Google Launches Gemini Decision 1, a Model That Scores 1,000 Questions in One Pass

TL;DR

Google launched the Gemini Decisions API on Nov 11, running on a new model, Gemini Decision 1. The model scores closed-set choices instead of writing text: send one shared input (text, images, audio, video or PDFs, up to 256K tokens) plus up to 1,000 yes/no, multiple-choice or rubric questions, and get back calibrated probabilities in 150-250 ms end to end. Input costs $0.10 per million tokens, output tokens cost nothing, and the shared input is billed once no matter how many questions you attach. TypeSafe's Jev 1.14 holds this category today; Google's chart puts Gemini ahead, and independent tester TrueStandard found a tie on English accuracy, worse calibration than the chart claims, and text bills 1.2 to 6.7 times Jev's.

What shipped

The Decisions API is a new endpoint on the Gemini API, in preview, announced on the Google blog by Min Choi and Felix Yu of Google DeepMind, with keys and a playground in Google AI Studio. It takes typed questions rather than prompts. Three types exist: noul returns the probability that a statement holds, always between 0.01 and 0.99, so the docs warn that a threshold above 0.99 never matches; choice picks from 2 to 255 fixed options and returns normalized probabilities for each; score grades input against an ordered rubric and returns a weighted score plus per-level probabilities. Every answer carries a confidence value your code can use to escalate to a human.

The smallest call, from the official docs:

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})

result = client.decisions.create(
    model="gemini-decision-1",
    input="The blue mug I ordered on Monday arrived cracked on Wednesday.",
    questions={
        "is_damaged": {
            "type": "noul",
            "instructions": "Did the item arrive damaged?",
        },
    },
)

print(result.answers["is_damaged"].noul)  # 0.99

Google lists four internal customers. Gemini Notebook routes incoming queries to the right model tier and thinking level, cutting latency up to 40% on everyday queries. The rapid canvas experiment in Stitch applies voice and tap edits to playable wireframes in seconds. CC from Google Labs picks emoji reactions from 200-plus options instantly, the most overqualified emoji picker ever shipped. Google Search runs the model as a safety layer over MCP tools. Day-0 ecosystem support covers OpenRouter, LangChain, the Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, with Gemini on Google Cloud listed as coming soon.

Why a model that refuses to talk

Agents burn most of their tokens on calls that never needed prose: should this ticket go to billing or support, is this image safe, which tool fits. A chat model answers those by generating a sentence, and then your code parses the sentence back out. A decision model returns a number with a confidence attached. A generative model writes you a memo; a decision model holds up a scorecard, and you only pay for the scorecard.

The category already has residents. TypeSafe sells Jev, which TrueStandard bills at $0.042 per million tokens, and Latent Space's roundup of the launch names Drex Max and GLiDE as ahead of Gemini Decision 1 on overall quality and cost efficiency in independent evals. The same roundup reports Cloudflare's talk at the AI Engineer CODE event, where a hybrid fast-decision-plus-reasoning agent design cut one flow from 40K tokens to 1.8K. Google arriving in a market a startup opened is a second-mover play, with distribution on day 0 through eight named integrations and a playground.

The scoreboard, and who keeps it

Google's launch chart compares Gemini Decision 1 with Jev 1.14 on nine tests, one calibration score and price. On the JevBench public split it claims 92.2% accuracy against Jev's 90.0%, calibration error of 0.0250 against 0.0494 across the seven tests the models share, and 1.3 times Jev's total price on those tests. The chart also claims the API is nearly 5x cheaper than Gemini 3.5 Flash-Lite on the shared suite, which is a comparison against Google's own model, not against the incumbent.

TrueStandard ran its own suite on Nov 12, the day after launch: 108 claims checked against source passages, three runs, four languages, plus a support-ticket task and screenshot reading. On English claims the two models tied at 96.3%. Japanese went to Gemini, 98.1% against 91.7-92.6%; Spanish went to Jev, 94.4-95.4% against 88.9%; Hindi tied at 95.4%.

Calibration is where the accounts disagree. Google's chart gives Gemini half of Jev's error; TrueStandard measured the reverse on its English and Spanish claims, 0.110 against 0.079-0.086 in English and 0.140 against 0.073-0.086 in Spanish, gaps it says luck does not explain. The tasks differ, so compare direction rather than size. The measured error comes from doubt: in English, every answer Gemini scored between 0.7 and 0.9 turned out right, so the scores ran low.

That hedging has a direct cost in any pipeline built on cut-offs. TrueStandard counted how many of 56 true English claims each model scored at 0.9 or higher: 46-47 for Jev, 18 for Gemini. Forty-nine of Gemini's 108 answers landed between 0.1 and 0.9, against 21-23 for Jev. A rule that auto-accepts at 0.9 and escalates the rest sends 38 true claims to human review with Gemini, and 9-10 with Jev. One consolation: neither model was ever confident and wrong, with no answer at 0.95 or above landing on the wrong side. Gemini also returned identical numbers on every repeat run, while Jev's varied.

The bill and the clock

On text, Gemini cost more in every configuration TrueStandard tried, 1.2 to 6.7 times Jev's bill, because it charges 2.4 times as much per token and bills about 49 tokens per question against Jev's 17. On screenshots the order flips: Gemini bills a third of Jev's image tokens, $0.136 against $0.181 per 1,000. Speed is a wash up to 100 questions per call; at 1,000 questions Jev finished in 0.7 seconds and Gemini took 2.6. Google's 150-250 ms figure holds inside its own envelope: contexts up to 16K tokens, 1 to 100 questions.

The screenshot test found a failure mode worth knowing. Gemini was wrong on 30 of 1,080 checks on dense, ad-heavy pages where Jev was wrong on none, and TrueStandard's advice for screenshot work is to test brand mentions that appear only inside comparisons, as in "Often compared with Altra", which Gemini missed about a third of the time.

The catches

• Google's chart is vendor-run, on tests Google picked, and the one independent test published so far flips its calibration claim on a different task. Treat both as task-specific until more results land.
• TrueStandard's suite is 108 claims per language, and three of the four languages were machine-translated then back-checked. Small sample, honestly reported with ranges.
• The API is in preview, with keys in AI Studio only. Google Cloud support is listed as coming soon, so enterprise procurement waits.
• The 5x-cheaper claim compares against Gemini 3.5 Flash-Lite, another Google model. Against Jev, text calls cost more.
• Audio, video and PDF inputs have no third-party test yet; the independent evidence covers text and screenshots.

Why builders should care

Anywhere your code generates text just to parse it back out is a candidate: request routing, moderation, triage queues, rubric grading, tool selection. The confidence field gives you an explicit escalate path, and batching up to 1,000 questions against one billed input is a pricing model built for exactly those jobs.

If you already run Jev or another decision model, the migration cost is not the API, it is your cut-offs. The same 0.9 threshold admits different answers under a model that hedges, so re-tune on your own traffic before switching anything production-facing.

If this category is not on your roadmap, watch it anyway. A frontier lab shipping a scoring-only model as a first-class API says where agent infrastructure is heading: the expensive general model stays for the hard steps, and a cheap calibrated one takes every branch point on the way there.

Key Takeaways

• Google launched the Gemini Decisions API on Nov 11, in preview, powered by Gemini Decision 1: one shared multimodal input (up to 256K tokens), up to 1,000 questions per call, calibrated probabilities back in 150-250 ms end to end.
• Pricing is $0.10 per million input tokens with zero output tokens, and the shared input is billed once per request regardless of question count. Google's nearly-5x-cheaper claim is against its own Gemini 3.5 Flash-Lite, not against Jev.
• Three question types: noul (yes/no probability, capped at 0.01-0.99), choice (2-255 options with normalized probabilities) and score (ordered rubric with per-level probabilities), each with a confidence value for escalation.
• Google's chart claims 92.2% accuracy and 0.0250 calibration error against Jev 1.14's 90.0% and 0.0494; TrueStandard's independent run tied English accuracy at 96.3% and measured worse calibration for Gemini, 0.110 against 0.079-0.086.
• The operational finding: Gemini hedges. Only 18 of 56 true English claims scored 0.9 or higher against Jev's 46-47, so an accept-at-0.9 rule escalates about four times as many true claims. It did repeat identically across runs, which Jev did not.
• Cost splits by modality: text calls ran 1.2 to 6.7 times Jev's bill (about 49 tokens per question against 17), screenshots ran cheaper ($0.136 against $0.181 per 1,000), and a 1,000-question batch took 2.6 seconds against Jev's 0.7.

Sources: Google, Gemini Decisions API launch post (Nov 11, 2026), Decisions API documentation, Google AI Studio, Gemini Decision 1 model page, TrueStandard, Gemini Decision 1 vs Jev 1.14, measured (Nov 12, 2026), TypeSafe, Latent Space, AINews: D-Day (Nov 12, 2026), SiliconANGLE (Nov 11, 2026), Sundar Pichai on X