ENPT

Gemini Decision 1

Last updated:

Beta211–211ms$0.1/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity6.4/10

    Gemini Developer API (v1beta), free and paid. Changelog says Released, not GA; Cloud is still coming. Guide, pricing, limits and the google-genai SDK. LangChain classifier is beta; OpenRouter and Vercel AI Gateway list the model. No SLA found.

  • Capability7.2/10

    Yes/no, choice (2–255) and score. Docs call probabilities calibrated; confidence is the top probability, with no method or ECE. AA v2.1: 211 ms/task, index 65.98 (10th of 19), below Jev 1.14 at 66.23. No Decision Index Full score.

  • Adoption13/100

    Blog and @googledevs on 11 Nov; like counts not collected. LangChain, OpenRouter and Vercel are the checked surfaces; Vercel is counted in integrations, and this partial stays 18. Google's own-product examples stay vendor claims, so production is 0.

Vendor claims

News covering this model

Read 11 Nov 2026 from Google’s blog, the AI Studio docs, the pricing page, the rate-limit page, the changelog and the Artificial Analysis Decision Index page.

The product surface is the Gemini Decisions API. The model is Gemini Decision 1 (gemini-decision-1). Google’s blog post is dated 11 Nov 2026 (Min Choi and Felix Yu, Google DeepMind). The changelog for 11 Nov 2026 says the model and the API were released. @googledevs posted the same announcement at 18:26 UTC (15:26 BRT). Logan Kilpatrick’s earlier post (18:16 UTC) matches the price and the modality list; the figures below come from the docs and the blog, not from the tweet images.

Specs

Attribute Value
Company Google (Google DeepMind on the blog)
Model id gemini-decision-1
Endpoint POST https://generativelanguage.googleapis.com/v1beta/decisions (CreateDecision)
Where it runs Gemini Developer API in Google AI Studio. The blog says Gemini in Google Cloud is still coming
Weights Closed. No parameter count on the model page
Question types noul (yes/no, probability 0.01–0.99), choice (2–255 options), score (2–255 ordered levels)
Input Text, image, audio, video and PDF, one shared input per request. Context window 262,144 tokens across the input and all questions. The blog rounds this to 256K
Questions 1 to 1,000 per request
Output tokens 0. Pricing page: output price “Not applicable (0 output tokens)”
Price $0.10 per 1M input tokens on the paid tier, any of text, image, video, audio or PDF. Free tier: input “Free of charge”. Shared input is billed once; each question adds its own instruction and criteria tokens
Free-tier quota 200 requests/min, 50,000 input tokens/min, 10,000 requests/day, per project (rate limits)
SDK google-genai 3.2 or later, client.decisions.create, api_version: "v1beta"
Not supported Context caching, Batch API, Flex, Priority. No prompt cache price

What the docs say about the numbers

The Decisions API guide says each answer includes probabilities and a confidence score. For choice and score, confidence equals the top outcome’s probability. For yes/no, it equals max(noul, 1 - noul). The guide calls those probabilities calibrated. It does not name a calibration method and it does not publish an ECE.

The catalog latency is 211 ms, the timePerTaskMs on Gemini Decision 1 in the Artificial Analysis page payload read on 11 Nov 2026 (210.85, rounded). That is one number, so min and max are the same. It is an independent per-task time, not a published p50/p99 pair.

Vendor claims (UnverifiedClaim)

The blog states these. They are not the Artificial Analysis table and they are not a Hugging Face Decision Index row.

  • End-to-end latency 150–250 ms, model compute 50–100 ms, nearly flat from 1 to 100 questions up to 16K tokens. The AA time of 211 ms sits inside the 150–250 ms band. The band itself is still Google’s claim.
  • Gemini Decision 1 “leads across the suite, with the lowest calibration error.” The blog text does not include the table (the page shows a figure). No ECE number is in the text we read.
  • “Nearly 5× cheaper” than Gemini 3.5 Flash-Lite with the Interactions API on Google’s shared suite.
  • Gemini Notebook, Stitch, CC and Google Search already call the API. Those are Google’s own products, described by Google.
  • Integrations “are rolling out” across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.

Artificial Analysis Decision Index v2.1 (11 Nov 2026)

Read from the live page titled Decision Index v2.1 (evaluations). This board is separate from the Hugging Face Decision Index and from the v2 table in the 8 Nov note. Scores are not mixed.

Gemini Decision 1 is 10th of 19 decision models on that page, index 65.98 (accuracy index 45.53). Jev 1.14 on the same payload is 66.23, 9th, 92 ms per task. Drex Max v1.1 is first at 72.62. SubQ Snap 1.2 is second at 72.58; that row was not in the 8 Nov table. No other catalog score changes with this reading.

AA v2.1 field Gemini Decision 1
Index 65.98
Document decision 56.83
EvalSafe-O*NET 87.05
When2Call 80.50
Omniscience 66.28
HLE 48.71
Invariance / symmetry 70.47
Time per task 211 ms
Cost per task $0.000345 (about $345 per 1M tasks)

There is no Hugging Face Decision Index Full score for this model, so the accuracy subscore does not use 8 × Full / 59.22. Accuracy stays 6: one independent board, and on that board the model is below Jev 1.14. Accuracy 9 would need two independent boards above Jev, or an untied first place. A second read of the same page on 12 Nov 2026 still shows index 65.98. Capability stays 7.2.

Integrations checked

  • LangChain GoogleGenAIClassifier, read 11 Nov: beta, langchain-google-genai>=4.5.0, model gemini-decision-1, question types Noul, Choice and Score. The page says the API may change without notice.
  • OpenRouter google/gemini-decision-1: the page exists. Its meta description repeats $0.10 per 1M input and $0 output. The catalog price is the Google pricing page, not OpenRouter.
  • Vercel AI Gateway, changelog dated 11 Nov 2026 (datePublished 2026-11-11). The page shows experimental_decide with model google/gemini-decision-1. PR #24938 in vercel/ai, merged 11 Nov 2026 at 18:22 UTC, switches google.decisionModel('gemini-decision-1') to the native decisions endpoint. This was on Google’s “rolling out” list on 11 Nov and had not been checked then.

Score (11 Nov 2026, integrations updated 12 Nov)

Formula in the quadrant doc. Weights unchanged.

  • Maturity 6.4 = availability 6 × 0.30 + docs 9 × 0.25 + integrations 7 × 0.25 + support 3 × 0.20 = 1.80 + 2.25 + 1.75 + 0.60 = 6.40. Availability 6, docs 9 and support 3 are unchanged from 11 Nov. Integrations 6 → 7: LangChain is still beta, and Vercel AI Gateway is now a checked production listing. Not 8: LiteLLM, Pydantic AI and the rest of the blog’s list were not rechecked. Adoption stays 13. The ecosystem partial stays 18; the new listing is counted in the integrations subscore, not as a second adoption bump.
  • Capability 7.2 = types 10 × 0.30 + calibration 5 × 0.25 + latency 7 × 0.25 + accuracy 6 × 0.20 = 3.00 + 1.25 + 1.75 + 1.20 = 7.20. Calibration 5: probabilities are documented and called calibrated, with no published method. Latency 7: 211 ms is inside the “p50 ≤ 300 ms” band and is not a p50 ≤ 100 ms with a p99. Accuracy 6: see the AA section.
  • Adoption 13 = engagement 20 × 0.35 + ecosystem 18 × 0.35 + production 0 × 0.30 = 7.00 + 6.30 + 0 = 13.30 → 13. Engagement 20 because the blog and the @googledevs post exist and like counts were not collected. Ecosystem 18 for the two surfaces above. Production 0: the in-house examples stay vendor claims.