Decision Node Pricing, API, and CLI Access Explained
Decision Node charges a few cents per million tokens, output is free, and batch jobs cut costs further. Here's how pricing, API, and CLI work.
What is Decision Node and why does its pricing model look different?
Decision Node is a model built to make decisions, not to chat. Instead of generating paragraphs of text, it takes in text or images along with a set of typed questions and returns structured answers: a choice, a score, a probability, a number, or even pixel coordinates pointing at an object in an image. Every answer ships with a calibrated confidence value.
That design choice is also why its pricing looks unusual compared to a typical LLM API. Because the model never writes out free-form text, there’s effectively no output generation cost in the traditional sense, and output tokens are free. You pay for what goes in, not what comes out. According to the model’s own developer walkthroughs, the price lands at a few cents per million tokens, with further discounts available for batch processing.
TL;DR
- Output tokens are free because Decision Node returns typed answers (choices, scores, probabilities, coordinates) instead of generated text, so you’re not billed for response length.
- Input pricing runs a few cents per million tokens, which is cheap enough to support high-frequency use cases like routing, moderation, and triage that fire thousands of times a minute.
- Batch jobs support up to 10,000 requests at once at a lower per-token price, making it practical to run large backlogs of classification or scoring work cheaply.
- Access comes through three paths: a REST API for direct integration, a CLI installable with a single command line, and an official Python SDK where one line of code returns a probability.
- Boosters act as specialist pre-processors (human age detection, face detection, AI-image detection, OCR) that feed evidence to the main model before it decides, and they factor into how a request is processed.
- A knowledge base feature lets you upload your own rules (PDF, text, or other formats) so the model grounds its decisions in your documentation via retrieval rather than needing those rules pasted into every request.
- Response times are fast, with one demonstrated prompt-injection test resolving in under 78 milliseconds, which matters for anything gating a live user action.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How does Decision Node’s pricing actually work?
The core idea is that you’re charged for input tokens, the text, image, and context you send in, at a rate described as a few cents per million tokens. Output is free because the model isn’t generating prose. It’s returning compact structured data: a true/false flag, a 0 to 1 confidence score, a category label, or a bounding box. Even a complex multi-question request, like asking a model to classify urgency, flag a callback, and estimate hours late from a single complaint, still resolves to a handful of typed fields rather than a wall of text.
For teams running high request volumes, that structure matters more than it might first appear. A moderation pipeline, a routing system, or a triage queue might issue thousands of calls per hour. If every one of those calls carried the cost of a generated explanation, the bill would scale with verbosity. Here it scales with input size and request count instead.
Batch jobs add a second discount layer. Decision Node supports submitting batches of up to 10,000 requests at a lower per-token price than real-time calls. This fits workloads like backfilling labels on an old dataset, scoring a large archive of support tickets, or running compliance checks across historical records, cases where you don’t need millisecond latency and can tolerate asynchronous processing in exchange for a lower rate.
How do you access Decision Node: API, CLI, or SDK?
There are three supported entry points, and which one makes sense depends on how you’re building.
REST API. You send a JSON payload containing your input (text, image, or both) and a set of typed questions, and the API returns typed answers with confidence scores. A demonstrated example sent a site foreman’s complaint along with four questions (which team should handle it, how urgent, whether to call back immediately, how many hours late) and got back a structured response: routing decision at 0.99 confidence, an urgency score of 2.82 out of 3, and a callback probability of 0.92, derived correctly from the text. This is the path for backend integration into existing applications, pipelines, or automation logic.
CLI. Decision Node ships a command-line tool installable with a single one-liner. After installing and authenticating with an API key, you can run commands like dn ask to pose a yes/no or scored question directly from the terminal. In one example, asking the CLI to judge how believable a kid’s homework excuse was returned a calibrated probability of 0.1, meaning roughly a 1-in-10 chance the model considered it true. The CLI is suited to quick checks, scripting, and testing prompts before wiring them into production code.
Python SDK. An official SDK is also available, and a single line of code is enough to get a probability back from the model. This is the most direct path for developers who want to embed decision calls inside existing Python applications without manually managing HTTP requests.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
All three paths hit the same underlying model and support the same typed output formats: choices, scores, probabilities, numeric values, and image coordinates.
What are boosters and do they affect cost or accuracy?
Boosters are specialist sub-models that run before the main decision model sees the input. Each booster reads the raw input (an image, for instance) and hands its findings to the main model as evidence. Demonstrated boosters include human age estimation, face detection, AI-generated image detection, and OCR for reading text embedded in images.
In one test, an AI-generated image of a person holding a sign was fed through boosters for age, face detection, AI-image detection, and OCR simultaneously. The AI-detection booster flagged the image as likely generated, the OCR booster read the text on the sign correctly, and the main model combined that evidence to reach a final verdict, correctly identifying the image as fake despite it looking convincing. Boosters are selected per request, so you only enable the ones relevant to your use case, and they’re part of what makes the model’s confidence scores meaningful rather than guesses.
Is Decision Node worth it for production use?
For workloads that genuinely need a decision rather than a conversation, the economics look favorable. Free output tokens combined with a few-cents-per-million input rate make it cheap to run at the volumes typical of routing, moderation, and triage systems, where a model might be called on every incoming message, image, or event. The addition of batch discounts extends that to offline or backlog processing.
The tradeoff is that Decision Node isn’t a general chat model. It won’t write you an email or summarize a document in prose. It’s built specifically to answer typed questions with calibrated confidence, which is a narrower but arguably more auditable use case for systems where code needs to act on a number, a label, or a coordinate rather than parse free text. Demonstrated tests, including a prompt-injection attempt where a hidden instruction inside a customer message tried to get the model to approve a refund, showed the model correctly identifying the injection and refusing to comply, all within roughly 78 milliseconds. That combination of speed, typed output, and confidence scoring is the core value proposition, and the pricing structure is built to support exactly that kind of high-frequency, low-latency decision workload.
Frequently Asked Questions
How much does Decision Node cost per request?
Pricing is based on input tokens at a rate described as a few cents per million tokens. Output tokens are free since the model returns structured answers (scores, probabilities, labels) rather than generated text.
What is the batch job discount?
Decision Node lets you submit up to 10,000 requests in a single batch job at a lower per-token price than standard real-time calls, which suits large offline processing tasks like backfilling labels or scoring historical data.
Can I use Decision Node without writing code?
Yes, through the CLI. After a one-line install and logging in with an API key, you can issue commands like dn ask to pose questions directly from the terminal and get back typed, confidence-scored answers.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
What’s the difference between the API, CLI, and SDK?
The API is for direct integration into applications or pipelines via HTTP requests. The CLI is a terminal tool for quick one-off queries or testing. The Python SDK wraps the same functionality into a library where a single line of code returns a result like a probability.
Does Decision Node support custom rules or private data?
Yes, through a knowledge base feature. You can upload documents (PDF, text, and other formats), and the model indexes them automatically so it can ground its decisions in your specific rules instead of relying only on what’s included in each individual request.