Google's Gemini Decision 1 gives you a decision model at $0.10 per million input tokens with no output charge, and it can answer up to 1,000 questions against one shared input in a single call. For now it runs only on the v1beta Gemini Developer API, with no Vertex AI route, no caching and no batch. On the one independent leaderboard that has scored it, it ties TypeSafe's Jev on accuracy while costing more than twice as much per task. If you run a model router, a policy classifier or an eval grader on a generative model, test it this month against what you already use, and keep regulated data off it until Google ships it on Google Cloud.
Google DeepMind announced the Gemini Decisions API and the gemini-decision-1 model on November 11. Vercel listed it on AI Gateway the same day. That puts Google alongside the other large platforms that already sell decision models, including OpenAI's Decisions API, Microsoft's Decision-1 on Foundry and Salesforce's Koa inside Agentforce.
What Gemini Decision 1 Actually Does
Gemini Decision 1 scores a fixed set of answers in one forward pass and returns probabilities, so it never writes text. A decision model is a model that picks among options you define (yes or no, one of N, a rubric level) and reports how confident it is, instead of generating a free-form reply you then have to parse.
Google's launch post describes three question types. A noul question returns the probability that a condition holds. A choice question picks from 2 to 255 options and returns the winner plus normalized probabilities. A score question grades against an ordered rubric and returns a weighted score. Every answer carries a confidence value between 0 and 1, which is what lets your code send the ambiguous cases to a fallback model or a person.
The shape of the request is where it differs most from the competition. The Decisions API documentation allows 1 to 1,000 questions per request against one shared input of text, images, audio, video or PDF, inside a 262,144-token context window. OpenAI's Decisions API caps a request at 400 questions. Google says end-to-end latency runs 150 to 250 ms and stays nearly flat from 1 to 100 questions, for inputs up to 16K tokens.
Google's own examples are internal. It says Gemini Notebook uses the model for real-time routing with up to 40% lower latency on everyday queries, and that Google Search uses it as a safety layer that detects adversarial behavior in MCP tools. Treat both as vendor claims: the post gives no figures for the Search use, and no outside customer is named.
How the Price Compares to OpenAI, Microsoft and Jev
On list price per token, Google matches OpenAI and costs about 2.4 times what Microsoft and TypeSafe charge. Google's pricing page lists Gemini Decision 1 at $0.10 per million input tokens on the paid tier, free on the free tier, with no output charge.
| Decision model | Input, per 1M tokens | Output | Max questions per request | Caching |
|---|---|---|---|---|
| Gemini Decision 1 | $0.10 | free | 1,000 | not supported |
| OpenAI Decisions API (GPT-6 Luna) | $0.10 uncached | free | 400 | yes, discounted hits |
| Microsoft-Decision-1 | $0.042 | free | not stated | not stated |
| TypeSafe Jev 1.14 | $0.042 | free | no separate limit | not stated |
Sources: Google pricing and Decisions API docs, OpenAI Decisions guide, Microsoft, TypeSafe models page. All checked November 14, 2026.
Google's billing has one feature that changes the arithmetic for multi-question work. The launch post says the shared input is billed once per request, and each extra question adds only its own instruction and criteria tokens. If you grade one document against 50 rubric items, you pay for the document once.
Input that repeats across requests gets no such break. The documentation says context caching is not supported and that "Repeated input across requests counts in full." A policy classifier that sends the same 6,000-token policy with every request pays for those 6,000 tokens every time. At $0.10 per million, that is $600 of every million requests spent on text that never changes (our arithmetic, at list price). OpenAI caches a repeated prefix by default and bills cache hits at a discount, per its Decisions guide. The right choice depends on which way your traffic repeats: many questions about one input favor Google, and one long fixed prompt across many inputs favors OpenAI.
What the Independent Benchmark Shows
Artificial Analysis ranks Gemini Decision 1 12th of 22 decision models, exactly at the median, which is a weaker result than Google's launch post implies. Google says the model "leads across the suite, with the lowest calibration error," but its post shows the benchmark numbers only as chart images, on a suite Google chose.
The Artificial Analysis page for Gemini Decision 1 gives it a Decision Index score of 66, against a field median of 66. It takes 211 ms per task, and it costs $345 per million tasks, which ranks 19th of 22 on cost against a median of $137. That same page lists text and image as its input modalities, narrower than the audio, video and PDF Google advertises, so the index has not yet tested the multimodal case Google is selling.
The closest comparisons on the same index:
- TypeSafe Jev 1.14 also scores 66, runs in 92 ms per task and costs $156 per million tasks.
- Microsoft-Decision-1 scores 62, runs in 274 ms and costs $143 per million tasks.
So on this index, Jev matches Google's accuracy at less than half the cost per task and less than half the latency. The index measures tasks one at a time, though, and that is exactly where Google's shared-input billing gives it no help. If your workload asks dozens of questions about one input, the per-task figure overstates what you would pay. Run your own traffic before you trust either number, the same advice we gave when Liquid's d1-ultra topped the index on text-only tests.
Why It Is Not Ready for Regulated Data Yet
Gemini Decision 1 is available only on the v1beta Gemini Developer API, so it does not yet come with Google Cloud's enterprise data terms, regions or support. The documentation says it is "Available on the Gemini Developer API (v1beta) with Gemini Decision 1 only," and Google's launch post says support in Gemini in Google Cloud is "coming soon," with no date. Vertex AI is not mentioned anywhere.
First, free-tier inputs can be used by Google and read by its reviewers. Google's pricing page marks free-tier usage of Gemini Decision 1 as "used to improve our products," and the Gemini API terms say human reviewers "may read, annotate, and process your API input and output" on unpaid services and tell users not to submit confidential or personal data there. An engineer who prototypes a ticket router in AI Studio with real customer tickets has just sent them to that tier. Paid usage is not used for product improvement, although the terms say Google logs paid prompts for "a limited period of time" to detect abuse, with no stated duration.
Second, the European rule is strict. The terms say you may use only paid services when making API clients available to users in the European Economic Area, Switzerland or the UK.
Third, the operational features enterprises lean on are missing. The documentation lists batch, context caching, and Flex or Priority inference as unsupported. The rate limits page gives Tier 1 projects 1,500 requests per minute and 1,000,000 tokens per minute, a ceiling a high-volume router can reach. Compare that with OpenAI, whose Decisions API is generally available and supports zero data retention, HIPAA use for eligible customers, and regional processing in the US and Europe. If zero retention is a hard requirement for you, our zero data retention comparison covers how each provider handles it.
The strongest argument for Google is reach. Gemini Decision 1 already works through OpenRouter, LangChain, LiteLLM, Pydantic AI and Vercel's AI Gateway, per the launch post, and Vercel exposes a TypeSafe-compatible endpoint. A team already on those layers can point a shadow test at it with a config change. If your company standardized on Gemini and Google Cloud, a decision model from the same vendor will be the easiest one to approve once it reaches that cloud, and Google has not said when.
What to Do About Gemini Decision 1
Treat Gemini Decision 1 as a candidate for your decision layer, test it on your own traffic, and keep production data off it until it reaches Google Cloud.
This Week:
- Inventory every place a generative model answers a closed question: routers, tool selection, guardrail checks, intent classification, LLM-as-judge graders. For each, note how many questions it asks per input and how long the fixed part of the prompt is. Those two numbers decide whether Google's billing helps or hurts you.
- Tell your engineers in writing that AI Studio's free tier is not for customer, employee or regulated data, and point them at the pricing page line that says free usage improves Google's products.
This Month:
- Run a shadow test on one workload. Send the same 1,000 to 5,000 labeled cases to Gemini Decision 1 and to your current option (Jev, Microsoft-Decision-1 or OpenAI's Decisions API) and compare accuracy, calibration on the cases you route to humans, latency and cost per decision at your real question count. Our guide to decision models versus fine-tuned classifiers has a test design you can reuse.
- Ask your Google account team three questions: the date for Vertex AI availability, which regions it will serve, and whether caching or batch is planned. Get the answers in an email.
Before Production:
- Do not move a regulated workload until Gemini Decision 1 runs on Google Cloud under your existing data terms, or until your legal team has signed off on the paid Developer API terms as written.
- If you route through a gateway, set a budget alert on the new model first. Vercel says decision calls count toward AI Gateway budgets alongside other model requests.
The Bottom Line
Google now sells a decision model on the same terms as everyone else (free output, a cheap input meter, calibrated probabilities), and it differs in two places: 1,000 questions per call with the input billed once, and a v1beta-only release with no cloud route. Since we covered TypeSafe's Jev in September, OpenAI, Microsoft, Cloudflare, Liquid, Salesforce and now Google have all shipped decision models. On the independent index, Gemini Decision 1, Jev and Microsoft-Decision-1 sit within four points of each other, so the deciding factors are how your traffic is shaped, which contracts you already hold, and how fast each vendor gets its model onto enterprise terms.
If most of your decision calls ask many questions about one document or image, put Gemini Decision 1 in the shadow test this month. If they send one long fixed prompt per call, OpenAI's caching or Jev's lower token price will likely cost less.
Continue Reading
- Long Rubrics Can Make OpenAI Decisions API Cost 4x More
- Microsoft Brings Decision-1 to Foundry and OpenRouter at Jev's Price
- Liquid d1-ultra charges $0.042 per million input tokens, like Jev
- Cloudflare's Clef Beats Jev at Routing and Loses at Judgment
- Agentforce Users Should Test Salesforce's Koa Decisions Upgrade
- Gemini 4 Hydrogen Costs $1.87 a Task vs $7.77 for 3.8 Flash