What are the downsides of ScaleDown Constrained Decoding Models (Classify, Extract, Summarize)?

Launch claims of 467x cheaper and 77.2% on DecisionBench lack a published baseline; billed input includes label criteria and schemas, so it exceeds your prompt; each Decisions question is billed as a separate call; /classify caps labels at 20; summarize drifts off the question and skipped half of a 369K-token input.

ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) review (updated)

What is ScaleDown Constrained Decoding Models (Classify, Extract, Summarize)?

Constrained Decoding Models (CDMs) are ScaleDown's family of single-task hosted models, launched on 13 October 2026 by co-founder and chief executive Neal Patel. Classify returns a probability per label you define, Extract pulls fields you describe into JSON, and Summarize writes a summary of at most 1,000 tokens. An older model, Compress, trims context before an LLM call. All four take up to 1M tokens of input, and the three CDMs also read images and PDFs.

Classify has two interfaces. POST /classify takes labels with rubrics and supports batches of up to 50 inputs. POST /v1/scaledown is what ScaleDown's docs call "a drop-in replacement for the Jev Decisions API": it copies the request and response shape of TypeSafe's Jev, with choice, noul (yes/no) and score questions. ScaleDown is a separate company and does not serve Jev; it built its own model behind the same request format. Extract and Summarize are called through an OpenAI-compatible /v1/chat/completions endpoint at https://api.scaledown.xyz.

Text input costs $0.05 per million tokens and images $0.15, with output free. On OpenRouter, Classify-1, Extract and Summarize list at $0.04 under a 20% discount.

Is ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) worth it?

ScaleDown's price is the claim that survives checking; most of its launch arithmetic does not. The launch post promised models "200x faster, 467x cheaper than frontier LLM APIs" and 77.2% on DecisionBench, against 72% for Jev. ScaleDown's own benchmark docs are far more modest: 2.2 to 13.9 times faster than each provider's lineup average, and 89 times cheaper than the OpenAI average and 161 times cheaper than the Anthropic one. No public record of the 77.2% DecisionBench result has surfaced, a gap a Bitroot thread flagged, according to an independent write-up. All of ScaleDown's comparisons are self-reported.

The best independent evidence is a 1,000-call test by Ishank Choudhary on DEV Community, which is one engineer's run rather than a peer-reviewed evaluation. On 300 CLINC150 intents the Jev-compatible Decisions endpoint scored 97.3% (95% interval 94.8 to 98.6) against 90.7% for a local Qwen3 1.7B; on 110 Bitext support tickets it was 92.7% against 62.7%. Extract tied the local model at 89.5% field accuracy on 50 SROIE receipts but read totals correctly every time, against 78%, and every value comes back with a confidence score and character offsets into the source.

The bill is larger than your prompt. ScaleDown counts label criteria, schemas and its internal prompts as input: Choudhary's classify calls billed about 481 tokens where the same task as a plain prompt was 172. On the Decisions endpoint every question is a separate model call billed separately, where Jev answers all questions in one pass. At list prices that made short classification twice the cost of gpt-5-nano ($0.024 against $0.012 per 1,000 calls) and slightly dearer on extraction ($0.069 against $0.057). The savings are large only against flagship models: 45 to 102 times against gpt-5.5 in the same test. The /classify batch mode helps, since 50 inputs share one billed label set.

Summarize is the weakest of the three. It invented no facts in six QMSum meetings, but it often summarised the wrong part of a meeting when asked a question, and on a 369,000-token document it covered only one of two meeting series. The 1M-token window accepts the input; it does not guarantee the model reads all of it.

Against Jev itself, the compatibility is real but partial. The docs list two model names (classify and classify-1), no box questions, at most 10 levels per score question, and a 20-label cap on /classify, which choice questions run on, against Jev's 255 options. A Jev routing layer with long option lists or many questions per request will not port at the same cost.

Where it earns its place

A platform engineer with a Jev routing layer in production
Run the same decision questions on a second vendor for price and failover
Point the existing Jev request shape at /v1/scaledown and compare answers on logged traffic. Requests with several questions bill one call per question, and choice questions above 20 options need splitting, so measure the bill before switching.
An accounts-payable lead at a mid-size distributor scanning 5,000 receipts a week
Pull merchant, date and total from photos into the ledger with an audit trail
Extract reads the image, returns each field with a confidence and offsets back into the text, and low-confidence rows go to a clerk. In the independent test it got receipt totals right 50 times out of 50, at about $0.069 per 1,000 calls.
A support-operations analyst at a bank tagging historic tickets
Classify a backlog of ten million tickets against 15 intents overnight
Send 50 tickets per /classify request so the 15 rubrics are billed once per batch rather than once per ticket. Batch requests skip the one-sentence reasoning call, which on single requests re-sends the rubrics at twice the input price.

Before you commit

Setup: The site offers sign-up with 1M free tokens and no card; the docs' introduction still points new users to a sales call for a key. Requests use an x-api-key header against https://api.scaledown.xyz; api.scaledown.ai is not an API host.
Integrations: Custom /classify and /compress/raw/ endpoints, the Jev-shaped POST /v1/scaledown, and OpenAI-compatible chat completions for extract and summarize. Also on OpenRouter (Classify-1 through its Decisions API), LiteLLM and LangChain's langchain-scaledown middleware.
Data and privacy: ScaleDown's homepage advertises zero data retention, no training on customer data and self-hosting, and says it is built to SOC 2, HIPAA, GDPR and ISO 27001 standards, with a SOC 2 Type 1 badge. The docs do not say which plans get zero retention or publish retention periods.
Cost at scale: At the independent test's measured token counts, a million classify calls cost about $24, a million extract calls about $69 and a million 6K-token summaries about $317. The launch-window offer of 50M free tokens has ended; new accounts now get 1M.
Limits: 1M input tokens; summaries capped at 1,000 output tokens; /classify takes at most 20 labels (16 multi-label) and 50 chunks; score questions take 2 to 10 levels. Out of credit returns 402 Insufficient credits. Rate limits are not published.

What's good

97.3% on CLINC150 and 92.7% on Bitext in an independent 1,000-call test
$0.05 per 1M input tokens, output free, one flat rate across four models
Extract returns confidence and character offsets for every value
Reads images and PDFs on Classify, Extract and Summarize

What's not

Launch claims of 467x cheaper and 77.2% on DecisionBench lack a published baseline
Billed input includes label criteria and schemas, so it exceeds your prompt
Each Decisions question is billed as a separate call; /classify caps labels at 20
Summarize drifts off the question and skipped half of a 369K-token input

Best for

Teams paying flagship-model prices for routing, tagging or field extraction who want a cheap hosted replacement and will test it on their own labelled data first.

ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) alternatives

Jev

TypeSafe AIDecision models

3.8/5

TypeSafe AI's hosted decision model: typed choices, scores and yes/no probabilities for $0.042 per million input tokens, output free.

Perplexity Decider (pplx-decider)

PerplexityDecision modelsOpen weights

4.5/5

Perplexity's open 27B and 4B decision models, also sold as a $0.02-per-million-token API: text and images in, probabilities out.

Laya

Convai InnovationsDecision modelsOpen weights

3.4/5

Apache 2.0 decision model on ModernBERT that answers typed questions in tens of milliseconds, but needs fine-tuning to be accurate.

GLiNER2.5-Decide

FastinoDecision modelsOpen weights

3.6/5

Fastino's sub-billion-parameter zero-shot classifier: pass any labels at call time and get one back on a CPU, no prompt or generation.

d3 (vLLM Semantic Router)

vLLM Semantic RouterDecision modelsOpen weights

4.0/5

vLLM Semantic Router's open 27B decision model with video input and stock vLLM serving, already overtaken by its own d3-turbo.

ScaleDown Constrained Decoding Models (Classify, Extract, Summarize): common questions

Is ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) worth it?: The prices are real and the classifier is good, but the launch numbers outran the evidence and the Jev compatibility costs more than it looks. We rate it 3.7 out of 5.
How much does ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) cost?: $0.05 per 1M text input tokens, $0.15 per 1M image input tokens, output free; 1M free tokens for new accounts. OpenRouter: $0.04 with a 20% discount. Enterprise: dedicated instance or self-host licence, by quote.
What are the downsides of ScaleDown Constrained Decoding Models (Classify, Extract, Summarize)?: Launch claims of 467x cheaper and 77.2% on DecisionBench lack a published baseline; billed input includes label criteria and schemas, so it exceeds your prompt; each Decisions question is billed as a separate call; /classify caps labels at 20; summarize drifts off the question and skipped half of a 369K-token input.
Who is ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) best for?: Teams paying flagship-model prices for routing, tagging or field extraction who want a cheap hosted replacement and will test it on their own labelled data first.
How much does ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) cost for a team or in production?: At the independent test's measured token counts, a million classify calls cost about $24, a million extract calls about $69 and a million 6K-token summaries about $317. The launch-window offer of 50M free tokens has ended; new accounts now get 1M.
Is ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) safe for company data?: ScaleDown's homepage advertises zero data retention, no training on customer data and self-hosting, and says it is built to SOC 2, HIPAA, GDPR and ISO 27001 standards, with a SOC 2 Type 1 badge. The docs do not say which plans get zero retention or publish retention periods.
What are the best ScaleDown Constrained Decoding Models (Classify, Extract, Summarize) alternatives?: The closest alternatives we have reviewed are Jev, Perplexity Decider (pplx-decider), Laya, GLiNER2.5-Decide and d3 (vLLM Semantic Router).

Answers checked Oct 26, 2026.