Skip to main content
TokenCost logoTokenCost
Model ReleaseNovember 17, 2026·7 min read

Gemini Decision 1 charges $0.10 in and nothing out, the same as OpenAI's Decisions API. OpenAI has since added a cache, so a long rubric costs 4.6x more on Google.

Google matched OpenAI's price to the cent and skipped the one feature that made OpenAI's endpoint cheap for real classifiers. Whether that matters depends on how long your prompt is.

Traffic light at night with red, yellow and green lamps all glowing against a black background

Photo by Drew Beamer on Unsplash

Google shipped gemini-decision-1 and a Decisions API on November 11. Like every decision model we've covered since October, it doesn't write text. You send content plus questions, and it returns probabilities: yes or no, one of up to 255 options, or a score on a scale. Google's price page lists input at $0.10 per million tokens for every modality and output as "not applicable".

That's exactly what OpenAI charges for GPT-6 Luna on its own Decisions endpoint. The difference is a line further down. Gemini Decision 1 has no context caching. OpenAI's endpoint has had one since it left beta on October 22, eight days after our October 14 comparison, and it flips that post's conclusion about rubrics.

  • Small prompts: same price as Luna Decisions, six points higher on Artificial Analysis.
  • A long fixed rubric in front of every call: $230 per million calls on Google, $50 on OpenAI with the cache warm.

The card, next to the two it's chasing

Gemini Decision 1GPT-6 Luna DecisionsMicrosoft-Decision-1
Input / 1M$0.10, all modalities$0.10$0.042
Outputfreefreefree
Cached inputnot available$0.01 read, $0.125 writenot listed
Questions per request1 to 1,000up to 400several, no stated cap
Context262,1441.05M32,768
Batchnonot in the guidenot listed
Statusv1beta, Nov 11GA since Oct 22Foundry preview

Google's Decisions API docs and pricing page, OpenAI's Decisions guide and pricing, read November 17, 2026. Microsoft's price and multi-question support are from third-party listings; we couldn't find them on a Microsoft page.

Google bills shared input once per request no matter how many questions ride on it, and each question adds only its own instruction and option tokens. OpenAI works the same way in practice: the shared input is the prefix, questions are billed as ordinary input. So "ask 50 things about one document" costs the same on both, as long as it fits in one request. Gemini's ceiling is 1,000 questions to OpenAI's 400, which only matters if you're running very long checklists.

Images, video, audio and PDFs all bill at the same $0.10. That's a third of Gemini 3.5 Flash-Lite's $0.30, and under the $0.15 Cloudflare charges for Clef-omni, the other decision model that takes audio and video. Audio goes in directly, without a separate transcription step, and the docs say only input tokens are billed.

The cache OpenAI added on October 22

When Luna Decisions launched in beta it billed every input token at $0.10, cached or not. The October 22 changelog entry that made it generally available also added prompt caching of the shared input. The guide doesn't print its own rates; it points to Luna's standard ones, so a cache read costs $0.01 and a 30-minute write $0.125. The prefix has to be at least 1,024 tokens.

Most production classifiers aren't a bare ticket. They carry label definitions, examples and edge-case rules, and that part never changes. Here's the case we priced last month: a 2,000-token rubric, a 300-token item, one question.

Per million callsCostHow
GPT-6 Luna Decisions, rubric cached$50.002,000 at $0.01, 300 at $0.10
Microsoft-Decision-1$96.602,300 at $0.042
Gemini 3.5 Flash-Lite, explicit cache$207.502,000 at $0.03, 300 at $0.30, 23 out at $2.50
Gemini Decision 1$230.002,300 at $0.10
GPT-6 Luna Decisions, no cache$230.002,300 at $0.10

Our arithmetic at list prices, cache assumed warm. The Flash-Lite row is an ordinary generative call with 23 output tokens and leaves out cache storage ($1 per million tokens per hour, about $0.002 an hour for this rubric).

$230 against $50. Gemini Decision 1 is 4.6 times the price of Luna Decisions here, and it even loses to Google's own Flash-Lite run as a plain chat model with an explicit cache. We didn't expect that row. Flash-Lite is badly calibrated by Google's own numbers, so it's not a real substitute, but it shows how much the missing cache costs.

The rule is easy to remember. With a warm cache, every million fixed tokens costs $0.09 more on Gemini than on Luna Decisions. If your reusable prefix is under 1,024 tokens, OpenAI can't cache it either and the two cost the same.

Short prompts: same bill, better model

The 356-token support ticket we've used across this series costs $0.0356 per 1,000 on Gemini Decision 1 and on Luna Decisions, and $0.015 on Microsoft-Decision-1. Nothing to cache at that size, so the tie is real.

On quality, Gemini wins that pair. Artificial Analysis has it at 66 on its Decision Index against Luna's 60, with nearly the same cost per task ($345 against $339 per million) and latency (211 ms against 206 ms). The cost gap is just tokenizer noise, because both charge $0.10. If you're already paying OpenAI's rate for short items, Google is a free six points.

One OpenAI wrinkle to check if you send long one-off documents: the guide says cache writes are charged and that the breakpoint sits at the end of the shared input by default. If a document you'll never send again gets written at $0.125, Luna is up to 25% dearer than Gemini on that call. Look at the usage block on your first few requests before you budget.

Fourteenth of 24, at OpenAI's price

ModelIndex$ / 1M tasks$ per pointMedian ms
Drex Max v1.1 (launch price)73$103$1.41283
GLiDE v1.273$780$10.68284
Mammoth 27B71$264$3.72106
Liquid d1-ultra71$133$1.87301
TypeSafe Jev 1.1466$156$2.3692
Gemini Decision 166$345$5.23211
Microsoft-Decision-162$143$2.31274
Mercury Decide (launch price)62$70$1.13245
GPT-6 Luna Decisions60$339$5.65206
Cloudflare Clef60$863$14.38354

Artificial Analysis Decision Index model pages, read November 17, 2026. 24 models listed, median $135 per million tasks. Dollars per point is our division. Two rows are on launch pricing that AA says normally doubles.

Gemini Decision 1 lands 14th of 24 and costs two and a half times the board's median. TypeSafe Jev 1.14 posts the same 66 for $156 at less than half the latency. Liquid's d1-ultra scores five points more for $133. On this board Google is paying OpenAI prices for a mid-table score.

For scale: AA's v2 write-up has general models above all of them. Claude Haiku 5.5 at max effort scores 77 for $6,240 per million tasks and GPT-6 Luna at max 76 for $1,530. Gemini Decision 1 gives up 10 or 11 points to those and costs 4.4 and 18 times less per task.

Google's "nearly 5x cheaper", checked

The launch post says Decision 1 is nearly five times cheaper than Gemini 3.5 Flash-Lite across a shared benchmark suite. The post only shows charts. DataCamp transcribed the numbers: $0.88 for Decision 1 across seven evals, $4.14 for Flash-Lite, $0.93 for GPT-6 Luna and $0.68 for Jev 1.14. That's 4.7x, so the claim holds against Flash-Lite.

Read the same table the other way and Jev is 23% cheaper than Google's model on Google's own chart. Google's case is calibration: top-1 calibration error of 0.025, against Jev's 0.049, Luna's 0.072 and Flash-Lite's 0.21. If you threshold on the probability, say auto-closing tickets above 0.9, that matters more than a few cents. If you only take the top answer, it doesn't.

We treat those figures as vendor-reported. They aren't in the text of Google's post or on a model card we could find.

What else to know before you switch

  • The free tier is generous: 200 requests a minute, 50,000 tokens a minute and 10,000 requests a day, per Google's rate limits. Free-tier data is used to improve Google's products, though, so keep customer content on the paid tier.
  • It's Gemini API only. Google says support in Google Cloud is "coming soon", it wasn't in the Vertex AI model docs as of November 12, and the endpoint sits under v1beta. The changelog never calls it GA, so expect changes.
  • Repeated input across requests counts in full, in Google's words. If the same document gets asked new questions later, send them in one request instead of three.
  • The yes/no type is spelled noul in the docs and SDK, and returns a probability clamped between 0.01 and 0.99. Choice and score types also return a confidence from 0 to 1.
  • Inline payloads go up to 100 MB (50 MB for PDFs). YouTube links and gs:// URIs aren't accepted, so video has to be uploaded or linked from elsewhere.

Who should use it

Teams on Luna Decisions with short prompts. Same bill, better score, and the free tier is enough to test it properly.

Anyone classifying audio or video, where Gemini is the cheapest decision model we know of that takes those inputs, and anyone who needs well calibrated probabilities rather than a top label.

Not teams with a long fixed rubric, who should stay on Luna Decisions with caching on. And if you're choosing fresh for text, start with Jev 1.14 or d1-ultra, which match or beat Gemini's score for less than half the cost per task.

Our first decision-model roundup explains how these models bill and why output was a quarter of a classifier's cost. See GPT-6 Luna's pricing page for its full card, compare generative models on the pricing table, or put your own token counts into the calculator.

Sources