core API Capabilities

The API accepts text, images, audio, video, and PDFs in a single shared input up to 262,144 tokens. Inline base64 files reach 100 MB, while external URLs are limited to 15 MB. Questions support yes/no answers, single-select choices from 2 to 255 options, or ordered rubric scores, with up to 1,000 independent queries evaluated in one forward pass. Results include probabilities between 0.01 and 0.99 for yes/no items plus per-answer confidence scores from 0 to 1.

End-to-end latency stays between 150 and 250 milliseconds, with model compute at 50 to 100 milliseconds for contexts of 16,000 tokens or less. Latency holds steady as question counts rise from one to over 100. The service runs on the v1beta endpoint via the google-genai SDK or REST calls, and a Decisions Playground exists inside Google AI Studio. Vertex AI integration remains listed as coming soon, with no open weights or self-hosting options available.

Google reports top-1 accuracy leading seven of nine benchmarks against comparators including Jev 1.14 and GPT-6 Luna. Scores include 92.2 percent on JevBench, 79.8 percent on PubMedQA, and 71.8 percent on MMAU audio tasks, alongside the lowest calibration error of 0.0250 Top-1 ECE. Internal uses at Google include dynamic routing in Gemini Notebook and safety filters in Search.

Pricing, Rate Limits, and Benchmark Comparisons

Paid pricing charges $0.10 per million input tokens across all modalities, with zero cost for output tokens. The shared input counts once per request regardless of question volume. A free tier in Google AI Studio supports prototyping but notes that data may be used for product improvement. No context caching, Batch API, or priority tiers exist at launch.

Free tier rate limits allow 200 requests per minute, 50,000 tokens per minute, and 10,000 requests per day. Paid Tier 1 raises these to 1,500 requests and 1 million tokens per minute, with limits applied per project. Safety filters remain active, and high-stakes applications require human oversight according to Google documentation.

Independent indices place Gemini Decision 1 around the middle of decision-model rankings, with an Artificial Analysis Decision Index near 66. The model does not support open-ended generation or multi-step reasoning, directing those tasks to the separate Interactions API.

Access Options and Documented Limitations

Developers access the service through the client.decisions.create method or direct REST calls. OpenRouter also lists the model. Free tier users face daily request caps, while paid accounts scale with project limits and no separate OCR or transcription fees apply. Google self-reports all benchmark numbers, and independent verification remains limited at this early stage.

Common user feedback on developer forums notes strong calibration on safety and routing tasks but occasional prompt sensitivity when question sets exceed 500 items. No major outages have been reported since the November 11 launch. The absence of Vertex AI support and context caching currently restricts enterprise-scale deployments compared with established generative models.