One support ticket can need three independent checks: which team should receive it, how urgent it is, and whether the message describes a working alternative. Gemini Decision 1 is designed for those bounded judgments, returning typed answers and probabilities rather than generating answer text.
AIHubMix lists Gemini Decision 1 on a separate decision endpoint, not Chat Completions. The integration below is documentation-based and has not been executed. No authenticated request, output measurement, or latency test was run for this article.
Choose the answer type before writing the request
Google's model specifications describe three answer types:
| Type | Question to express | Answer value |
|---|---|---|
| noul | Does this condition hold? | Probability of yes |
| choice | Which route fits? | One option plus probabilities |
| score | How severe is this? | Expected rubric level |
On Google's native API, noul ranges from 0.01 to 0.99. A choice accepts 2–255 options; a score accepts 2–255 ordered levels, with 2–10 recommended for most rubrics. A score can be fractional because it is the probability-weighted average of level indices, not necessarily the most likely level.
The native API supports up to 1,000 questions within a 262,144-token window covering the input and all questions. It documents text, images, audio, video, and PDFs. These are native Google capabilities, not verified gateway limits or media support.
Questions share input but do not share answers. One question cannot read another question's result. Combine independent outcomes in application code; use another request when a later decision genuinely depends on an earlier one.
Count all billed input
The official pricing page sets Decision's standard price at $0.10 per million input tokens, with no output tokens. Shared input is billed once per request, while each question adds instructions, criteria, and formatting tokens.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Gemini Decision 1 | $0.10 | No output tokens |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 |
These are official standard direct-API prices in USD. Flash-Lite's output price includes thinking tokens. AIHubMix's listed rates matched these prices in the November 12, 2026 check.
For an illustrative, not measured workload, assume 100,000 tickets per month, each sent once with three questions, and 2,000 billed input tokens per request including those questions:
- Decision: 200 million input tokens × $0.10/M = $20.
- Gemini 3.5 Flash-Lite: the same assumed input costs $60. Another assumed 40 output tokens per request produces four million output tokens, costing $10, for $70 total.
The estimate excludes caching, retries, taxes, and downstream review. Equal assumed billed input does not establish equal tokenization or equal quality. More routing errors can outweigh a smaller token bill.
The Decisions API has no context caching or Batch API. Shared-input billing helps within one request; repeating the state in another request bills it again.
Use the gateway's request envelope
The AIHubMix model instructions document POST /v1/systemone, Bearer authentication, and the fields model, state, and questions.
Google's native API instead uses /v1beta/decisions and input. The envelopes are not interchangeable. Changing a base URL on a native SDK example does not establish that its method and fields will translate.
This is the unexecuted documentation example, using a fictional ticket and text only. Media behavior through the gateway has not been verified here.
curl --fail-with-body https://aihubmix.com/v1/systemone \
-H "Authorization: Bearer $AIHUBMIX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-decision-1",
"state": "The export button fails in Safari. Chrome works, but I need to send the report today.",
"questions": {
"department": {
"type": "choice",
"instructions": "Which team should handle this support request?",
"criteria": {
"billing": "Charges, invoices, or refunds.",
"technical": "A product feature is not working.",
"other": "The request does not fit either department."
}
},
"urgency": {
"type": "score",
"instructions": "How time-sensitive is the request?",
"criteria": [
"No deadline is stated.",
"A future deadline is stated, but not today.",
"The user needs the task completed today."
]
},
"has_workaround": {
"type": "noul",
"instructions": "Does the message explicitly describe a working alternative?"
}
}
}'
choice.criteria is an object; score.criteria is an ordered array. Put the complete question in instructions, rather than expecting the question key to supply its meaning.
The current AIHubMix instructions contain a response example naming a Jev model. That is not a Gemini transcript. It establishes neither Gemini's actual response shape and routing nor its output-token accounting, score semantics, or confidence semantics. Validate the real returned types and usage before integration. Full native media support, context limits, and question-count limits through this gateway remain unverified here.
Failure modes belong in the routing policy
The Google Decisions API guide recommends selecting thresholds against labeled examples from your own traffic, then measuring errors on a separate held-out set.
A valid-looking result still needs validation
Validate the response and required fields before reading an answer. Check that the selected route belongs to the allowed options. Send uncertain, out-of-scope, or malformed results to review. Include an other option when no route may fit: a closed option list still distributes probability even when every listed option is unsuitable.
High confidence is not the probability of yes
On Google's native API, confidence is the highest outcome probability. For a yes/no question it is max(p, 1 − p), not a second model independently confirming the result. A high-confidence no is still no. Do not use confidence in place of the probability that the condition holds.
For scores, decide whether the application needs the expected score or the single most likely level. They can lead to different actions. These native semantics are not established for Gemini by the gateway's Jev response example.
A missing answer must not become approval
Google's documentation shows blocked inputs returning HTTP 400 rather than decision answers. That is a native documentation example, not a measured AIHubMix response. Handle non-success responses separately using the gateway's HTTP error guidance.
A typed answer is not authorization to execute an action. Authorization and high-stakes review remain application responsibilities.
What this integration does not demonstrate
Google reports median end-to-end latency of 150–250 ms for inputs up to 16K tokens and nearly flat latency from 1 to 100 questions. These are vendor-reported figures, not AIHubMix measurements, and not a promise for 1,000 questions or the full context window.
Decision is a plausible fit for routing, policy checks, candidate ranking, and narrow agent stop/escalation decisions. Keep a generative model when the application needs an explanation, arbitrary extracted values, tool arguments, or multi-step reasoning.
The next integration step is to validate actual response types and usage, then evaluate a routing policy against labeled traffic. The example above demonstrates a documented request structure, not a working integration or a measured result.
Keep reading: the Gemini Decision 1 series
- To compare routing policies rather than just request formats, continue with Gemini Decision 1 vs Jev 1.14: Evaluate the Routing Policy, Not One Score.
Sources
Prices reflect the November 12, 2026, 06:43 UTC check. Specifications and prices are documentation snapshots, not inference measurements.
- Gemini Decisions API launch announcement — Google DeepMind.
- Gemini Decision 1 model specifications and Decisions API guide — Google.
- Gemini API pricing — Google, standard USD prices for Decision and Flash-Lite.
- Gemini Decision 1 gateway instructions and Flash-Lite listing — AIHubMix.
- HTTP status-code reference — AIHubMix.
Top comments (0)