At the System One track of the second annual AIE CODE event, Google executives including Sundar Pichai and Logan Kilpatrick introduced Gemini Decision 1 and a new Decisions API. The model is positioned as a calibrated, multimodal decision engine that scores closed‑set choices in a single pass rather than generating text token‑by‑token.
Capabilities
- Inputs: accepts text, images, audio, video and PDF files.
- Questions per request: up to 1,000.
- Answer types: yes/no, multiple‑choice (up to 255 options) and rubric scoring.
- Context window: specified as 256K.
- Latency: Google and third‑party measurements report end‑to‑end response times of roughly 150–250 ms (the announcement cites ~200 ms).
- Confidence: each answer includes a confidence score.
Pricing: Google states an input cost of $0.10 per 1 million input tokens and $0 billed for output; Sundar Pichai said this is about 5× cheaper than Flash‑Lite.
Origin and release
According to Google, the project began as an internal Google DeepMind prototype. A small team delivered the model and API in about six weeks. Google’s launch reached production before a comparable offering from Anthropic.
Performance and benchmarks
Google’s internal comparisons show Decision 1 outperforming TypeSafe’s Jev 1.14 on five of seven shared benchmarks; for example, JevBench scores were 92.2% for Decision 1 versus 90.0% for Jev. Independent evaluations such as those from Artificial Analysis give more mixed results.
On the Decision Index, Decision 1 scores 66, tying Elephant 9B and Jev 1.14 and trailing Drex Max v1.1 and GLiDE v1.2 by seven points.
Strengths and weaknesses: Decision 1 leads on some tasks (AA‑Omniscience‑Decision 66, HLE‑MultiChoice 49) but underperforms on others; for instance, on AA‑Paperclip it answers every question correctly on only 33% of documents, compared with 63% for Drex Max.
Cost and latency tradeoffs reported by Artificial Analysis show Decision 1 with a high per‑task cost in that evaluation: $345 per 1M tasks versus $103 for Drex Max and $156 for Jev. Its Brier score is 0.18 and average time per task is 211 ms, slower than Jev’s 92 ms in the cited measurements.
Early adopters and ecosystem support
- Stitch used the Decisions API to build a rapid design experiment that converts a rough prompt into a multi‑screen product flow in about five seconds.
- LangChain shipped day‑one support with a base class compatible with TypeSafe classifiers.
Presentations at the event also highlighted a hybrid agent architecture pattern: fast decision models perform tool selection and log extraction while a larger reasoning model handles diagnosis. Cloudflare’s demo demonstrated reducing LLM calls from six to one and shrinking token usage from about 40K to 1.8K using this pattern.
Why it matters
Gemini Decision 1 represents Google’s move to provide production‑grade decision models tailored to closed‑set classification, with a focus on low latency and usage pricing geared toward input‑heavy workloads. Single‑pass decision models can reduce cost and latency for applications such as tool selection, rubric scoring, and high‑volume multiple‑choice tasks—but independent benchmarks and cost/latency comparisons indicate the tradeoffs depend on the workload.
The AIE CODE System One day also showcased related developments across the ecosystem — from Cloudflare models and frontier model evals to DeepMind’s verification and genomics work — underscoring that decision models and agent‑oriented architectures are becoming a central focus for both research and production deployments.
Key numbers and facts
- Model: Gemini Decision 1 (Google).
- Supported modalities: text, image, audio, video, PDF.
- Questions per request: up to 1,000.
- Context: 256K.
- End‑to‑end latency: ~150–250 ms (Google cites ~200 ms).
- Price: $0.10 per 1M input tokens; $0 output.
- Origin: internal Google DeepMind prototype; small team shipped model/API in ~6 weeks.
- Independent evaluations: mixed; Decision Index = 66.
Note
The announcement is a notable step for production decision models, but organizations evaluating Gemini Decision 1 should weigh independent benchmark results, task‑specific accuracy, and observed cost/latency tradeoffs before adopting it in production.
