Chat models are built to write for people, not to flip yes-or-no switches for software. Every time your app asks a generative LLM for a routing decision, you pay for text generation. On November 11, 2026, Google DeepMind launched a cleaner path: the Gemini Decisions API, powered by Gemini Decision 1 (gemini-decision-1). It reads multimodal inputs and outputs zero text tokens. It returns calibrated probabilities in 150 to 250 milliseconds at $0.10 per million input tokens.
What the Gemini Decisions API Actually Is
Instead of writing a sentence word by word, gemini-decision-1 scores your candidate answers in one fast pass. You send a single shared payload of up to 262,144 tokens. That payload can mix text, JSON, PDFs, images, audio, and video together. In that same request, you can ask up to 1,000 independent questions about the material. Every question uses one of three simple formats:
- Yes/No (noul): Returns the probability that a statement is true, from 0.01 to 0.99, along with a confidence score.
- Multiple Choice (choice): Picks the best fit from 2 to 255 named options and shows the probability for each one.
- Ordered Rubric (score): Grades an input across a custom scale, returning a weighted score and per-level probabilities.
Why It Matters for Builders and ISVs
When output tokens drop to zero, the economics of AI routing change immediately. You never pay for generated text, hidden reasoning tokens, separate OCR passes, or audio transcription. Better yet, Google bills your shared input once per request, whether you ask one question or a thousand.
On Google’s benchmark suite, the Decisions API cost nearly 5x less than Gemini 3.5 Flash-Lite on the Interactions API ($0.88 versus $4.14). Model compute stays between 50 and 100 milliseconds at the median. End-to-end latency also stays nearly flat as you scale from 1 to 100 questions.
Calibration matters just as much as speed when software acts on AI judgments. In Google’s launch tests, Gemini Decision 1 posted a calibration error of 0.0250. That’s roughly half of TypeSafe Jev 1.14’s 0.0494. Because the confidence score closely tracks real-world accuracy, your code doesn’t have to guess. High-confidence items run automatically. Meanwhile, uncertain edge cases route to a human or a heavier reasoning model.
Google already uses this endpoint in production across several products. Gemini Notebook uses it as a live router to pick the right model tier and thinking budget. That routing cuts routine query latency by up to 40%. Meanwhile, Google Search runs it as a fast safety gate to catch adversarial inputs aimed at MCP tools. Stitch Rapid uses it to apply voice and tap wireframe edits almost instantly. Finally, CC from Google Labs uses it to pick an emoji reaction from 200+ options in a split second.
Who It’s For and Who Should Skip It
This API fits any team building AI agents, support triage, or document workflows. Agent builders get fast pre-tool guardrails and intent routing without slowing down the conversation. Contact centers can bundle a call recording, order data, and a customer photo into one request. That lets them route tickets instantly. Claims and trust-and-safety teams can also score huge inspection checklists across PDFs and video clips in one shot.
However, it won’t replace generative models everywhere. It can’t draft an email, summarize a meeting, or pull an arbitrary invoice number out of a file. Questions inside a single call also run in isolation and can’t see each other’s answers. If step two depends on step one, your application code still needs to chain those calls.
How It Compares to Other Decision Models
Google stepped into a category that already has several specialized contenders. Here’s how Gemini Decision 1 compares to the main alternatives:
| Model / API | Modalities | Base Pricing (Input / Output) | Context / Batch Limit | Where It Wins |
|---|---|---|---|---|
| Google Gemini Decision 1 | Text, JSON, Image, PDF, Audio, Video | $0.10 / $0 per 1M | 262K tokens / 1,000 questions | Native audio, video, and PDF bundling; lowest calibration error |
| TypeSafe Jev 1.14 | Text, JSON, Image (opt-in) | $0.042 / $0 per 1M | 64K (262K exp.) | Lower price for short text/JSON routing; experimental bounding boxes |
| OpenAI Decisions (GPT-6 Luna) | Text, JSON, Image, Tool calls, Live voice | $0.10 / $0 per 1M ($0.01 cached) | 400 questions | Prompt caching for static policy docs; per-question image attachments |
| Nace AI Drex Max v1.1 | Text, JSON, Image | $0.03 / $0 per 1M | Varies | Strong text and document scores on Artificial Analysis Decision Index v2.1 |
| Cloudflare Clef-Omni | Audio, Video (+ multimodal) | $0.15 / $0 per 1M | 64K tokens | Open weights available for self-hosted media evaluation |
| Amazon Strands Decider | Text | Self-hosted compute | 4K+ tokens | Apache 2.0 open weights for strict VPC data residency |
Look past the headline benchmarks, and each competitor’s advantage is fairly narrow. TypeSafe Jev 1.14 and Nace AI Drex Max v1.1 cost less for short text routing. Yet neither handles native audio, video, or full PDFs out of the box. OpenAI’s GPT-6 Luna offers prompt caching at $0.01 per million cached tokens for repeated static manuals. Still, it caps batches at 400 questions and skips native audio and video.
Meanwhile, Amazon Strands Decider and Cloudflare Clef-Omni offer open weights if data residency forces you to self-host. In exchange, you take on your own GPU operations and capacity planning. For teams handling real-world multimodal data, Gemini Decision 1 delivers the most complete managed package on the market.
How to Get Started
You can test gemini-decision-1 for free today in Google AI Studio. The no-cost tier gives you 200 requests per minute and 50,000 input tokens per minute. Note that unpaid traffic allows Google to use data for product improvement. Paid API tiers exclude prompts and responses from training, and Google Cloud support is coming soon.
To start building, grab an API key in Google AI Studio and update to google-genai SDK v3.2 or later. You can also call the model through OpenRouter, Vercel AI Gateway, LangChain, LiteLLM, Pydantic AI, Pipecat, or LiveKit. From there, you pass your shared input alongside a dictionary of questions and branch on the returned probabilities.
Before shipping to production, keep three guardrails in place. First, keep hard permissions and spending limits inside your own application code. Decision models can still get tricked by prompt injection. Second, always add an explicit fallback option or confidence threshold to multiple-choice questions. The API forces option probabilities to sum to 1.0 even when nothing fits. Finally, run gemini-decision-1 in shadow mode on live traffic first. That lets you tune your escalation thresholds before it touches customer records.