Deep Dive — Gemini Decisions API: $0.10 per million, free output

Google shipped an API endpoint this week that cannot write a sentence. The Gemini Decisions API, released on November 11 alongside a single-purpose model called Gemini Decision 1, takes one multimodal input — text, images, audio, video, PDFs, up to 262,144 tokens — and returns probabilities instead of prose. Up to 1,000 independent yes/no, multiple-choice or rubric questions can be asked about that same material in a single request. The price is $0.10 per million input tokens and zero for output, because there is no output to bill, per Google's launch post.

That is the whole product, and it is a bigger deal than the launch coverage suggested. Not because Google built something new — it did not — but because the largest model vendor in the world just agreed, in public and with a price list, that most of what agents ask a model is a closed-set question, and that generating a paragraph to answer it was always waste.

The latency claim is the one Google leads with: 50 to 100 milliseconds of model compute and 150 to 250 milliseconds end to end on inputs up to 16,000 tokens, nearly flat whether you ask one question or a hundred. A developer testing with a free key measured 0.24 seconds median for one and ten questions, rising to 0.41 seconds for a hundred, according to Zhidx's write-up of the test. Same input, same pass, 100 judgments, less than half a second.

The model that doesn't write

Under the hood the mechanism is simple and that is the point. A normal chat model decodes an answer one token at a time, and every token is another forward pass and another line on the invoice. Gemini Decision 1 skips decoding entirely: it scores the candidate options you supplied during the prefill pass, then returns normalized probabilities and a confidence score for each question. Nothing is generated, so nothing is billed as generated.

Three question shapes cover the space, and they can be mixed in one request. A yes/no question ("noul", in the schema Google copied from TypeSafe's Jev) returns a calibrated probability between 0.01 and 0.99 plus confidence, with optional descriptions of what true and false each look like. A choice question picks from 2 to 255 named options and returns the winner with a probability for every candidate. A rubric question grades the input against 2 to 255 ordered levels and returns a probability-weighted expected score. This is the plumbing of an agent's control flow — which tool, which queue, which policy rule tripped, how urgent is this — exposed as a first-class API instead of improvised as a chat completion whose JSON you have to parse and hope about.

Google's own numbers put the model's median compute at 50 to 100 milliseconds and total round trip at 150 to 250 for inputs up to 16,000 tokens. If you have ever watched an agent take four seconds to decide whether to call search or lookup, that gap is the product.

What the confidence score is actually for

The interesting design decision is not the probability output — classification models have returned probabilities for a decade. It is that Google treats confidence as an instruction to the caller: set a threshold, and anything below it goes to a fallback model or a human.

That reframes the whole reliability conversation. An agent stack built on chat models has two states: the model answered, or it errored. An agent stack built on decision scores has three: answered confidently, answered uncertainly, refused to answer at all — and the middle state is where real work actually goes, into a review queue with a priority number attached.

Google's application example is a photo of a breakroom counter plus a text note: does liquid need cleaning, which of three actions applies, how dirty is the scene, all in one call over one image. Change the domain and the same request shape is a claims intake system: a scanned form, damage photos, a call transcript, and twenty checks about queue, missing evidence and urgency. That is exactly the workflow where the industry has been paying for a chatbot to write "Based on the image provided, it appears that…" before the actual answer.

Google says the endpoint already runs inside the company: Gemini Notebook uses it to route queries between models and reasoning levels, with up to 40% lower latency on routine queries, and Google Search uses it as a safety layer for MCP tools, screening for adversarial behavior without adding a generation step. Both are in VentureBeat's account of the launch.

The arithmetic: 50 cents per thousand records

Here is where the pricing model gets interesting, and where the "5x cheaper" headline needs unpacking. Google's comparison is against Gemini 3.5 Flash-Lite through the Interactions API on a shared benchmark suite: $0.88 of inference with Decision 1 against $4.14 with Flash-Lite, roughly 4.7 times. True, and also the least interesting comparison available, because Flash-Lite is a general model doing a job it was not built for.

The batching is the real saving. The shared input is billed once per request no matter how many questions hang off it; each additional question costs only its own instructions and criteria. Take a 4,000-token customer record and 20 checks at 50 tokens each. One request costs about $0.0005, or $0.50 per thousand records. The same twenty checks as twenty separate calls resend the record each time: roughly $8.10 per thousand records, before retries. Sixteen times the bill for identical work, per VentureBeat's worked example.

That arithmetic kills a whole category of glue code. Teams have been paying twice for the same evidence — once when the agent reads the ticket, again when the router reads it to decide who handles it. Billing the evidence once and the checklist per item makes the marginal check almost free: the twenty-first check costs a fraction of a cent instead of another model call.

It also does something quieter to the market: it makes the checklist the asset. The prompts and rubrics that define what "urgent" or "policy-violating" means become the thing worth writing down, reviewing and versioning, rather than the model choice. We made a version of this argument in our deep dive on token prices halving while the cost to make them didn't: when the per-token price collapses, the value moves to whoever controls the specification. A model that charges nothing for output pushes that collapse to its logical end — the output is not a product at all.

Google's benchmark table vs. everyone else's

Google reports leading scores across agent routing, reasoning, ranking and multimodal understanding, with five wins of seven shared tests. The specifics, per the launch post: 92.2% on JevBench against 90.9% for GPT-6 Luna and 90.0% for Jev 1.14; 71.8% to 67.7% on the audio benchmark MMAU against Gemini 3.5 Flash-Lite; 69.1% to 63.3% on Video-MME against the same. Jev still wins Mind2Web Multimodal, 86.8% to 86.2%, and When2Call, 82.0% to 79.9% — two narrow losses Google disclosed rather than buried.

Independent testing tells a less flattering story. Artificial Analysis's Decision Index v2.1, which combines six evaluations including document review, tool calling, knowledge and reasoning, ranked Gemini Decision 1 ninth of 18 at launch. Its score was 65.98, behind Jev 1.14's 66.23 at 92 milliseconds a task and $156 per million tasks, behind Drex Max v1.1's 72.62 at $103 per million, and well ahead of GPT-6 Luna's 60.25. On the same suite, Gemini took 211 milliseconds per task and $339 per million tasks — four times Jev's cost for a slightly lower score, according to VentureBeat's comparison.

The gap widens on all-or-nothing measures, which are the ones that matter operationally. On AA-Paperclip, which asks 15 questions about each résumé, paper, invoice or receipt, Gemini answered every question correctly on 33% of documents. Jev managed 35.5%, Drex Max 63%. A workflow where a missed field means a human reopens the file cares about that number, not the average per-question accuracy. When2Call, which tests whether to call a tool at all, went 94.6% for Drex Max, 80.8% for Jev and 75.2% for Gemini.

So the honest read of launch week: Google did not win the decision-model category on quality. It won the multimodal corner of it, and it matched everyone's price. Jev costs $0.042 per million input tokens — less than half — but takes text and structured JSON only, with opt-in images and a 64,000-token standard limit. GPT-6 Luna matches Google's $0.10 and can attach a different image to each question, and its cached input reads at $0.01 per million reward teams that resend the same policy manual. Gemini's differentiator is the media: one request that reads a call recording, four damage photos and a scanned claim is a workflow neither competitor can run as a single call today.

The category is crowded enough that this is now a comparison question whose answer depends on your input type, not a winner. A healthy market is a terrible pitch deck — which is why the launch post leads with latency and price.

Google shipped Jev's API, down to the vocabulary

The design debt is worth naming. TypeSafe's Jev shipped first with the same three question types, the same single-pass scoring model and the same flat per-input pricing. OpenAI followed on October 22 with its Decisions API on GPT-6 Luna, covering up to 400 questions per request. Cloudflare's Clef and Clef-Omni went open-weight. Microsoft built a Decision 1. Google's entry reuses Jev's name for the yes/no type — "noul" — verbatim.

Simon Willison, who tested the endpoint against a video file within hours of launch, noted that the API is "almost identical to Jev", down to the terminology. His four-question test on a robot video returned 0.92 for "does this video have a robot" and 0.92 for "is this set in the USA" — not the 1.0 and 0.0 that Jev and OpenAI return, but a hedge in both directions.

Two things follow. The generous reading is that the schema is converging on a standard the way chat completions did in 2023, and a standard with four vendors behind it is good for everyone writing against it. The less generous reading is that the category was defined by a startup eight weeks ago and is now being standardized by the companies that can afford to give it away. Google's pricing at parity with OpenAI and 2.4 times Jev's rate per token is not a price war; it is a floor set where the commodity can be absorbed into a cloud bill.

There is also a calibration question hiding in Willison's 0.92. Google's headline technical claim is the lowest calibration error in its comparison — 0.0250 against Jev's 0.0494, roughly half. Calibration measures whether a stated 90% is right 90% of the time, and it is genuinely the right metric for a system that routes on thresholds. But a model that never says 1.0 will beat one that does on expected calibration error while being less decisive at a fixed cutoff. If your policy is "route to a human below 0.9", a confident-but-hedged model sends more correct cases to the review queue. Google's confidence output is a 0.01-to-0.99 range by design; whether that is principled modesty or a ceiling that costs you throughput is something every team will have to measure on its own data.

Where it breaks

The failure mode is the one every classifier has and few vendors volunteer. Asked to classify inputs that fit no category — the capital of Australia, random characters, a request for an autumn poem — Gemini Decision 1 returned probabilities of 0.01, 0.07 and 0.01 that these were support requests, correctly saying no. With no "other" option in the choice question, it still assigned all three to technical support, at confidence 0.71, 0.64 and 0.84. Add an "other" option and it picked that, at 0.99, 0.69 and 0.98, per the same Zhidx test.

That is not a bug in the model so much as a property of closed-set scoring: every input gets distributed across the options you gave it, because there is nothing else to do with the probability mass. The confidence score does not save you — 0.84 is well above most thresholds. The mitigation is a catch-all option in every choice question and a separate yes/no gate in front, which is a design rule, not a setting. Anyone shipping this into a triage pipeline without one will find out from their support queue.

The second limitation is structural. Questions in a request are evaluated independently; if decision two depends on decision one, you handle the dependency in code or make a follow-up call. There is no chain of reasoning inside the request and no way to ask "if yes, then which of these three". Rubric scores give you a graded answer where a chain would have gone, and for most policy checks that is enough — but multi-step deduction is still a chat model's job.

Third, no structured outputs, no function calling, no batch API, and no caching on the Gemini side. Resending a 40,000-token policy manual with every inbound email pays full freight, where OpenAI's cached reads cost a tenth of the input rate. Google lists Cloud support as coming soon.

What to watch

The pricing is the part competitors will copy first and the part that matters least in a year. Watch three things instead.

Whether video and PDF become table stakes. Google's only durable advantage today is the media it can read in one pass. If OpenAI adds PDFs to the Luna endpoint and cache reads keep it cheaper, Google is left arguing about a benchmark its own numbers win and independent numbers don't.

Whether anyone publishes an all-or-nothing decision benchmark with a per-document pass rate rather than per-question averages. Google's 92.2% on JevBench and Artificial Analysis's 33% all-correct on AA-Paperclip can both be true, and only one of them predicts whether your claim file gets processed without a human. Right now a buyer cannot compare vendors on the number they actually care about.

And whether the Google Cloud integration arrives with the data-use terms enterprises need. VentureBeat notes that paid-service prompts are excluded from product improvement while unpaid-tier terms permit it — a detail that decides whether the free tier in AI Studio is a prototype sandbox or a compliance problem. For a model whose entire job is to read customer records, that is not a footnote.

The decision-model category now has five vendors, two price points and no agreed yardstick. We covered the launch briefly in Wednesday's roundup; it deserves more than a paragraph because it changes what an agent's control flow costs, and price decides which architecture ships. When a bounded question costs a fraction of a cent and returns a probability you can audit, the router stops being infrastructure you build and becomes a line item you buy. If you want the mechanics of a single scoring pass first, our explainer on what AI inference actually is is the shorter read.

Would you trust a 0.7 confidence score to route a customer's claim, or does anything below 0.9 go to a person? Write to us via our contact page.