Google's Gemini Decision 1 Scores 1,000 Questions in One Pass

Google DeepMind on Wednesday released a model that scores up to 1,000 questions in a single forward pass, returning probabilities and confidence scores in 150 to 250 milliseconds end to end. The model, Gemini Decision 1, powers a new Gemini Decisions API aimed at bounded choices: tool selection, policy checks, model routing and content grading. It does not write text at all, which is rather the point.

A single request takes one multimodal input of up to 256,000 tokens, mixing text, images, audio, video and PDFs, plus as many as 1,000 independent questions. The API is available now through the Gemini API in Google AI Studio, with a no-cost tier for prototyping.

What can developers actually ask?

Three question types can be mixed in one call, according to Google:

• Yes/no checks that estimate the probability a condition holds.
• Multiple choice across 2 to 255 options, with normalized probabilities for every candidate.
• Ordered rubric scores that return a weighted score plus per-level probabilities.

Every answer carries a confidence score from 0 to 1, so code can route ambiguous cases to a fallback model or a human reviewer. Google's worked example is a photo of a breakroom counter, checked for a spill, matched to a cleanup action and rated for severity, all in one request.

How does the pricing work?

Gemini Decision 1 costs $0.10 per million input tokens, with zero output-token charges, and the shared input is billed once per request regardless of how many questions ride on it. Google claims this makes the model nearly five times cheaper than Gemini 3.5 Flash-Lite on its shared benchmark suite, and says it leads that suite — which spans agent routing, reasoning, ranking and multimodal understanding — with the lowest calibration error. Those are Google's own benchmarks, so readers may apply the usual discount.

The median model-compute time of 50 to 100 milliseconds was measured on contexts up to 16,000 tokens, a long way short of the 256,000-token maximum. Google says compute stays nearly flat from 1 to 100 questions per request. Flat pricing on shared input is the quiet headline here; per-question costs at this rate border on rounding errors.

Where is Google already using it?

The company has deployed the API internally in several places:

• Gemini Notebook routes each incoming query to a model tier and thinking level, which Google says cuts latency by up to 40% on everyday queries.
• Stitch uses it in an experimental canvas that turns voice or tap edits into wireframes almost instantly.
• Google Labs' CC uses it to pick emoji reactions from more than 200 options.
• Google Search uses it as a safety layer to detect adversarial behavior in MCP tools.

An internal routing deployment claiming 40 percent latency cuts is a stronger endorsement than any benchmark chart. Google is, notably, its own first customer.

Who else is shipping scoring models?

Google is the latest large company to ship this kind of model. The Decoder reported that OpenAI, Cloudflare and Microsoft released similar offerings in recent weeks, after the startup Jev introduced its version. Integrations with OpenRouter, LangChain, Vercel and other platforms are rolling out, according to the same report.

The price matches the $0.10 per million input tokens that Meta set for Muse Ember, its small model for high-volume agent work, though Muse Ember is a text generator rather than a scorer. Alphabet also released Gemini 4 Carbon broadly a day earlier, which suggests the company's release cadence now outpaces most vendors' naming conventions.

What comes next?

With four major vendors and a startup now competing on scoring endpoints at identical price points, differentiation will shift to latency, calibration accuracy and integration depth. Google's head start in production routing gives it evidence; rivals will need benchmarks nobody can dismiss.