Gemini Decisions API: $0.10 per million input tokens

What changed

Google announced the Gemini Decisions API on November 11, 2026, introducing Gemini Decision 1 model designed for bounded classification tasks processed in a single forward pass. Instead of generating text token by token, this API accepts up to 1,000 independent questions about multimodal inputs and returns probability scores and confidence metrics.

The API charges $0.10 per million input tokens with no output token costs. Input is billed once per request regardless of question count, since each additional question only adds its instruction and criteria tokens. A no-cost tier is available through Google AI Studio for prototyping.

Gemini Decision 1 returns decisions in 150-250 milliseconds end-to-end latency across multimodal inputs up to 256K tokens, with median model compute time of 50-100 milliseconds. The API supports three question types: yes/no evaluations with optional description criteria, multiple-choice selection from 2 to 255 options, and ordered rubric scoring that grades inputs on a continuous scale with per-level probabilities. Every answer includes a confidence score from 0 to 1 for flagging ambiguous cases for fallback routing or human review.

On benchmarks spanning agent routing, reasoning, ranking, and multimodal understanding, the model leads with the lowest calibration error. Latency remains nearly flat from 1 to 100 questions per request. Gemini Decision 1 with the Decisions API costs nearly 5x less across shared benchmark tasks than Gemini 3.5 Flash-Lite with the Interactions API.

The API is available through Google AI Studio and supports ecosystem integrations rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode, with support coming soon to Gemini in Google Cloud.

Why it matters

This launch addresses a gap for teams that need fast, reliable classification rather than open-ended text generation. Teams within Google are already using the API: Gemini Notebook uses it for real-time model routing to reduce latency up to 40 percent on everyday queries, Stitch generates playable wireframes in seconds, and Google Search employs it as a safety layer to detect adversarial behavior.

For developers building agents and workflows, the Decisions API removes the need to train custom classifiers, maintain routing logic, or wait for token-by-token generation on simple yes/no or ranking decisions. Since the API charges only for input tokens with no output token cost, the economics shift dramatically for high-volume classification workloads. The confidence scores enable graceful fallback to human review or escalation routes when the model is uncertain.

The multimodal capability matters for businesses that need to classify content across documents, images, and audio in one call. A financial services firm could score a claims photo with supporting documents. A content platform could route a submission through policy checks based on text and video. An HR department could grade interview recordings alongside resumes. These tasks previously required chaining multiple API calls or using slower text-generation models not optimized for bounded decisions.

The speed and single-pass design open use cases that were too slow before. Coding agents and automated workflows that depend on rapid routing decisions now have a dedicated tool instead of relying on inference latency from general-purpose models. CC from Google Labs uses the API to select the right emoji reaction from over 200 options in a fraction of a second, enabling reactions to appear instantly without latency delays.

What to test

Before adopting the Decisions API in production, validate these claims:

• Latency benchmarks in your workload. The API reports 150-250ms end-to-end latency, but measure with your own input sizes, question counts, and modality mix. Test whether latency remains flat as you scale from 1 to 100 questions.

• Confidence calibration. Evaluate whether confidence scores from Gemini Decision 1 match your actual error rates in your domain. A score of 0.8 should mean the decision is correct roughly 80% of the time.

• Cost per decision versus alternatives. Compare token usage when using the Decisions API versus general-purpose models like Gemini 3.5 Flash for the same classification task. The cost advantage depends on your specific use case and question complexity.

• Fallback handling. Set thresholds for when you escalate low-confidence results to human review. Test how often this happens and whether it aligns with your service level requirements.

• Multimodal fidelity. If combining image, video, audio, and PDF in one input, verify that the model captures the signals you need. Adjust image and video resolution settings to balance token cost against decision quality.

• Integration availability. If using LangChain, Vercel, or Pydantic AI, validate that the ecosystem integration matches your pipeline. The announcement lists some integrations as “rolling out” or “coming soon,” so check current availability for your stack.

The conclusion

The Gemini Decisions API represents a practical shift in how teams approach classification tasks. By decoupling bounded decisions from text generation, Google has created a tool that trades generality for speed and cost predictability. Pricing at $0.10 per million input tokens with zero output charges appeals to high-volume routing, grading, and policy-checking workloads where confidence matters as much as speed.

The adoption inside Google suggests real utility: model routing in Gemini Notebook, rapid prototyping in Stitch, and safety checks in Search. These are pragmatic applications that benefit from millisecond-scale latency and straightforward go/no-go decisions rather than reasoning or creativity.

The main question for practitioners is whether this tool solves your routing or grading bottleneck better than your current stack. If you are chaining multiple API calls or waiting on slower models to make simple choices, the speed and pricing math probably works in your favor. If your classification tasks require complex reasoning or open-ended output, the Decisions API is not designed for that job.

Watch for actual latency and cost data from early adopters, especially on multimodal requests. Track availability of ecosystem integrations, since some are listed as rolling out or coming soon to Google Cloud. Early testing in Google AI Studio can establish baselines against your current workflows before committing to production use.

Related reading

• Gemini 4 Carbon vs GPT-6.1 Sol: Pricing and Performance
• Gemini Omni 1.2 Flash: 30% price cut, 15-second video
• Soniox STT v6: Agent text input and echo filtering