Google Says Gemini Decision 1 Leads Its Benchmark Suite. Artificial Analysis Puts It Seven Points Behind Two Startups, at More Than Three Times the Cost per Task.
Google launched the Gemini Decisions API and gemini-decision-1 on November 11: one shared input of up to 256,000 tokens in text, image, audio, video, or PDF, plus 1 to 1,000 isolated questions answered in a single forward pass with probabilities and a confidence value and no generated text, at $0.10 per million input tokens with no output charge, no caching, and no Batch API. Google's post says it leads across its own suite with the lowest calibration error. Artificial Analysis scored it the same day on Decision Index v2.1: 66, level with Jev 1.14 and Elephant 9B, seven points behind Drex Max v1.1 and GLiDE v1.2 at 73, and $345 per million tasks against $103 for Drex Max, nearly the same as OpenAI's Decisions API at $339. We read Google's post, docs, and pricing page, and AA's post.
By T.W. Ghost
Google entered the decision-model market today, November 11, 2026, with the Gemini Decisions API and a model called gemini-decision-1, on the Gemini API's v1beta surface with the google-genai SDK at 3.2 or later. The shape is the one OpenAI set in September: one shared input, which here can be text, image, audio, video, or PDF up to 256,000 tokens, plus 1 to 1,000 questions, each isolated from the others, answered in a single forward pass. Three question types: yes or no, which Google calls noul; a choice among 2 to 255 options; and an ordered score against a rubric. Every answer comes back as a probability with a confidence value. Nothing is generated, so nothing is billed as output.
The price is $0.10 per million input tokens on the paid tier, every modality, the same figure OpenAI charges for its Luna-based Decisions API. Google's launch post says the model "leads across the suite, with the lowest calibration error," on a suite Google built, and that it is "nearly 5x cheaper" than Gemini 3.5 Flash-Lite on the Interactions API for the same work. Artificial Analysis scored it the same day on Decision Index v2.1 and used three words Google did not: "not yet frontier." Sixty-six, level with TypeSafe's Jev 1.14 and Elephant 9B, seven points behind Drex Max v1.1 and GLiDE v1.2 at 73, both from startups, and $345 per million tasks against $103 for Drex Max. We read Google's post, the AI Studio docs, the pricing page, and AA's post. Here is what held up.
What You Are Calling
| gemini-decision-1 (November 11, 2026) | |
|---|---|
| Where | Gemini API v1beta; google-genai SDK 3.2+; day-one on OpenRouter, Vercel, LangChain, LiteLLM; Vertex AI "coming soon" |
| Request shape | One shared input up to 256K tokens (text, image, audio, video, PDF); 1 to 1,000 isolated questions |
| Question types | noul (yes or no); choice (2 to 255 options); score (ordered rubric) |
| Output | Probabilities and a confidence value; no generated text |
| Price, paid tier | $0.10 per 1M input tokens, all modalities; output not applicable; shared input billed once per request, each question adds its own instruction and criteria tokens |
| Not available | Context caching; Batch API |
| Free tier | Exists; free-tier data is used to improve Google products |
| Latency | Median 50 to 100 ms model compute, 150 to 250 ms end to end, inputs up to 16K tokens, nearly flat from 1 to 100 questions (vendor) |
| AA Decision Index v2.1 | 66; $345 per 1M tasks |
Zero output tokens is not a cheap task. Drex Max bills $0.03 per million input tokens against Google's $0.10, and on AA's measure that becomes 3.3 times the cost per million tasks. The free-output framing is true, and it does not make the model cheap.
Google matched OpenAI to the cent, and AA's cost column agrees. $0.10 per million input on both, and $345 against $339 per million tasks through OpenAI's Decisions API. Within $6, and no undercut.
No caching, no Batch. A policy document re-sent on every request is billed in full every time. OpenAI's version offers cached input at $0.01.
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| "Leads across the suite, with the lowest calibration error" | Vendor (Google's own suite) | |
| "Nearly 5x cheaper" than Gemini 3.5 Flash-Lite on the Interactions API | Vendor | |
| Latency: 50 to 100 ms compute; 150 to 250 ms end to end | Vendor | |
| Decision Index v2.1 | 66; Drex Max v1.1 (Nace AI) 73; GLiDE v1.2 (Fastino) 73; Jev 1.14 66; Elephant 9B 66 | Independent (Artificial Analysis) |
| Cost per 1M tasks | $345; Drex Max $103; Jev $156; GPT-6 Luna via Decisions API $339 | Independent (AA) |
| AA-Paperclip, all 15 answers right | 33% of documents vs 63% for Drex Max | Independent (AA) |
| Brier score across the Index | 0.18; Drex Max 0.16; Jev 0.19 | Independent (AA) |
| Brier on When2Call | 0.14 vs 0.04 for Drex Max | Independent (AA) |
| AA-Omniscience-Decision; HLE-MultiChoice | Leads at 66; leads at 49 | Independent (AA) |
| $0.50 vs $8.10 per 1,000 records, one call against 20 | Press arithmetic (VentureBeat), not verified |
Three honest observations:
- Vendor suite against independent board. Google's "leads" and AA's "not yet frontier" can both be true, because they are different tests. The one you can read is AA's.
- The calibration claim does not survive the board. "Lowest calibration error" on Google's suite becomes a Brier of 0.18 against Drex Max's 0.16 on AA's Index, and 0.14 against 0.04 on When2Call. The model does lead two of AA's evaluations, Omniscience-Decision and HLE-MultiChoice, the knowledge rows. It is weakest on extraction, where the Paperclip all-15-right rate is 33% against 63%.
- v2.1 is not v2. AA's post is on Decision Index v2.1; our November 8 post was on v2. Drex Max's cost per million tasks read $102 then and $103 now. Check for rescoring before setting the two boards side by side.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
Launched Nov 11, 2026; model id; v1beta; SDK 3.2+ | Verified (Google blog; AI Studio docs) |
| Up to 1,000 questions; 256K shared input; questions isolated | Verified (AI Studio docs) |
| $0.10 per 1M input; no output charge; no caching; no Batch; free tier trains on data | Verified (Gemini API pricing page) |
| Shared input billed once; per-question tokens added | Verified (pricing page and blog) |
| Latency figures; "nearly 5x cheaper"; "leads across the suite" | Vendor |
| AA v2.1 score, costs, Paperclip, Brier, two evaluation leads | Verified (AA post; model listed on AA's index page) |
| Vertex "coming soon"; day-one partners | Verified (Google blog) |
| $0.50 vs $8.10 per 1,000 records | Press arithmetic, not verified |
Who Should Use It
Use Gemini Decision 1 if you are already on the Gemini API and want up to a thousand calibrated answers over one document in one call, at the per-token price OpenAI charges, with the knowledge-heavy rows as its best case.
Skip it for cost, where Drex Max does the Index at $103 per million tasks against $345 and scores seven points higher, and skip it for anything that must stay on Google Cloud terms until Vertex ships; the free tier trains on your data.
The frontier is elsewhere: on AA's decision board it belongs to two startups at 73, and the most calibrated thing AA has measured is still a general LLM at max effort. Our model comparison has the current standings side by side.
Sources
- Google: the Gemini Decisions API
- Google AI Studio docs: Decisions API
- Gemini API pricing
- Artificial Analysis: Decision Index evaluation page
- Artificial Analysis on X: Gemini Decision 1 on Decision Index v2.1
- VentureBeat: Google's Gemini Decision 1 scores up to 1,000 questions in one API call
*Not sure whether your classification job needs a decision model, a calibrated LLM, or neither? Take the free 2-minute quiz and get matched. Then read about the API Google priced to the cent: OpenAI's Decisions API at GA.*