Two dollars for 10,000 AI judgments: an illustrative calculation

An AI system sorting a customer request sometimes only needs to choose between a quote, a delivery question and everything else. Google's new Gemini Decisions API makes these bounded judgments a product of their own, priced at $0.10 per million input tokens. That creates a concrete economic opportunity: more ordinary requests could get an initial assessment, with longer reasoning reserved for the work that benefits from it.
This week's podcast draws on Hammer's provider research from November 9–15. Here we explore one question from that material: how much of a workflow really needs a long AI answer?
Gemini Decisions returns a choice with probabilities
Google launched the Decisions API on November 11. Its Gemini Decision 1 model evaluates predefined options and returns probabilities, yes/no judgments or scores against a supplied rubric. It can work with text, images, audio, video and PDFs. An API is the connection that lets other software use the model.
An office could assess the same incoming message for its request type and whether the customer mentions an order number. The questions are evaluated independently. Software then uses the results to decide what happens next. Selecting the label “delivery question” provides neither a delivery update nor access to an order by itself.
That suits a recurring job with clear options. A thoughtful reply to an unusual complaint needs more context and a different kind of work.
Source: Google: Gemini Decisions API.
An illustrative bill: 10,000 requests for $2 in model charges
Suppose a business wants to classify 10,000 requests. Each call averages 2,000 billed input tokens, including the instructions and questions. Tokens are the units a model uses to process content; for text, they are often parts of words.
That adds up to 20 million input tokens. At the published rate of $0.10 per million, the model charge is $2. The Decisions API bills zero output tokens.
This is a classification cost example. Integration, storage, any additional calls and staff time add to the bill. It also assumes each request fits within the stated input size. The calculation says nothing about how many judgments will be correct.
The small model charge still changes where attention belongs. If routing works, the business could assess the full incoming workload and find requests that otherwise sit unanswered. If it frequently sends customers to the wrong person, the cost of correcting those mistakes could dominate.
Source: Google: Gemini API pricing. Price checked November 16, 2026.
Hydrogen and Carbon divide the longer work
Google launched Gemini 4 Hydrogen on November 12 for tasks including coding and workflows where AI uses tools. In Antigravity, Orchestrator can switch models during the same assignment: Carbon or Argon can plan, Hydrogen can carry out the work, and the first model can take over again when the plan needs to change.
This is a concrete example of dividing AI work by task. Decisions supplies bounded judgments; Hydrogen and Carbon can generate content and carry out longer workflows. The products fill different roles and need to be compared through the result of the whole assignment.
For a business, the useful question is which part deserves more reasoning. In the hypothetical office, sorting can be cheap while a complicated customer reply gets more time. A simple setup with one model may still be the best choice when volume is low or integration is expensive.
Source: Google Antigravity: Gemini 4 Hydrogen and Orchestrator.
What becomes worth assessing more often?
Hammer's thesis is that cheap, bounded AI judgments could make it reasonable to assess more ordinary requests. That might mean checking whether a request is complete, which service it concerns or whether it needs closer attention. The benefit comes when that judgment improves the work that follows.
Choose a recurring flow that someone already sorts by hand. Use past requests with known outcomes and compare how many reach the right destination, how much correction they need and what the whole process costs. That lets you judge whether a cheap assessment actually frees up time.
The English podcast provides the broader review of OpenAI, Anthropic Claude, Google Gemini/DeepMind, SpaceX/xAI Grok and Mistral. The research also includes recovered older updates with their original dates preserved.
The episode is an AI-generated masterclass made with NotebookLM from Hammer's daily deep research into AI providers' updates and features. This article draws on the research material and a written NotebookLM synthesis.
The Forge newsletter
Get new articles in your inbox
Pick the topics you care about. No noise, at most one email a week.
We follow GDPR. Unsubscribe anytime.


