Editorial illustration for Google Launches Gemini Decisions API at USD 0.10/M Tokens, 50-100 ms Compute
Google's Gemini Decisions API: $0.10/M Tokens
Google rolled out a new API on its Gemini Developer platform this week that does not write a single word of output. The Gemini Decisions API, built on a model called Gemini Decision 1, takes one shared input and up to 1,000 yes/no or multiple-choice questions, then returns a probability score for each. No generated text, no output tokens billed, just numbers your code can branch on directly.
The pitch is aimed squarely at the plumbing inside AI agents, the routing decisions, policy checks, tool selection and grading steps that typically get bolted onto a general-purpose chat model as an afterthought. Google is pricing the new model at $0.10 per 1 million input tokens, with a context window of 262,144 tokens covering both the shared input and every question asked against it. The model reads text, images, video, audio and PDFs in a single prefill pass, and it's available only through Google's hosted API, there are no open weights and no option to self-host.
Google published benchmark numbers alongside the launch, putting Gemini Decision 1 up against rival models on accuracy and calibration across nine separate tests.
The Gemini Decisions API does not generate text. It scores up to 1,000 typed questions against one shared input and returns calibrated probabilities. For developers, this targets the small branching calls inside agents: routing, policy checks, tool selection and grading.
Why this matters
For teams building agents, the cost math here is the real story. Routing, policy checks and tool selection happen constantly inside any agentic loop, and until now developers were either paying full generation prices for yes/no decisions or hand-rolling classifiers to avoid it. At $0.10 per million input tokens with no output billed, and compute holding steady from 1 to 100 questions against a single shared input, Google is betting that branching logic is frequent enough and cheap enough to become its own product category rather than a side effect of a chat model.
We'd want to see how Gemini Decision 1's calibrated probabilities actually behave under adversarial or ambiguous inputs before trusting it for anything like content moderation or financial routing. A scoring model that's wrong with confidence is arguably worse than a slow one that hedges. The 262K context window and flat latency up to 100 questions are the numbers worth testing first. If those hold up outside Google's own benchmarks, this could change how cheaply agents make small decisions at scale.
Related Reading
Browse all Google news →
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI
From Archives US corporate AI spending hits a new high in October, growth slows to 1.65%
From Archives TSMC board approves USD 500M OpenAI and USD 500M Anthropic investments
From Archives DeepMind's AlphaProof Code Verified Signal's Cryptographic Core in Four Weeks
Research & Benchmarks
From Archives Gemini Decisions API cuts costs nearly 5x by billing zero output tokens
LLMs & Generative AI
Common Questions Answered
What is the main difference between the Gemini Decisions API and traditional text-generating APIs?
The Gemini Decisions API does not generate any text output. Instead, it scores up to 1,000 yes/no or multiple-choice questions against a single input and returns calibrated probability scores that code can branch on directly. This eliminates the need to pay for output tokens since no generated text is produced.
How does the pricing model of Gemini Decisions API at $0.10/M tokens benefit developers building AI agents?
The $0.10 per million input tokens pricing with zero output token charges makes routing, policy checks, and tool selection decisions significantly cheaper than using traditional text-generating APIs. Previously, developers either paid full generation prices for simple yes/no decisions or had to build custom classifiers to avoid these costs, making the Decisions API a more cost-effective solution for frequent branching logic inside agentic loops.
What specific use cases within AI agents does the Gemini Decisions API target?
The Gemini Decisions API is designed for the internal plumbing of AI agents, specifically targeting routing decisions, policy checks, tool selection, and grading tasks. These branching operations happen constantly within agentic loops and require fast, efficient decision-making without generating text output.
What is the performance range for the Gemini Decisions API's compute time?
The Gemini Decisions API maintains consistent compute performance between 50-100 milliseconds regardless of whether you're scoring 1 to 100 questions against a single shared input. This fast and predictable latency makes it suitable for real-time decision-making within AI agent workflows.
How many questions can the Gemini Decisions API process simultaneously?
The Gemini Decisions API can handle up to 1,000 yes/no or multiple-choice questions processed against one shared input in a single request. This capability allows developers to batch multiple routing, policy, and selection decisions efficiently while maintaining consistent compute performance.
Further Reading
- Gemini Decisions API: Fast, calibrated multimodal decisions - Google Blog
- Gemini Decision - Google AI for Developers
- Decisions API - Google AI for Developers
- Google Introduces Gemini Decisions API - MarketScreener
- Gemini Decision 1 now available on AI Gateway - Vercel
