Gemini Decision 1

Google's Gemini Decision 1 answers up to 1,000 yes/no, multiple-choice and rubric questions about one photo, call recording, video or PDF in a single pass, in 150 to 250 milliseconds, and Google only bills what you send it. It shipped November 11, 2026, as the model behind the new Gemini Decisions API.

Gemini Decision 1 is a different kind of model from the rest of the Gemini line. You hand it one input (text, images, audio, video or PDFs, up to 256K tokens) and a list of questions, and it scores the candidate answers in one forward pass instead of writing a reply. It handles three question types: yes/no, multiple choice with 2 to 255 options, and ordered rubric scores. Every answer comes back with probabilities and a confidence score from 0 to 1, so your code can send the unsure cases to a bigger model or a person. It costs $0.10 per million input tokens with nothing billed for output, the shared input is billed once no matter how many questions ride on it, and there's a free tier in AI Studio. It's live in the Gemini API and on Vercel's AI Gateway as google/gemini-decision-1, with OpenRouter, LangChain, LiteLLM, Pydantic AI and others rolling out.

Google pitches it for the small calls agents make constantly: which tool to use, which model tier a query needs, whether content breaks a rule. Google says it already runs inside its own products, routing queries to the right model tier in Gemini Notebook (up to 40% lower latency), picking emoji reactions from 200-plus options in CC, and screening MCP tool traffic in Search. Google quotes 50 to 100 ms of model compute and says it's nearly 5x cheaper than Gemini 3.5 Flash-Lite across its own benchmark suite. The catch is the company it keeps: this is a crowded lane of small, dedicated decision models (TypeSafe's Jev, OpenAI's GPT-6 Luna Decisions API, Perplexity's Decider), and the independent numbers below put Gemini Decision 1 in the middle of that pack overall, slower and pricier per task than most of it.

The benchmarks

Decision models score fixed answer options instead of writing text, so they sit on their own Decisions boards on handyai.news. They don't count in the overall weigh-in or next to Gemini, GPT or Claude chat models. Only Artificial Analysis's runs are charted below. Google's own table stays on this page, since no other launch reports those benchmarks in decision mode.

Google's numbers

Google ran Gemini Decision 1 on its Decisions API against TypeSafe Jev 1.14, GPT-6 Luna on OpenAI's Decisions API and Gemini 3.5 Flash-Lite on the Interactions API with minimal thinking. All self-reported. It beats Jev on 5 of the 7 benchmarks both ran: JevBench 92.2% (Jev 90.0%, GPT-6 Luna 90.9%), Amazon ESCI 58.7% (Jev 56.7%), New Yorker Caption 79.7% (Jev 72.0%, GPT-6 Luna 68.4%), PubMedQA 79.8% (Jev 78.2%) and Global-MMLU-Lite 86.0% (Jev 85.0%). Jev wins Mind2Web Multimodal (86.8% to 86.2%) and When2Call (82.0% to 79.9%). Only Gemini Decision 1 and Flash-Lite ran the audio and video tests: MMAU 71.8% (Flash-Lite 67.7%) and Video-MME 69.1% (Flash-Lite 63.3%). Its calibration error is the lowest in the table at 0.0250 (Jev 0.0494, GPT-6 Luna 0.0721). By Google's own total price for the seven shared evals, though, Jev is cheaper: $0.68 against $0.88, with GPT-6 Luna at $0.93 and Flash-Lite at $4.14.

Independent numbers

Artificial Analysis scored it 66 on its Decision Index v2.1, ninth of 18 decision models, compared here with the top five, TypeSafe's Jev 1.14 and OpenAI's GPT-6 Luna Decisions. It took 211 ms a task (Jev 1.14 92 ms, Celeris-1.1 Decision 61 ms, Drex Max v1.1 283 ms) and cost $345 per million tasks (Drex Max v1.1 $103, Decider v1.2 $68, Celeris-1.1 Decision $14).

• Artificial Analysis Decision Index v2.1 (six decision evals): 66.0. −6.6 under Drex Max v1.1 (72.6), −6.5 under GLiDE v1.2 (72.5), −5.4 under Mammoth 27B (71.4), −4.6 under d1-ultra (70.6), −3.8 under Decider v1.2 (69.8), −0.2 under Jev 1.14 (66.2), +5.8 over GPT-6 Luna Decisions (60.2).

• AA-Omniscience-Decision v1.1 (knowledge, wrong answers cost): 66.3. +6.0 over Drex Max v1.1 (60.3), +5.4 over GLiDE v1.2 (60.9), +7.4 over Mammoth 27B (58.9), +1.9 over d1-ultra (64.4), +6.2 over Decider v1.2 (60.1), +6.2 over Jev 1.14 (60.1), +6.1 over GPT-6 Luna Decisions (60.2).

• HLE-MultiChoice v1.1 (expert questions, can decline): 48.7. +2.1 over Drex Max v1.1 (46.6), +4.1 over GLiDE v1.2 (44.6), +1.3 over Mammoth 27B (47.4), +1.3 over d1-ultra (47.4), +4.3 over Decider v1.2 (44.4), +4.2 over Jev 1.14 (44.5), +6.1 over GPT-6 Luna Decisions (42.6).

• AA-Paperclip (calls on business documents): 56.8. −16.0 under Drex Max v1.1 (72.8), −14.6 under GLiDE v1.2 (71.4), −7.4 under Mammoth 27B (64.2), −7.9 under d1-ultra (64.7), −10.5 under Decider v1.2 (67.3), −0.7 under Jev 1.14 (57.5), +3.9 over GPT-6 Luna Decisions (52.9).

First impressions

It's under an hour old, so most of the reach belongs to Google's own accounts. Outside voices so far are Artificial Analysis and a handful of developers.

The positives

Artificial Analysis had numbers out within minutes and put it at the top of the decision models on knowledge and reasoning.

That's the one place it clearly leads, and it's the part Google's own pitch leans on least. Those two evals are 30% of AA's index, which is why the overall 66 sits mid-pack.

waximabbax on Hacker News made the practical case for it.

Its bit more expensive than other other Decision Models in the same lane. However, it can accept Audio and Video as input as well and for those specific cases it can be extremely useful.

That matches Google's table, where only its two Gemini models have audio and video scores at all.

The negatives

Artificial Analysis also ran the bill, and this is where it looks worst.

$345 per million tasks is more than triple Drex Max v1.1, which outscores it on AA's index, and about 25 times Celeris-1.1 Decision. Google's "nearly 5x cheaper" is against its own Gemini 3.5 Flash-Lite, not against the dedicated decision models, and Google's own table has Jev 1.14 cheaper across the shared evals too.

iamdelirium on Hacker News, who posted the launch, summed up how it's being read.

Looks like Google announced their own Jev inspired API and model

Google is late to a category TypeSafe and OpenAI already have products in, and on AA's index it lands level with Jev rather than past it.