Most agent delays happen before the real work starts. A user sends a request, and your system calls a generative model just to answer a closed question: which tool should run, which model tier fits, or whether the input passes policy.
Generating text token by token to pick one option from a fixed list wastes time and budget. You wait on decoding, pay for output tokens, and parse JSON that only needed to be a probability score.
On November 11, 2026, Google DeepMind introduced the Gemini Decisions API (opens in a new tab), powered by the new gemini-decision-1 model. Instead of decoding prose, it evaluates closed-set choices in a single forward pass across text, images, audio, video, and PDFs. Here is what shipped, how it performs, and how fast decision gates sharpen an enterprise agent stack.
1. Why generative calls slow down agent control flow
An enterprise workflow makes dozens of small routing and policy checks around every generation step.
TRADITIONAL GENERATIVE ROUTING
Input + options -> Decode tokens one by one -> Parse JSON string -> Route
Cost: Input + thinking + output tokens | Latency: Hundreds of ms to seconds
SINGLE-PASS DECISION SCORING (gemini-decision-1)
Shared multimodal input + up to 1,000 questions -> Single forward pass -> Calibrated probabilities
Cost: $0.10 / 1M input, $0 output tokens | Latency: 150-250 ms end-to-end
When every branch waits on text generation, latency stacks up across steps. Generative models are also poorly calibrated when asked to state certainty in words. A control plane needs a calibrated probability and a confidence score so code can decide automatically when to proceed and when to hand off to a person.
2. What Google shipped in Gemini Decision 1
The Gemini Decisions API exposes gemini-decision-1 as a dedicated endpoint for bounded logic: tool selection, model routing, policy checks, and multimodal rubric grading.
You pass one shared input block of up to 256K tokens (combining text, images, audio, video, or PDFs) and up to 1,000 independent questions in a single call. The shared input is billed once per request regardless of how many questions you attach.
Each question uses one of three typed primitives, which can be mixed in the same request:
noul(Yes/No). Estimates the calibrated probability that a binary condition holds, with optional true and false criteria descriptions.choice(Multiple Choice). Evaluates 2 to 255 candidate options in parallel, returning the winning option and normalized probabilities for every candidate.score(Ordered Rubric). Grades the input against an ordered rubric starting at 0, returning a continuous weighted score alongside per-level probabilities.
Every answer returns a confidence score between 0 and 1. Median model compute runs in 50 to 100 ms (150 to 250 ms end-to-end) on contexts up to 16K tokens, and latency stays nearly flat from 1 to 100 questions.
3. How it compares on accuracy, calibration, and cost
At $0.10 per million input tokens and $0 for output tokens, gemini-decision-1 removes output-token charges, variable thinking-token bills, and separate OCR or audio transcription fees. Across Google's seven-benchmark shared suite, it is nearly 5x cheaper than Gemini 3.5 Flash-Lite on the Interactions API while posting the lowest calibration error in the table.
| Benchmark or metric | Gemini Decision 1 | TypeSafe Jev 1.14 | GPT-6 Luna (OpenAI Decisions) | Gemini 3.5 Flash-Lite |
|---|---|---|---|---|
| JevBench (Public split) | 92.2% | 90.0% | 90.9% | 86.6% |
| Amazon ESCI (Macro-F1) | 58.7% | 56.7% | 54.7% | 47.2% |
| New Yorker Caption (Text-only) | 79.7% | 72.0% | 68.4% | 73.7% |
| Mind2Web (Multimodal) | 86.2% | 86.8% | 81.8% | 78.0% |
| PubMedQA | 79.8% | 78.2% | 74.4% | 77.6% |
| Global-MMLU-Lite (Non-English) | 86.0% | 85.0% | 79.6% | 78.8% |
| When2Call (Tool selection) | 79.9% | 82.0% | 70.4% | 60.9% |
| MMAU (Audio) | 71.8% | Not supported | Not supported | 67.7% |
| Video-MME (No subtitles) | 69.1% | Not supported | Not supported | 63.3% |
| Top-1 ECE calibration (lower is better) | 0.0250 | 0.0494 | 0.0721 | 0.2104 |
| Total price (7 shared evals) | $0.88 | $0.68 | $0.93 | $4.14 |
Google's own teams are already using the endpoint as a real-time control plane. Gemini Notebook uses it to classify query intent and pick the right model tier and thinking level, cutting latency on everyday queries by up to 40% while saving deep reasoning models for hard prompts. Google Search runs it inside the latency budget as a safety layer for MCP tools to spot adversarial content, and Stitch Rapid uses it to turn voice and tap edits into playable wireframes in seconds.
4. Putting fast decision gates to work in CommandLyne
Google's internal deployment pattern mirrors how CommandLyne (opens in a new tab) separates decision routing from heavy generation across an enterprise workspace.
- Dynamic model routing before generation. Through Governance and Multi-Model Control (opens in a new tab) and Dynamic Routing (opens in a new tab), CommandLyne matches each step to the right model based on task complexity, user role, latency, and cost. Fast, closed-set classification and triage belong on a lightweight decision step, while deep code synthesis or 1M-token analysis routes to frontier models under your own keys.
- Decision branches and human approval gates. In Agents and Workflows (opens in a new tab), multi-step workflows branch on structured choice, score, and boolean decisions. When a step touches production systems or carries a low confidence score, the workflow routes directly to a human approval checkpoint with Preview, Approve, Adjust, Cancel, and Rollback controls.
- Scoped MCP tool boundaries. Just as Google Search pairs MCP tool calls with a fast safety check, CommandLyne enforces MCP and tool policies (opens in a new tab) per agent and logs every prompt, data access, and tool execution in the audit trail.
| Workflow stage | Generative-only approach | Decision-gated orchestration |
|---|---|---|
| Intent and tier routing | Full LLM prompt before the real call | Sub-250 ms classification to pick the right tier |
| Multimodal intake check | Separate OCR, transcription, and grading calls | One shared input scored across up to 1,000 rules |
| MCP tool safety check | Slow prompt guard or unscoped execution | Fast pre-execution policy check + scoped tool access |
| Ambiguous edge cases | Model guesses in prose | Confidence threshold routes to human approval |
| Token spend | Output tokens billed on every routing step | $0 output tokens on decision gates + spend tracking |
5. Where to start
Audit your existing agent workflows and separate choosing from writing. Anywhere an agent selects a tool, classifies a ticket, grades a document against a rubric, or decides which model tier should answer next, that step is a closed-set decision rather than a text-generation task.
Putting fast, calibrated decision gates in front of your frontier models cuts latency on every turn and saves your reasoning budget for the steps that genuinely need it.
Try CommandLyne to orchestrate multi-model routing, decision branches, and governed MCP workflows in one platform, or book a demo with our engineering team.
Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

