Google’s Decisions API turns Gemini into a control plane, not a chatbot

The next AI cost fight is moving from answers to gates: which model runs, which tool opens, which case gets escalated. Google’s new Gemini Decisions API says it can evaluate up to 1,000 bounded questions against one multimodal input, with 150-250 ms end-to-end latency and pricing at $0.10 per 1M input tokens. The common read is that this is another small Gemini endpoint. That misses the point. Google is carving out a separate layer for machine judgment before generation, where milliseconds and calibration matter more than prose.

Reporting note: this analysis is based on Google’s announcement. The performance, pricing and internal-use examples below are Google’s claims, not independently verified benchmarks.

Google split bounded decisions away from generative answers

The Decisions API is built around closed-set choices: yes/no checks, multiple choice and ordered rubric scoring. Google says Gemini Decision 1 does not generate answer text token by token. Instead, it scores candidate options in a single forward pass and returns probabilities and confidence scores.

That distinction matters because a large share of production AI work is not “write the answer.” It is decide whether a prompt needs a larger model, whether an input violates a rule, whether a document matches a rubric, or whether a tool call should be allowed. Those jobs are classifier-shaped, but teams have often used general chat models because they were available and flexible.

Google is now giving that work its own product surface. One request can combine text, images, audio, video and PDFs, with one shared input and many independent questions. Google says the input can run up to 256K tokens, while its benchmark latency claim is described across contexts up to 16K tokens.

The sentence to carry into the room: Google is trying to make Gemini the thing that decides which AI system gets to act.

The timing is about latency budgets, not model demos

The announcement leans hard on speed and price because control-plane decisions are only useful if they fit inside existing product latency. Google claims median model compute of 50-100 ms and 150-250 ms end-to-end latency, remaining nearly flat from 1 to 100 questions per request in its benchmark description.

That shape is the product. If a developer can ask many policy, routing and scoring questions in one pass, the decision layer stops being a tax on every interaction and becomes infrastructure. Google also says there are zero output tokens billed and that Gemini Decision 1 through the Decisions API is nearly 5x cheaper across a shared benchmark suite than Gemini 3.5 Flash-Lite through the Interactions API.

Those are vendor claims. Still, the direction is clear. The frontier-model story has trained buyers to compare intelligence at the point of response. Google is pushing a different procurement question: how many small judgments can be made cheaply enough to sit in front of every response?

This changes the buyer for AI tooling inside companies

The first users named by Google are not generic chatbot builders. They are teams building routers, safety layers, graders and product controls. Google says Gemini Notebook uses the API to route incoming queries to model tier and thinking level, with everyday queries seeing up to 40% lower latency. It says Google Search uses the Decision API to detect adversarial behavior and content against adversarial evals within a latency envelope.

That points to a different internal owner. Platform teams, trust and safety teams, support automation groups and AI infrastructure leads are the natural audience, because they own the decision points where generative systems become expensive or risky.

For startups building evaluation, routing or policy middleware, this is also a warning shot. Google is packaging a native multimodal scorer into the same ecosystem where developers already call Gemini. The defensible layer shifts from “we can call a model to classify this” to workflow design, auditability, policy management and proof that the decisions hold up in a customer’s domain.

The missing proof is calibration outside Google’s own frame

Google says Gemini Decision 1 leads across benchmarks spanning agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. The announcement does not provide enough independent detail here to treat that as established fact.

The unknowns are the important part. We do not know how calibration behaves in messy enterprise taxonomies, how probabilities drift across domains, how customers should audit rubric scores, or how the API performs at the stated 256K-token ceiling. We also do not know whether developers will trust a probabilistic decision layer for high-stakes policy calls without clearer monitoring and appeal mechanics.

Confidence scores are useful only if they are reliable at the boundary. A bad answer is visible. A bad gate can be silent.

Watch whether decisions become the new default wrapper around agents

The trigger is adoption beyond Google’s own examples. If developers start putting Decisions API calls before tool use, model selection, safety checks and content grading, then Google will have created a low-friction layer between user intent and AI action.

That is more consequential than another chatbot feature. Agents need brakes, routers and judges. Google just made the case that those should be fast, cheap, multimodal and sold as a primitive.

What remains unverified

• Google’s benchmark claims have not been independently verified in the supplied source.
• The supplied source does not show how calibration performs in customer domains outside Google’s described benchmark suite.
• The supplied source does not specify production availability terms beyond access through Google AI Studio and the Gemini API.
• The supplied source does not explain auditing, appeal or monitoring mechanics for high-stakes decisions.

Sources and reporting note

• Gemini Decisions API: Fast, calibrated multimodal decisions

Reporting note: This analysis was generated by an automated workflow using the cited public primary sources, source checks, and a separate AI review. Company statements are attributed claims; no independent product testing or interviews were conducted.

Originally published at Sift & Signal.