Key Takeaways
- ScaleDown announced Constrained Decoding Models on October 13, 2026: three task-specific models called Classify, Extract and Summarize that read input once and skip the token-by-token generation loop.
- Classify returns a calibrated probability and a reason per label, Extract returns JSON with confidence scores and citations, and Summarize returns a consistent summary under 1,000 tokens from long text, images or PDFs.
- ScaleDown claims up to 1M tokens of multimodal input, 77.2% on DecisionBench versus 72% for Jev, and 200x faster and 467x cheaper than frontier LLM APIs, at $0.05 per million input tokens with output free.
- No public build, download or official product page exists yet, so every figure is a vendor claim from the launch announcement and should be verified before you plan production work around it.
Quick Facts
| Type | Task-specific AI models: classify, extract, summarize |
|---|---|
| Devices | No public build yet; access details unconfirmed |
| Price | $0.05 per million input tokens; output free (announced) |
| Languages | Not stated in the announcement; check official site |
| Highlight | Skips the token-by-token loop; claims 200x faster |
| Learning curve | Low for API users; your prompt is labels or entities |
| Category | AI & Models |
| Creator | Neal Patel |
| Heat | 🔥 1.7M Heat |
| Last updated |
What ScaleDown's Constrained Decoding Models Actually Are
ScaleDown introduced Constrained Decoding Models on October 13, 2026. The name describes the core idea: instead of letting a model write free-form prose one token at a time, a CDM is constrained to a single narrow job and returns a structured answer. ScaleDown frames the family as a direct response to a common engineering complaint, namely that most production calls to a large language model are not really writing tasks at all. They are classification, extraction or summarization jobs dressed up as chat.
The pitch is blunt. Large language models were built to write, but most production workloads just need an answer that is fast, consistent and affordable. A conventional model generates output token by token, and that generative loop is where latency and cost accumulate. A CDM reads your input once and skips the loop entirely, which is the mechanism behind every performance claim in the announcement. Whether that trade-off holds inside your own workload is the question worth testing.
Availability matters here. As of the announcement, ScaleDown has not published a public build, a download, an API key or an official product page. The company has only shown the models in a short preview and announcement, so you cannot sign up and run a CDM today. Treat everything below as a description of what has been claimed, and confirm current status, access and limits on ScaleDown's official site before you plan around it.
- One model, one job
Each CDM is purpose-built for a single task rather than general conversation, so it does not drift into open-ended writing you did not ask for.
- Read once, not token by token
The model consumes your input in a single pass and returns a structured result instead of generating a long response.
- Announced, not shipped
No public build, download or official page exists yet, so this is a preview of an architecture rather than something you can try today.
Inside the Three CDMs: Classify, Extract and Summarize
The family currently lists three models. Classify takes a set of labels you define and returns a calibrated probability plus a reason for each label. ScaleDown suggests it for request routing, tagging and intent detection, the kinds of jobs where you need a decision and a short justification rather than a paragraph of prose. Because the output space is bounded by your own label set, the model never has to invent wording, which is exactly where a generative model burns tokens and time.
Extract works differently. You describe the entities you want in natural language, and the model returns well-formed JSON with confidence scores and citations. The suggested uses are web scraping, document processing and PII redaction, all of which need predictable structure and an audit trail back to the source text. Citations are the interesting part, because they let you verify a field against the passage it came from, usually the hardest step in trusting an automated extraction pipeline.
Summarize accepts long text, images or PDFs and returns a consistent summary under 1,000 tokens. ScaleDown positions it for context compaction, metadata generation and transcript digests, three jobs where the goal is a stable, bounded output rather than a creative rewrite. The 1,000-token ceiling is a deliberate constraint that keeps cost and downstream context predictable. Note that the announcement describes the family as three models live today, yet no public build has been released, so read these as specifications rather than tested behavior.
- Classify
Define your labels and get a calibrated probability plus a reason for each one, aimed at routing, tagging and intent detection.
- Extract
Describe entities in natural language and receive well-formed JSON with confidence scores and citations, aimed at scraping, document processing and PII redaction.
- Summarize
Pass in long text, images or PDFs and get a summary capped at 1,000 tokens, aimed at context compaction, metadata generation and transcript digests.
The Performance Claims: 1M Tokens, DecisionBench and $0.05 Pricing
ScaleDown's launch numbers are specific and worth reading carefully. Input capacity is listed at up to 1M tokens of multimodal text and images, which the announcement describes as 31 times larger than Jev, the model it compares against. On DecisionBench, CDMs are said to score 77.2% against 72% for Jev. Against frontier LLM APIs, the claim is 200 times faster and 467 times cheaper. Every one of these figures comes from ScaleDown's own announcement, so they are vendor claims rather than independently verified results.
The pricing model is the most striking part. ScaleDown lists $0.05 per million input tokens with output free. That structure only makes sense because a CDM is not generating long text: if the output is a label, a JSON object or a bounded summary, charging by output token would be pointless. For teams whose bills are dominated by generated tokens, the arithmetic is easy to sketch. For teams whose bills are dominated by input, the saving is smaller than the headline suggests.
Two caveats belong next to those numbers. First, DecisionBench appears to be ScaleDown's own evaluation, so the comparison against Jev is not neutral. Second, a 200x speed claim depends entirely on what you compare it to and how you measure it, since a constrained model returning one label is doing far less work than a chat model writing a paragraph. Ask for methodology, task definitions and latency percentiles before you treat any of it as a budget line.
- 1M token input
Multimodal text and images, described in the announcement as 31x larger than Jev's input capacity.
- DecisionBench 77.2%
Compared against 72% for Jev, on a benchmark named in ScaleDown's own launch material.
- 200x faster, 467x cheaper
Claims measured against frontier LLM APIs, with methodology not yet published.
- $0.05 per million input tokens
Output is listed as free, a structure that fits bounded, non-generative outputs rather than long text.
Where CDMs Fit, How They Compare and What Remains Unproven
The strongest fit is high-volume, low-creativity work: routing a support ticket, tagging a catalog item, pulling fields out of invoices, redacting personal data, or compressing a long transcript before it goes into another prompt. These are jobs where a general chat model works but wastes output tokens on explanation nobody asked for. If your pipeline already ends with a parser that digs JSON out of a chat response, a constrained model is aimed squarely at you.
Compared with a frontier LLM, a CDM gives up flexibility. It cannot draft an email, reason through a multi-step problem or hold a conversation, and it will not help with tasks whose output shape you cannot define in advance. Compared with classic machine learning classifiers, it asks for far less training data because you describe labels and entities in natural language. That middle position, more structured than a chatbot and more flexible than a trained classifier, is the whole argument.
The open questions are practical. Nothing has been published about latency percentiles, rate limits, supported file types, data retention or whether weights will ever be offered. The announcement says three models are live today, yet no public build exists, so the gap between claim and access is the first thing to resolve. Anyone evaluating this for production should wait for the official page, then run their own labeled sample through all three models before trusting a benchmark.
- Good fit
Routing, tagging, intent detection, scraping, document processing, PII redaction, context compaction and transcript digests.
- Poor fit
Open-ended writing, multi-step reasoning, conversation and any task whose output shape you cannot specify up front.
- Versus a trained classifier
You describe labels and entities in natural language instead of collecting a large labeled dataset first.
- Still unresolved
Latency percentiles, rate limits, file types, data handling and whether weights will be released are all unconfirmed.
Step-by-Step Guide
Read the announcement and note the three task types
Start with ScaleDown's preview material and write down which of the three models, Classify, Extract or Summarize, maps to a job you already run. Because no public build exists yet, this is a scoping exercise rather than a setup guide.
Audit your current model calls
Pull a week of production logs and tag each call as classification, extraction, summarization or genuine generation. Teams usually find that a large share of calls are structured tasks in disguise, which is precisely the workload a CDM is designed to replace.
Prototype the task with a general model today
Build the pipeline now using a conventional LLM with strict output schemas, so you have a baseline and a labeled test set ready. When a CDM becomes available, you can swap the engine and compare accuracy, latency and cost on identical inputs.
Define your labels, entities and summary limits
Write the exact label set, the entity list in plain language, and the maximum summary length your downstream system can accept. These definitions become the model's interface, and vague definitions will produce vague results no matter which engine runs them.
Watch the official site and verify before committing
Check ScaleDown's official site for a public build, API access, confirmed pricing and usage limits. Until that appears, keep the architecture on your roadmap rather than in production, and ask for benchmark methodology once details are published.
Frequently Asked Questions
What are Constrained Decoding Models?
They are a family of AI models from ScaleDown, announced on October 13, 2026, that are built for one narrow task each instead of open-ended writing. A CDM reads your input once and skips the token-by-token generation loop, returning a structured answer such as a label, a JSON object or a bounded summary.
Can I try ScaleDown CDMs right now?
Not yet. No public build, download, API key or official product page has been published, and ScaleDown has only shown the models in a short preview and announcement. Check the official site for release news, and in the meantime prototype the same tasks with a general model so your evaluation is ready.
How much do CDMs cost?
The announced price is $0.05 per million input tokens, with output listed as free. That structure works because the models return short, bounded results rather than long generated text. The figure comes from ScaleDown's own launch material and has not been confirmed on a public pricing page, so verify it before budgeting.
How do CDMs compare with a regular LLM?
ScaleDown claims CDMs are 200 times faster and 467 times cheaper than frontier LLM APIs, and reports 77.2% on DecisionBench against 72% for Jev. Those are vendor numbers without published methodology. The real difference is scope: a CDM cannot write, reason through multi-step problems or hold a conversation, so it complements rather than replaces a general model.
What are the main limitations?
Each model handles one task, so you still need a general LLM for anything open-ended. Access is unconfirmed, benchmark methodology is not published, and details such as latency percentiles, rate limits, supported file types and data handling have not been stated. Treat the architecture as promising but unproven until you can test it yourself.
