FULL STORY

Google Launches Gemini Decision 1: Launch and First Benchmarks

Google launched the Gemini Decisions API and its first frontier decision model, Gemini Decision 1, on Nov 12. Early benchmarks from Artificial Analysis show top knowledge and reasoning but high costs.

2026-11-12 ~ 2026-11-12 · 2 episodes · 30 posts

Episode 1 · Google Launches Gemini Decisions API with Gemini Decision 1 Model (2026-11-12, 24 posts)

On November 12, Google officially launched the Gemini Decisions API, powered by a new frontier decision model, Gemini Decision 1. Both Sundar Pichai and Google AI Studio made the announcement; the product is live in Google AI Studio and available via the Gemini API. Rather than token-by-token text generation, it is purpose-built for closed-ended judgment scenarios.

Confirmed

  • Aggressive pricing: $0.10 per million input tokens, output tokens free of charge, no extra fees
  • Three decision types: yes/no, multiple choice (up to 255 options), and rubric-based scoring
  • Multimodal input: native support for text/JSON, PDF, images, audio, and video, with a 256K-token context window
  • A single request can deliver fast, well-calibrated decisions on up to 1000 questions, with end-to-end latency of 150-250ms
  • According to Logan Kilpatrick, head of Google AI Studio, Gemini Decision originated as a grass-roots project inside Google DeepMind; the team built a prototype, opened it to internal testing, and went from prototype to launch in roughly 6 weeks
  • Per @philschmid (who says he was involved in designing the API), it is built for yes/no, multiple-choice, and rubric-type judgments, returning answers in about 150ms

Why it matters

  • The product shifts LLMs from open-ended text generation toward high-throughput, low-cost, structured batch judgment, with pricing and latency metrics directly attractive to content moderation, evaluation, and data labeling applications
  • The path from internal DeepMind tool to launch in 6 weeks also showcases a new model of small Google teams rapidly productizing internal projects

4 more related posts →

Episode 2 · Gemini Decision 1 debuts: tops decision benchmarks but costs 3x rivals (2026-11-12, 6 posts)

Google has released its first decision model, Gemini Decision 1, and Artificial Analysis published a full evaluation on November 12: a Decision Index v2.1 score of 66, short of frontier level, on par with models like Elephant 9B and Jev 1.1.

Confirmed

  • Overall score 66; it tops the knowledge-reasoning subcategory (AA-Omniscience-Decision), ahead of d1-ultra (64) and the GLiD series, though @ArtificialAnlys notes its edge comes from calibration rather than raw accuracy
  • The evaluation table shared by @Angaisb shows: Jevbench 92.2%, Amazon ESCI 58.7%, New Yorker Caption 79.7%, PubMedQA 79.8% — leading on multiple benchmarks with a calibration error of just 0.025
  • Latency: averaging 211ms per 21 tasks, faster than Drex Max v1.1 (283ms) and GLiDE v1.2 (284ms) but slower than small models, mid-tier overall
  • Cost: $345 per 1 million tasks, close to GPT-6 Luna (Decisions, $339) and more than 3x that of leaderboard leader Drex Max

Why it matters

This marks Google's first entry into the decision-model arena. @ArtificialAnlys's evaluation shows it is competitive in reasoning quality and calibration, but its value for money clearly lags — it would need major cost reductions to challenge the top of the leaderboard.