Google Unveils Gemini Decision 1, a Model Built to Process 1,000 Questions for $0.10 per Million Input Tokens
Image to help understand the article
Google has introduced Gemini Decision 1, a specialized artificial intelligence model designed to evaluate large batches of independent questions in a single operation, with pricing starting at 10 cents per million input tokens and no charge for output tokens.
Announced Nov. 11, 2026, the model is available through Google's new Gemini Decisions API under the model ID gemini-decision-1. Google says it can process up to 1,000 independent questions in a single forward pass, potentially reducing the cost and latency of applications that need to classify, rank or evaluate many items simultaneously.
Low-cost processing and developer access
Developers can experiment with the model through the free tier of Google AI Studio. Google also plans to add support through Gemini on Google Cloud. The model is additionally available through OpenRouter and Vercel AI Gateway.
Google reports end-to-end response times of 150 to 250 milliseconds, with median model computation times of 50 to 100 milliseconds. Based on a benchmark package it shared, the company claims Gemini Decision 1 through its Decisions API costs nearly five times less than Gemini 3.5 Flash-Lite through the Interactions API.
For comparison, the stated input prices are 4.2 cents per million tokens for Jev 1.14 and 10 cents for uncached input tokens with GPT-6 Luna.
Designed for decisions, not open-ended generation
Gemini Decision 1 supports up to 16,000 tokens of context and can accept as many as 255 answer choices for a single multiple-choice question. Google also says that all questions in one request share a context window of 262,144 tokens.
The model has important limitations. It cannot handle tasks requiring explanations or arbitrary data generation. Each question is evaluated independently, so questions whose answers depend on the results of other questions must be submitted in separate API calls.
Google's benchmarks show strong results, but not a clean sweep
Google says Gemini Decision 1 ranked first in five of seven tests included in its launch benchmark results, covering agent routing, reasoning, ranking and multimodal understanding. Its JevBench score was 92.2%, ahead of GPT-6 Luna at 90.9% and Jev 1.14 at 90.0%.
The model also recorded a calibration error of 0.0250, compared with 0.0494 for Jev 1.14. Calibration error measures how closely a model's confidence matches the actual reliability of its answers, with lower scores indicating better calibration.
However, Google's own results showed competitors outperforming it in some categories. Jev scored 86.8% on Mind2Web Multimodal, compared with 86.2% for Gemini Decision 1. On When2Call, Jev also led with 82.0%, against Google's model at 79.9%.
Gemini Decision 1 scored 71.8% on the MMAU audio benchmark, exceeding Gemini 3.5 Flash-Lite's 67.7%. On Video-MME, it achieved 69.1%, compared with 63.3% for the comparison model.
Independent evaluation raises questions about cost competitiveness
A separate assessment by Artificial Analysis, using its Decision Index v2.1, placed Gemini Decision 1 ninth among 18 decision-focused models available at launch, with a score of 66. The model achieved the highest score among decision models in the index's knowledge and reasoning evaluations.
But it trailed competitors in several other independent tests. Drex Max v1.1 achieved a 63% pass rate on AA-Paperclip, compared with 33% for Gemini Decision 1. In Artificial Analysis' own repeat measurement of When2Call, Drex Max scored 95.3 against Google's model at 80.5. Mammoth 27B also outperformed Gemini Decision 1 on AA-Symmetry, scoring 88.2 versus 70.5.
Cost measurements presented a further challenge to Google's value proposition. Artificial Analysis estimated Gemini Decision 1's cost at $345 per million tasks, substantially above Drex Max v1.1 at $103 and Jev 1.14 at $156. Its measured processing time per task was 211 milliseconds.
The results suggest Gemini Decision 1 is aimed at a specific segment of the AI market: applications that need rapid, inexpensive evaluations across large numbers of independent questions rather than conversational responses or open-ended content generation. Google's pricing and batch-processing approach could appeal to developers building automated decision systems, although independent benchmarks indicate that performance and operating costs vary considerably by task.
Comments
Post a Comment