Google DeepMind launches Gemini Decisions API and Gemini Decision 1 for thousand-question scoring
Google DeepMind launched the Gemini Decisions API and Gemini Decision 1 for batched closed-ended multimodal scoring across text, images, audio, video, and PDFs rather than open generation. It evaluates up to one thousand questions on one shared multimodal context, returning calibrated probabilities for routing, policy checks, tool selection, and grading. According to Google, responses take about 150 to 250 milliseconds at roughly five times lower cost than Gemini 3.5 Flash-Lite, with MarkTechPost citing $0.10 per million input tokens and no output billing on Google AI Studio, Gemini Developer API v1beta, and Vercel's google/gemini-decision-1 gateway route. Teams may replace generative branch calls with this scoring layer. Excerpts mark latency, cost, and accuracy as vendor-reported, cite API-only hosting without open weights, and flag 58.7 percent Macro-F1 on Amazon ESCI before high-volume adoption.
Google DeepMind launches Gemini Decisions API and Gemini Decision 1 for thousand-question scoring
Google DeepMind is introducing the Gemini Decisions API for fast closed-ended decisions across text, images, audio, video, and PDFs. Gemini Decision 1 processes up to 1,000 independent questions in a single pass and returns results in 150 to 250 milliseconds at roughly five times lower cost than Gemini 3.5 Flash-Lite, according to Google, with availability through Google AI Studio.
Key takeaway
Gemini Decision 1 reframes high-volume agent branching as a cheap batched scoring API, but packet evidence treats Google's speed, cost, and accuracy claims as unverified.
What happened
According to The Decoder and the packet headline, Google DeepMind launched the Gemini Decisions API with Gemini Decision 1 to handle fast closed-ended decisions across text, images, audio, video, and PDFs, processing up to one thousand independent questions in one pass with availability through Google AI Studio.
MarkTechPost and Techmeme excerpts in the packet say the API returns predefined answers with calibrated probabilities rather than generated text, priced at $0.10 per million input tokens, targeting routing, policy checks, tool selection, and grading inside agents via the Gemini Developer API v1beta without self-hosting or open weights.
Evidence
Gemini Decision 1 can evaluate up to 1,000 closed-ended questions in one multimodal pass with Google-cited 150–250 ms latency.
The Decoder · attributed
Gemini Decision 1 returns results in 150 to 250 milliseconds at roughly five times lower cost than Gemini 3.5 Flash-Lite, according to Google.
The Gemini Decisions API scores questions and returns calibrated probabilities without generating text.
MarkTechPost · attributed
The Gemini Decisions API does not generate text. It scores up to 1,000 typed questions against one shared input and returns calibrated probabilities.
Google prices the API at $0.10 per 1M input tokens and bills zero output tokens.
MarkTechPost · attributed
Google prices it at $0.10 per 1M input tokens, with zero output tokens billed.
Techmeme reports the API returns predefined answers with confidence scores for up to 1,000 yes/no or choice options.
Techmeme · attributed
Evaluate up to 1,000 yes/no, choice
Vercel lists Gemini Decision 1 on AI Gateway for routing, policy checks, and tool filtering without output tokens.
Vercel AI · attributed
Gemini Decision 1 from Google is now available on AI Gateway. The multimodal decision model works with text, images, audio, video, and PDF documents, returning Boolean probabilities, choices, and scores.
Why it matters
Teams building agents may replace repeated generative classification calls with a dedicated decision endpoint, shifting cost and architecture choices for routing and policy layers.
Limits and uncertainties
Packet excerpts do not establish independent benchmarks or confirm production-wide general availability beyond Google-hosted API access.
Latency, cost, and accuracy figures cited in the packet are vendor-reported; MarkTechPost notes mixed benchmark results including 58.7% Macro-F1 on Amazon ESCI and trailing Jev 1.14 on When2Call.
Practical implications
Operators can pilot google/gemini-decision-1 for batched yes/no and choice scoring inside agents, but should benchmark their own tasks before cutting over from general models.
Builders needing on-prem or open weights cannot self-host per MarkTechPost; adoption depends on accepting Google API-only hosting and v1beta access paths.
What to watch
Independent latency, calibration, and accuracy tests on representative agent routing and policy workloads.
Google publication of GA status, regional availability, and updated pricing relative to competitors such as Jev on the shared benchmark suite cited in MarkTechPost.