EN AR

Google launches Gemini Decisions API, pricing output tokens at zero

Sukaina Khalid

Key Points

  1. Google released Gemini Decisions API on November 11, 2026, running a new model, Gemini Decision 1.
  2. The model scores up to 1,000 preset questions in one forward pass instead of generating text.
  3. Pricing is 0.10 dollars per million input tokens and nothing for output, targeting agent routing.

The latest:

A decision layer priced at 0.10 dollars per million input tokens and zero for output went live on November 11, 2026, as Google opened its Gemini Decisions API to developers. The service runs Gemini Decision 1, which Google says returns end-to-end answers in 150 to 250 milliseconds. Google product manager Min Choi and DeepMind chief scientist Felix Yu announced it.

Details:

  • How it works: Instead of decoding text token by token, the model evaluates one shared multimodal input against a predefined question set and scores candidate answers in a single forward pass, according to Google. It accepts text, images, audio, video and PDF files in the same request.
  • The scale: A single request can carry up to 1,000 independent questions over one shared input, Google said. Each returns probabilities plus a confidence score between 0 and 1, defined as the estimated likelihood that the top answer is correct.
  • Question types: Three formats are supported: yes/no, multiple choice with 2 to 255 options, and rubric scoring with 2 to 255 levels, though Google recommends 2 to 10. Question keys are capped at 64 characters.
  • The limits: Context is 262,144 tokens covering input and questions together, with inline input up to 100 megabytes and 50 megabytes for PDFs. External HTTPS links are capped at 15 megabytes per file. YouTube links and gs:// addresses are rejected.
  • Latency claims: Google puts end-to-end response at 150 to 250 milliseconds and median model compute at 50 to 100 milliseconds for inputs up to 16,000 tokens, saying performance stays nearly flat from one question to 100 in a single request.
  • The cost math: The shared input is billed once no matter how many questions ride on it, with each question adding only its instruction and criteria tokens. There are no separate charges for thinking tokens, audio transcription or OCR, Google said.
  • Benchmarks: Google says the model leads its own benchmark suite on agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. Per SiliconANGLE, Google claims it beat OpenAI’s Decisions API and TypeSafe AI’s Jev on 7 of 9 tests. Scores were not published.
  • The competition: Interesting Engineering reported OpenAI launched its own Decisions API in October with GPT-6 Luna, claiming up to 10 times the speed of its Responses API, support for 400 questions per request, and pricing also starting at 0.10 dollars per million input tokens.
  • Internal use: Gemini Notebook uses it to route queries to the right model tier, cutting latency by up to 40% on everyday queries, Google said; Search runs it as a safety layer against adversarial behavior in MCP tools; Labs’ Project CC picks emoji reactions from over 200 options.
  • Availability: It is live through the Gemini API and Google AI Studio with a free testing tier, plus integrations including OpenRouter, LangChain, Vercel AI SDK and LiteLLM. Vertex AI support is described as coming soon, with no date. Rate limits have not been published.

Between the lines:

Google is also claiming cost advantage, saying the model is roughly 5 times cheaper than Gemini 3.5 Flash-Lite via the Interactions API on its own benchmark set. But both headline prices start at the same 0.10 dollars per million input tokens, and neither company has released the underlying scores that would let developers test the comparison independently.

What’s next

Watch for Vertex AI and Google Cloud support, which carries no announced date, and for published rate limits on the free and paid tiers. Independent benchmark results would test Google’s claim of beating OpenAI on 7 of 9 tests.

What to read next