Google priced this model before it praised it. The Gemini Decisions API costs $0.10 in US pricing per 1M input tokens, which is $0.14 CAD, and it bills $0 for output tokens. The model behind it, Gemini Decision 1, writes no text at all. It scores up to 1,000 bounded questions against 1 shared input in a single pass, and it returns a probability and a confidence score for each answer. The held detail is the limit that decides where it fits in a Canadian firm. It runs only on Google’s hosted API today, and support inside Google Cloud is still listed as coming soon.
The direct answer: Gemini Decisions API launched on November 11, 2026 through the Gemini API in Google AI Studio. It takes 1 multimodal input of up to 256K tokens across text, images, audio, video and PDFs, plus up to 1,000 independent questions. It answers in 150 to 250 ms end to end, at $0.14 CAD per 1M input tokens, with a no-cost tier for prototyping.
- What the Decisions API actually does
- Where the price holds and where it bends
- The work in a Canadian firm that fits this shape
- The test to run before any workflow moves
What the Decisions API actually does
A normal model call generates an answer token by token. Gemini Decision 1 skips that step. It reads the shared input once, then scores every question in 1 forward pass. Google calls the 3 question types noul for yes or no checks, choice for picking from 2 to 255 options, and score for grading against an ordered rubric that starts at 0.
Every answer carries a confidence score from 0 to 1. That score is the product. Your code can accept a clear answer, send a weak one to a fallback, and route the rest to a person. Google reports median model compute of 50 to 100 ms, with end-to-end latency of 150 to 250 ms, and says latency stays nearly flat from 1 to 100 questions per request at contexts up to 16K tokens.
The benchmark claims need the usual care. Google says the model leads its own suite across agent routing, reasoning, ranking and multimodal work, with the lowest calibration error, and that the pairing is nearly 5x cheaper on that shared suite than Gemini 3.5 Flash-Lite on the Interactions API. Those are the vendor’s figures, on the vendor’s tests. Gemini 4 Carbon is out. The token price is the small number made the same point for a bigger model this week. A claim earns trust when your own cases repeat it.
Where the price holds and where it bends
The shared input is billed once per request, no matter how many questions ride on it. Each question adds only its own instruction and criteria tokens. There are no output-token charges, no thinking-token bills, and no separate transcription or OCR line for audio and image inputs. For a firm that asks 40 checks against the same claim file, that billing shape is the saving.
The bends sit at the edges. Access runs through the Gemini Developer API, version v1beta, and the quickstarts need the google-genai SDK at version 3.2 or later. Google Cloud support is coming soon, which means the enterprise buying path many Canadian firms use is not the launch path. Open Google Vids before you buy Gemini Omni API is a useful habit here. Read what the free and hosted surfaces already do before a contract conversation starts.
The work in a Canadian firm that fits this shape
The fit is triage. A courier photo that needs a damage call. A claim PDF with photos that needs 6 policy checks. An invoice that needs a routing decision and a risk grade. A call recording paired with account notes that needs a yes or no on a disclosure line. Each case is 1 input, a fixed set of questions, and a person waiting on the answer that scores low.
The Canadian condition is data handling. A claim file or a call recording that enters this API enters Google’s hosted service under your firm’s account terms. Your firm stays accountable for that personal information under Canadian privacy law, and the confidence threshold is part of the control. Set the line where a score stops being an answer and becomes a referral, and write that line down before the first real file moves. Six frameworks want to run your agents. Pick the one your team can run half-asleep covers the wider control question. A decision endpoint is narrower than an agent, and that narrowness is its safety case.
The test to run before any workflow moves
No hands-on test was run for this post. Run this one before you move any workflow. Take 100 closed cases from your own operation, cases where a person already made the call and the outcome is known. Send each case with your real question set. Score the model on 3 lines only: how often its top answer matches the recorded call, how its confidence behaves on the cases it gets wrong, and what the 100 cases cost in total.
That third line matters more than the rate card. Stop counting tokens. Even the man who sells them calls the metric fake argues the same discipline from the other side. Here the unit is clean, 1 input read once, so count the finished decisions and the referrals. If the referrals land under your agreed line and the matches hold, the $0.14 CAD rate is a real saving. If the confidence scores bunch in the middle, you have bought a faster queue for your own staff.
The recap is short. Gemini Decision 1 trades writing for judging, at a price that makes high-volume triage cheap enough to test this month. Its limits are a hosted-only launch and vendor-run benchmarks. Both limits point at the same next step.
Run the 100-case test on your own closed files before any live workflow moves to the Decisions API, and read Gemini 4 Carbon is out. The token price is the small number for the cost-per-task check that pairs with it.
