Google Launches Gemini Decision 1, a Model That Picks Answers Instead of Writing Them

Google announced its Gemini Decision 1 model and dedicated Decisions API on November 11, 2026, according to the official Google blog. Built for fast classification and intent routing rather than text generation, the model (gemini-decision-1) scores predefined options without generating text tokens. Available through the v1beta Gemini API in Google AI Studio, it costs $0.10 per 1 million input tokens with zero output token fees, alongside a free prototyping tier.

Scoring Closed Sets Instead of Generating Text

Unlike models highlighted in our earlier news on Omni Flash that generate words sequentially, Gemini Decision 1 evaluates options in a single forward pass. Per the Google AI Studio docs, it supports a 262,144-token context window and evaluates between 1 and 1,000 independent questions per request. Median compute latency sits at 50 to 100 ms for inputs up to 16K tokens, yielding an end-to-end latency of 150 to 250 ms that stays nearly flat across 1 to 100 questions.

The shared input handles text, images, audio, video, and PDFs, billed once per request. Requests map to three question types:

• noul: Yes/no binary evaluations returning probabilities strictly between 0.01 and 0.99.
• choice: Multiple-choice questions selecting from 2 to 255 options, sorted in UTF-8 byte order.
• score: Ordered rubrics of 2 to 255 levels returning a continuous weighted average.

What It Costs and the Free-Tier Privacy Catch

Pricing is aggressive: $0.10 per 1M input tokens across all modalities, with output tokens permanently at zero. Google says the model is nearly five times cheaper than Gemini 3.5 Flash-Lite on its own benchmark suite, and that it leads that suite. Evaluating numerous questions against a single shared input makes high-density classification remarkably cheap.

However, the free tier carries a privacy trap. Google’s official Gemini API pricing page confirms that free-tier data is used to improve its products; paid-tier data is not. Teams handling customer messages should avoid the free tier entirely. Additionally, context caching and the Batch API are unavailable, so repeated inputs across requests are billed in full each time.

Infrastructure Plumbing Rather Than Conversational Chat

Gemini Decision 1 is strictly not for conversational chat, summaries, or JSON schema extraction. As reported by The Decoder, Google is joining an industry wave of decision models started by startup Jev, followed by OpenAI, Cloudflare, and Microsoft.

End consumers will not interact with the model directly. Google uses it internally for routing in Gemini Notebook, rapid wireframing in Stitch, emoji selection in Google Labs CC, and safety filtering in Google Search. Developers can access it using SDK version 3.2 or later, with integrations rolling out to OpenRouter, LangChain, and LiteLLM, and Google Cloud support coming soon.

Availability in the Arab World and Testing Arabic Dialects

Regional access is broad: Google’s available-regions page lists Saudi Arabia, UAE, Egypt, Qatar, Kuwait, Bahrain, Oman, Jordan, Lebanon, Iraq, Palestine, Morocco, Algeria, Tunisia, Libya, Yemen, Sudan, and Mauritania. Syria is excluded.

Crucially, no Google document lists supported natural languages, leaving Arabic unconfirmed. While question keys require ASCII characters matching [A-Za-z0-9_-], option keys permit UTF-8 strings up to 512 bytes. Arabic text can be passed into inputs and choices, but dialect accuracy remains unverified.

Developers should run a 15-minute test in AI Studio:

• Gather 20 customer messages mixing Modern Standard Arabic, Gulf, Egyptian, and Maghrebi dialects.
• Using our prompt-writing guide, write instructions in Arabic with Latin question keys.
• Create a choice question for intent (complaint, order, refund, or other) and a noul question for angry sentiment.
• Compare model outputs and confidence against human ground truth.
• Use made-up or anonymised messages: free-tier data is used to improve Google’s products.

The Daleel AI Verdict

Gemini Decision 1 is a welcome architectural correction. Paying generative token rates and enduring streaming latency to route tickets has always been inefficient. For regional developers, $0.10 per million tokens with sub-250 ms latency is compelling. But because Google has published no Arabic documentation, teams should test dialect accuracy on their own data, on the paid tier, before trusting it in production.