Google's Gemini Decision 1 Scores Up to 1,000 Questions in One Pass and Bills Zero Output Tokens

Google has opened a new API that is built to answer closed-set questions rather than write text. On November 11, 2026, Product Manager Min Choi and Principal Scientist Felix Yu of Google DeepMind introduced the Gemini Decisions API, powered by a new model called Gemini Decision 1 (gemini-decision-1). It is available through the Gemini API in Google AI Studio.

The design is aimed at the tiny branching calls that sit inside agents: tool selection, policy checks, model routing and context grading. A single request can carry one multimodal input of up to 262,144 tokens and up to 1,000 independent questions. Instead of decoding an answer token by token, Gemini Decision 1 scores the candidate options in one forward pass and returns probabilities and confidence scores. Google quotes 150–250 ms end-to-end latency, with 50–100 ms of median model compute for contexts up to 16K tokens, and says latency stays nearly flat from 1 to 100 questions.

Three question types, one shared input

Every request has a model ID, a shared input, and a map of questions. The input uses the same Content block format as Google’s Interactions API and can mix text, images, audio, video and PDFs. Image and video resolution can be tuned to trade visual detail against token cost. Questions cannot see one another’s answers.

Google documents three types:

• noul (yes/no) returns the probability that a condition holds, bounded between 0.01 and 0.99, plus a confidence score from 0 to 1.
• choice picks one of 2 to 255 options and returns normalized probabilities for every candidate.
• score rates the input on 2 to 255 ordered rubric levels and returns a probability-weighted score with per-level probabilities.

Google suggests routing low-confidence cases to a fallback or a human reviewer. Because questions are scored in parallel against the same prefill, a breakroom photo and a short note can be checked for a spill, a cleanup action and a severity grade in one call.

Where Google puts it on the board

Google’s own benchmark table compares Gemini Decision 1 with TypeSafe Jev 1.14, GPT-6 Luna on OpenAI’s Decisions API, and Gemini 3.5 Flash-Lite on the Interactions API. Refusals and safety blocks count as incorrect. On the public split of JevBench, Google reports 92.2% for Decision 1, against 90.9% for Luna, 90.0% for Jev and 86.6% for Flash-Lite. Decision 1 also leads Google’s table on Amazon ESCI Macro-F1 (58.7%), New Yorker Caption text-only (79.7%), PubMedQA (79.8%), Global-MMLU-Lite non-English (86.0%) and Video-MME without subtitles (69.1%). It trails Jev on Mind2Web Multimodal (86.2% vs 86.8%) and When2Call (79.9% vs 82.0%).

The calibration gap is the clearest one Google highlights. Top-1 expected calibration error on seven shared evals is 0.0250 for Decision 1, versus 0.0494 for Jev, 0.0721 for Luna and 0.2104 for Flash-Lite. On total price for those seven evals, Google quotes $0.88 for Decision 1, cheaper than Luna ($0.93) and Flash-Lite ($4.14) but above Jev ($0.68). Google says Decision 1 is nearly 5× cheaper across that shared suite than Flash-Lite with the Interactions API. All of those figures are company-reported.

Pricing that charges only for input

Paid pricing is $0.10 per million input tokens and $0 for output tokens. The shared input is billed once per request, no matter how many questions are attached; each question adds only its own instruction and criteria tokens. Google says there are no thinking-token, transcription or OCR surcharges. The model page marks Batch API and context caching as not supported, so repeated context is billed in full. A no-cost tier is available for prototyping in AI Studio. The docs list google-genai SDK 3.2 or later, with the Decisions endpoint on v1beta.

Already routing notebooks and picking emoji

Google says teams inside the company are already using the API as a low-latency control plane. Gemini Notebook routes everyday queries to cheaper tiers with what Google claims is up to 40% lower latency. Stitch’s rapid canvas experiment uses it to generate playable wireframes and apply voice or tap edits. CC from Google Labs picks emoji reactions from a palette of more than 200 options. Search uses it as a safety layer against adversarial behavior in MCP tools, within its latency budget.

Ecosystem support is rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. Support in Gemini on Google Cloud is listed as coming soon. Parameter counts are not disclosed. For agent builders, the pitch is narrow on purpose: keep generative models for open-ended work, and put bounded choices through a model that returns numbers your code can branch on without parsing prose.