Gemini Decision 1 answers 1,000 questions and writes nothing
Google's Gemini Decision 1 returns calibrated probabilities for up to 1,000 questions in one pass, with zero output tokens. The refusals define the product.
Vlad Makarovreviewed and publishedGoogle released the Gemini Decisions API on Wednesday, with a model built for exactly one job. Gemini Decision 1 reads text, images, video, audio and PDFs, accepts up to 1,000 typed questions about that material, and returns calibrated probabilities for each one without generating a single output token. Google's post sells the economics and the latency. The more interesting document is the model page, where the capabilities table lists text generation, structured outputs, function calling, thinking, code execution, batch API, caching, search grounding and the Live API as not supported.
That column is the product. Google is betting that a large share of enterprise AI work is classification wearing a costume, and that routing, triage, policy checks and rubric grading never needed a model that writes prose.
The refusals are the specification
Three question shapes are all it accepts. A yes/no check returns a calibrated probability between 0.01 and 0.99 plus a confidence score. A single-select question takes two to 255 named options and returns the top one with per-option probabilities. A rubric question takes two to 255 ordered levels and returns a probability-weighted expected score. Everything else is out of scope by construction.
The consequences are concrete. It cannot draft the customer email explaining a rejection, because there is no text-generation path. It cannot pick a tool and then invoke it, because function calling is absent from the model. It cannot chain a second decision onto the first: every question is evaluated independently over the same evidence, so a dependent judgment has to be handled in application code or pushed into a follow-up call. VentureBeat is blunt about the boundary, noting that drafting an explanation or pulling an arbitrary invoice number from a document still requires a separate generation or extraction step. Zero output tokens is simultaneously the cost trick and the ceiling. You are buying a scorer, not a worker.
A thousand questions, one pass over the evidence
The batch is where Google's version differs from the decision models that arrived in October. A single request carries up to 262,144 tokens of shared input — the record, the images, the call recording, the scanned form — and the instructions and options for every question are counted inside that same budget. The shared material is read once. Each question adds only its own instruction and criteria tokens.
The pricing follows the same shape: $0.10 per million input tokens, and output billed at zero. The free tier allows 200 requests and 50,000 input tokens per minute. VentureBeat works the arithmetic on a 4,000-token record with 20 checks at 50 tokens apiece. That is a 5,000-token request, roughly $0.50 per 1,000 records at Google's rate. Twenty separate calls that each resend the full record would cost $8.10 per 1,000 records, before retries or follow-ups. The saving is not a discount on the model; it is a discount on re-reading the same document twenty times.
| Published figure | Value |
|---|---|
| Input price | $0.10 per 1M tokens |
| Output price | $0 |
| Questions per request | up to 1,000 |
| Shared input ceiling | 262,144 tokens |
| Free tier | 200 req and 50K input tokens per minute |
| Reported end-to-end latency | 150-250 ms up to 16K tokens |
Five of seven on a scorecard Google built
Google's launch benchmarks compare Decision 1 against TypeSafe's Jev 1.14, OpenAI's Decisions API running on GPT-6 Luna, and Gemini 3.5 Flash-Lite on the media tests. All of it is vendor-run and self-reported.
| Test | Gemini Decision 1 | Rival |
|---|---|---|
| JevBench | 92.2% | GPT-6 Luna 90.9%, Jev 1.14 90.0% |
| Mind2Web Multimodal | 86.2% | Jev 86.8% |
| When2Call | 79.9% | Jev 82.0% |
| MMAU (audio) | 71.8% | Gemini 3.5 Flash-Lite 67.7% |
| Video-MME | 69.1% | Gemini 3.5 Flash-Lite 63.3% |
Decision 1 led five of the seven tests that all four models ran, a real lead on the shared workload. It reverses on the two tasks closest to agent work: Jev won Mind2Web Multimodal and When2Call, the browser and tool-use tests, where a wrong call costs something. The two media wins come against a Flash-Lite sibling, not the decision-model field.
Calibration is the claim worth taking seriously, because it sets how many cases reach a human reviewer at a given threshold. Google reports the lowest calibration error in its comparison, 0.0250 against Jev's 0.0494. VentureBeat flags a result running the other way: on Artificial Analysis's own When2Call run, Gemini scored 75.2% against Jev's 80.8% and Drex Max's 94.6%, and the same outlet notes the Decision Index puts Jev and Gemini close on overall quality.
150 milliseconds, measured by the company that sells it
Google reports 150-250 ms end-to-end for inputs up to 16K tokens, with little change between one question and a hundred, and median model compute of 50-100 ms under those conditions.
The most credible part of the launch is that Google is a customer. Gemini Notebook uses the model as a real-time router, sending each query to a model tier and thinking level, and Google reports up to 40% lower latency for routine queries while deep reasoning is reserved for hard ones. A vendor eating its own routing layer is better evidence than a benchmark table, because the failure mode — misrouting a hard query to a cheap model — is immediately visible to users.
What would settle it
Two things would move this from announcement to fact. The first is calibration measured on a customer's own distribution rather than Google's combined test set, since confidence that tracks accuracy on a curated benchmark can still drift on a queue where 3% of tickets are the hard kind. The second is whether 1,000 questions per request survive contact with a real checklist. Real compliance and claims workflows are full of dependent rules — this determination only matters if that field is present — and a model that evaluates every question independently pushes that logic back into application code, which is exactly the work the API was supposed to remove.
The competitive picture argues for shopping rather than switching. Microsoft and Liquid AI both shipped decision models in October, OpenAI has a Decisions API of its own, and Amazon's Strands Decider can run on your own hardware. Jev remains cheaper on text at $0.042 per million input tokens, though its 32K shared-material limit is a smaller envelope than Google's 262K, unless buyers enable an experimental setting that raises it to the same ceiling. What Google actually added is the batching: one document, one read, a thousand verdicts. Whether that is worth a migration depends on how many of your decisions were ever really questions in the first place.
