Tech Meridian ← LIVE FEED
PROMY MERIDIAN RU

RELEASE · MODELS · #3449

Google launches Gemini Decisions API powered by Gemini Decision 1

Google announced the Gemini Decisions API, a new endpoint powered by the gemini-decision-1 model that scores closed-set choices (yes/no, multiple choice, rubric) across multimodal inputs (text, images, audio, video, PDFs) and up to 1,000 questions in a single request. The company says the API returns calibrated probabilities and confidence scores with end-to-end latency around 150–250 ms, charges $0.10 per 1M input tokens with no output-token billing, and is integrated with Google AI Studio and several ecosystem partners.

KEY POINTS

  1. Google announced the Gemini Decisions API, a new endpoint powered by the gemini-decision-1 model that scores closed-set choices (yes/no, multiple choice, rubric) across multimodal inputs (text, images, audio, video, PDFs) and up to 1,000 questions in a single request.
  2. The company says the API returns calibrated probabilities and confidence scores with end-to-end latency around 150–250 ms, charges $0.10 per 1M input tokens with no output-token billing, and is integrated with Google AI Studio and several ecosystem partners.
  3. This introduces a purpose-built, low-latency decision primitive for multimodal closed-set tasks that Google positions as cheaper and better-calibrated than prior Gemini generative calls, enabling real-time routing, safety checks, and UI features.
MERIDIAN INTELLIGENCE

DECISION BRIEF

75/100
CONFIDENCE
01

WHAT CHANGED

Google launched the Gemini Decisions API, a new endpoint powered by the gemini-decision-1 model that scores closed-set choices (yes/no, multiple choice, ordered rubric) across multimodal inputs and up to 1,000 independent questions in a single request.

02

WHY NOW

This is a purpose-built, low-latency decision primitive for bounded choice tasks: Google positions it as returning calibrated probabilities and confidence scores in a single forward pass (not token-by-token), at end-to-end latency of ~150–250 ms and lower per-input cost than prior Gemini generative calls. Google frames it for uses such as tool selection, model routing, policy/safety checks, and grading/context scoring.

03

WHO IS AFFECTED

Directly affected: developers and teams building closed-set decision flows over multimodal data (text, images, audio, video, PDFs); users of Google AI Studio/Gemini API; tool and integration partners (articles note integrations rolling out with OpenRouter, LangChain, Vercel); Google products that Google reportedly already uses the API in (Gemini Notebook, Stitch, Google Search).

04

CONFIRMED

- Google announced the Gemini Decisions API powered by the gemini-decision-1 model. (4063) - The API evaluates up to 1,000 independent yes/no, multiple-choice, and rubric-scoring questions in a single request across multimodal inputs (text, images, audio, video, PDFs). (4063, 4074) - The model scores candidate options in a single forward pass (not token-by-token) and returns probabilities and confidence scores for answers. (4063) - Google states end-to-end latency of about 150–250 ms and median model compute latency of ~50–100 ms. (4063) - Google states pricing at $0.10 per 1M input tokens and that output tokens are not billed. (4063, 4074) - Google claims Gemini Decision 1 leads on a benchmark suite for agent routing, reasoning, ranking, and multimodal understanding and has the lowest calibration error (per Google). (4063) - The API is available through the Gemini API in Google AI Studio; third-party integrations are noted as rolling out (OpenRouter, LangChain, Vercel per the media write-up). (4063, 4074) - The Decoder reports Google says it uses the API internally for dynamic model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Google Search. (4074)

05

UNCERTAIN

- Token/context limits: one excerpt states a multimodal input limit of up to 256K tokens while another part of the Google excerpt refers to contexts up to 16K tokens; the sources conflict on exact context/token limits and do not reconcile them. (4063) - Real-world generalization of Google’s benchmark claims (latency, calibration, accuracy) beyond Google’s reported suite is not independently verified in these sources. (4063, 4074) - Exact availability scope and timing for third-party integrations (OpenRouter, LangChain, Vercel) and broader partner rollouts are described as "rolling out" but lack specific timelines or coverage details. (4074) - Operational behaviors under high question counts per request beyond the described 1–100 question flatness claim, and performance variability across different modalities or custom inputs, are not shown by independent measurements in the supplied excerpts. (4063)

06

WHAT TO WATCH

Observable signals to monitor: - Official Google AI Studio/Gemini documentation updates clarifying token/context limits and exact per-request constraints. (4063) - SDK, API changelog, or pricing pages confirming billing model and any regional or tiered pricing changes. (4063, 4074) - Announcements or release notes from OpenRouter, LangChain, Vercel (and other partners) showing integration availability, SDK examples, or usage limits. (4074) - Independent benchmarks and third-party tests reporting latency, calibration, and accuracy across modalities and workloads. (4063) - Google product updates or blog posts showing broader internal adoption beyond the initial examples (Gemini Notebook, Stitch, Search). (4074) - Community examples, case studies, or issue reports revealing behavior at high question counts or with large multimodal inputs. (4063)

WHY IT MATTERS

This introduces a purpose-built, low-latency decision primitive for multimodal closed-set tasks that Google positions as cheaper and better-calibrated than prior Gemini generative calls, enabling real-time routing, safety checks, and UI features.

EVIDENCE MAP

4

Editorial claims linked to specific sources, with support, contradiction and context shown separately.

The API can evaluate up to 1,000 yes/no, multiple-choice, and rubric-scoring questions in a single forward pass across multimodal inputs (text, images, audio, video, PDFs).

SUPPORTED

Checked 2026-11-11 · 2 supporting

SOURCES & TIMELINE

2