Gemini Decision 1 review, pricing and verdict
A purpose-built decision model for routing, policy checks and rubric scoring, separate from Google's generative Gemini models.
- Best for
- Routing a request to a model, queue, agent or tool
- Not for
- Writing summaries or explanations
Should you use Gemini Decision 1?
Google DeepMind released Gemini Decision 1 on 11 November 2026, and it can score up to 1,000 independent questions per request in 150 to 250 ms end to end, by Google's measure. It suits teams that route requests or check policy rules and want to branch on probabilities, not anything needing written output.
Best for
- Routing a request to a model, queue, agent or tool Google lists routing as a primary use and says Gemini Notebook uses the model as a real-time router.
- Checking one input against many policy rules Each rule becomes an independent yes/no question evaluated in the same request.
- Grading candidates against a rubric A score question rates the input on an ordered scale and returns a probability for each level.
Not for
- Writing summaries or explanations Text generation is marked unsupported on Google's model page.
- Multi-step reasoning or tool calls Thinking and function calling are unsupported; Google points those jobs to its Interactions API.
- Fully automatic high-stakes actions such as payouts Google warns that user-supplied text can argue for its own classification and advises a human or second check.
Can you use Gemini Decision 1?
- Gemini API (Decisions API): yes (Free tier at no charge, and a paid tier billed per 1M input tokens). Served on the v1beta API version with the Decisions endpoint only; Batch, caching, Flex and Priority are unavailable.
- Google AI Studio (Decisions Playground): yes (No-cost tier for prototyping, with sample apps).
- OpenRouter: yes (Pay per token, with Google AI Studio as the only provider).
- Free plan: yes.
Checked 12 Nov against Google DeepMind's own pages.
Gemini Decision 1 pricing
- Input, per 1M tokens
- $0.10/1M
- Output, per 1M tokens
- $0.00/1M
- Pricing model
- per-token
- Free tier
- Free of charge, limited to 200 requests a minute, 50,000 tokens a minute and 10,000 requests a day. Google may use free-tier inputs to improve its products.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0030 | $0 | $0.0030 |
| Support reply | $0.0002 | $0 | $0.0002 |
| One coding agent run | $0.020 | $0 | $0.020 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Paid tier: $0.10 per 1M input tokens and no output charge. Free tier: free of charge within rate limits. Caching, the Batch API, Flex and Priority inference are not available.
What Gemini Decision 1 can do
Gemini Decision 1 reads a shared input of up to 262,144 tokens and returns calibrated probabilities instead of text. Google DeepMind built it for the Decisions API, where yes/no, choice and rubric-score questions are answered in one forward pass, with a confidence value on every answer for routing and policy checks.
What goes in and what comes out
Reads
- text
- image
- audio
- video
Writes
- text
Capabilities
- Audio inputSupported
- Video inputSupported
- VisionSupported
- Audio outputNot supported
- Code executionNot supported
- Function callingNot supported
- Structured outputNot supported
- Tool useNot supported
- Web browsingNot supported
Key features
Typed questions, three answer shapes
A request mixes yes/no, choice (2 to 255 options) and score (ordered rubric) questions over the same input.
Confidence on every answer
Each answer carries a 0 to 1 confidence so code can run the action or escalate to a fallback.
Zero output tokens
The model generates no text, so output usage is always 0 and only input tokens are billed.
Multimodal shared input
Text, images, audio, video and PDFs can be combined in one input of up to 262,144 tokens.
Show all 6 features
Many questions in one pass
A single request carries up to 1,000 independent questions, scored in parallel and isolated from each other.
Low-latency scoring
Google reports 150 to 250 ms end to end, with latency staying nearly flat as questions are added (vendor-reported).
Known limitations
- Cannot write text, reason, call tools or return structured JSON: Google's model page marks all of them unsupported, so it cannot explain a decision.
- No context caching, Batch API, Flex or Priority inference, so repeated input across requests is counted in full.
- Question instructions accept text only, and YouTube and gs:// links are refused as media.
Who it is for
- Backend engineers building request routers and classifiers
- Trust and safety teams running policy checks across many rules
- Agent builders deciding which tool or model to call next
- Voice and video app developers who need real-time classification
Gemini Decision 1 benchmarks
Standout result
End-to-end latency per request, vendor-reported
Source: Google DeepMind launch post, 11 November 2026, read 12 November 2026Where it sits
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.075/M | #6 / 103 | peer median $1.71/M |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Strengths and limits
Strengths
- Google's docs show one request answering a yes/no check, a routing choice and a severity score on the same photo, so one call can feed several decisions.
- Output is a probability your code can threshold, which Google's docs contrast with generated text that has to be parsed.
- OpenRouter, an independent listing read on 12 November 2026, shows a single provider for it with 0.15 s latency, in line with Google's own latency claim.
- Scores many independent questions against one shared input in a single request, and Google says latency stays nearly flat as questions are added (vendor-reported).
Show 2 more
- Reads text, images, audio, video and PDFs in one input, with image and video resolution adjustable from low to ultra high to control token cost.
- Returns a calibrated confidence on every answer, so code can escalate low-confidence cases instead of acting on them.
Limits
- Google's launch post shows benchmark results only as images with no published scores, so Hokai cannot verify the claim that it leads the suite.
- Google's documentation lists no text output, reasoning, function calling or structured outputs, so it cannot replace a generative model for any job that needs an explanation.
- Free-tier inputs may be used to improve Google's products, per its pricing page, so sensitive data belongs on the paid tier.
Gemini Decision 1 alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Gemini Decision 1 This page | 262K | $0.10 / $0 | text, image, audio, video, pdf → text | |
 Strands Decider 2BPick Gemini Decision 1 for a hosted API that also reads audio, video and PDFs; pick Strands Decider 2B if you need open-source weights you can run yourself. | Amazon | -- | -- | text, image → text |
 Intern-DecisionPick Gemini Decision 1 for hosted multimodal scoring with audio and video input; pick Intern-Decision if you want open weights in 0.8B, 2B or 4B sizes. | Shanghai AI Lab | -- | -- | text, image → text |
 Gemini 3.5 Flash-LitePick Gemini Decision 1 when the result is a probability your code branches on; pick Gemini 3.5 Flash-Lite when you need generated text or reasoning. | 1M | $0.30 / $2.50 | text, image, video, audio, pdf → text |
Gemini Decision 1 specs
- Context window
- 262K tokens
- Architecture
- Specialised multimodal decision model (single prefill pass)
- Mixture of experts
- No
- Parameters
- Not disclosed
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| Gemini Developer API, Decisions endpoint (POST /v1beta/decisions) | https://aistudio.google.com/docs/decisions-api | API key (x-goog-api-key header) |
SDKs
- Python
- JavaScript
Prompting patterns
Show 2 prompting patterns
Ask narrow, single-purpose questions and combine the answers in code
Separate yes/no questions for is_damaged and needs_refund, joined by an if statement
Google's best practices say atomic checks compose better than one complex question.
Describe every choice option and score level
billing: Payments, invoices, and refunds.
Short descriptions make adjacent options clearly distinct, per Google's docs.
Known quirks
Show 4 known quirks
Yes/no probabilities stop at 0.99
Yes/no answers always fall between 0.01 and 0.99, so a check for a value above 0.99 never matches.
Workaround: Set thresholds inside the 0.01 to 0.99 range, for example 0.95.
Choice options are sorted alphabetically
The API sorts option keys in UTF-8 byte order before evaluating and breaks exact ties to the first key, so a catch-all named none_of_the_above sorts oddly.
Workaround: Name the catch-all option other and check the confidence before acting on the top option.
Input text can argue with the classifier
User-supplied text can argue for its own classification, which affects moderation and approval flows.
Workaround: Write explicit criteria, test with adversarial examples and keep a human or a second check on high-stakes actions.
Repeated input is billed in full
Context caching, Batch API, Flex and Priority inference are unavailable, so sending the same document in many requests costs the full input each time.
Workaround: Ask every question about one input in a single request and send only the fields the questions need.
Gemini Decision 1 facts
- Deployment
- API
- Licence
- Proprietary (Google Gemini API terms)
- Released
- 11 Nov 2026
- Current version
- gemini-decision-1
- Family
- Gemini Decision
- Variant
- Multimodal
- Data retention
- Google's pricing page says free-tier inputs may be used to improve its products and paid-tier inputs are not. The pages read do not state a retention period and do not describe how to opt out of retention.
Content restrictions
- Input safety filters block a request with an HTTP 400 and a code of safety, prohibited_content, blocklist, image_safety or content_blocked.
Safety controls
- System prompt
- No
- Content filter API
- No
- Moderation endpoint
- No
Data governance
- Trains on inputs free tier
- Yes
- Trains on inputs paid tier
- No
Version history
gemini-decision-1 · 11 Nov 2026
Initial release with the Decisions API.
What changed
- Release Gemini Decisions API: Fast, calibrated multimodal decisions source
Questions about Gemini Decision 1
How much does Gemini Decision 1 cost per million tokens?
Google charges $0.10 per 1M input tokens on the paid tier of the Gemini API, the same for text, image, video, audio and PDF input. There is no output charge because the model produces zero output tokens, and the shared input is billed once per request however many questions follow. Caching, Batch, Flex and Priority pricing do not apply because the model does not support them.
Is Gemini Decision 1 free or open source?
It is proprietary and its weights are not downloadable. Google offers a free tier in the Gemini API at no charge, limited to 200 requests a minute, 50,000 tokens a minute and 10,000 requests a day. Open decision models such as Strands Decider 2B and Intern-Decision are the route for teams that want to self-host.
Does Google train on what I send to Gemini Decision 1?
Google's pricing page says inputs are used to improve its products on the free tier and are not on the paid tier. The pages Hokai read state no retention period and do not describe how to opt out of retention. Input safety filters can also block a request, returning an HTTP 400 error with a code such as safety or blocklist.
Should I use Gemini Decision 1 or Gemini 3.5 Flash-Lite for routing?
Gemini Decision 1 returns a probability from a 262,144-token window and generates no text, while Hokai lists Gemini 3.5 Flash-Lite as a generative model with a 1,048,576-token window. Choose Decision 1 when the output is a closed-set answer your code branches on, and Flash-Lite when you need explanations, reasoning or tool calls. Google says Decision 1 is cheaper on its own benchmark suite but shows the figures only in charts.
What are the rate limits for Gemini Decision 1?
The free tier allows 200 requests a minute and 50,000 tokens a minute. Tier 1 raises that to 1,500 requests and 1,000,000 tokens a minute, and Tiers 2 and 3 allow 5,000 and 15,000 requests a minute with unlimited daily requests. Each request may carry at most 1,000 questions.
About Gemini Decision 1
Gemini Decision 1 (model code gemini-decision-1) is a specialised model from Google DeepMind, released on 11 November 2026 together with the Decisions API. Instead of writing an answer, it scores a fixed set of options in one forward pass and returns calibrated probabilities, so application code can branch on the result. It sits apart from the main Gemini line, for example Gemini 4 Carbon: Google's model page marks text generation, thinking, function calling and structured outputs as unsupported.
Read the full background
Google says the model leads its internal suite on agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error, but the launch post shows those results only as charts, so Hokai records no benchmark scores. Google also reports a median of 50 to 100 ms of model compute, or 150 to 250 ms end to end, all vendor-reported. OpenRouter, an independent listing, showed a 0.15 s latency for Google AI Studio when Hokai read it on 12 November 2026.
A request carries one shared input across text, images, audio, video and PDFs, plus a set of typed questions. A yes/no question returns a probability between 0.01 and 0.99, a choice question picks among 2 to 255 options, and a score question grades against an ordered rubric. Every answer includes a confidence value from 0 to 1, which code can use to send doubtful cases to a fallback or a person.
It is served only through the Decisions API on the Gemini Developer API (v1beta), with the Python and JavaScript SDKs or plain REST, and can be tried in Google AI Studio. Google names Gemini Notebook, which uses it to route queries, plus Stitch, Google Search and CC from Google Labs as internal users.
Do not use it for anything that needs written output, multi-step reasoning, tool calls, caching or batch discounts, and the Decisions API refuses YouTube and Cloud Storage links. A team that must run its own classifier should look at open models such as Strands Decider 2B, released under Apache 2.0, or Intern-Decision, which are both listed on Hokai as open decision models. A job that needs an explanation or a JSON object belongs on a generative model such as Gemini 3.5 Flash-Lite.
Google publishes no training cutoff, parameter count, architecture, system card or safety benchmark for this model on the pages Hokai read, and input safety filters can block a request with an HTTP 400 error.
Sources and guides
Official sources
- Gemini Decision 1 model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Google DeepMind's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Lists that include Gemini Decision 1
The provider
More AI Models on Hokai
Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.- Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
- Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.