ScaleDown: Classify-1

ScaleDown

•

Classify-1

•

Released

October 2026

ScaleDown
Classify-1

Calibrated per-label probabilities for routing, tagging and intent detection, with 1M-token text, image and PDF input.

Modality:

Text

model ID

scaledown/classify-1

Start for free

View API Docs

Output Speed *

N/A

tok/s

Intelligence Index *

N/A

/ 100

Context Window *

1000000

tokens

Input price

0.24

Anytoken

Output price

Anytoken

Calibrated Label Scores for Routing, Not Generated Text classify-1 is ScaleDown's task-specific classification model. It doesn't generate prose. It scores text, images or documents against labels you define in plain language and returns a probability for every label. It sits in ScaleDown's single-task model family beside Extract, Summarize and Compress. Its 1M-token input limit, rubric-based labels and optional one-sentence reason suit high-volume routing, tagging, moderation and intent detection that run ahead of a larger model. Add classify-1 to your routing layer via AnyAPI.ai.

Performance

Label Probabilities Over 1M Tokens, Without Generating Output

classify-1 reads the input once and returns a score per label instead of writing an answer token by token. In an independent hands-on test on 300 CLINC150 utterances, the Decisions endpoint reached 97.3% accuracy and rejected 49 of 50 out-of-scope queries, with a 162 ms median latency. That makes it a practical first-stage router, where fast decisions gate expensive generative calls. The sample is small and the tester's own, so validate on your taxonomy.

Benchmarks

classify-1 Benchmarks: Small Independent Tests, No Public Leaderboard Yet

No standardized leaderboard lists classify-1 yet, so the evidence comes from hands-on tests. One independent developer ran the Decisions endpoint on 300 CLINC150 utterances (97.3%, versus 90.7% for a local Qwen3 1.7B) and on 110 Bitext support tickets (92.7% versus 62.7%). Median latency was 162 ms and p95 was 397 ms. ScaleDown's own CLINC150 comparison put it level with small LLMs, within noise. Treat these as directional. The 200x speed and 467x cost multiples are vendor claims.

Output Speed

N/A

tok/s

Intelligence Index

N/A

/ 100

MMLU *

Broad world knowledge and problem-solving

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.

HLE *

Adherence to multi-step structured instructions.

LiveCodeBench *

Tool-calling reliability in long agentic loops.

Technical Specifications

What the model supports

classify-1 accepts text, images and files and returns label scores instead of generated text. The 1M-token input limit is stated by ScaleDown and has not been independently corroborated. The tighter constraint is label count: 20 labels per request, 16 in multi-label mode, and 50 chunks per batch. Plan taxonomies around those caps. Because the response is structured scores, maximum output length is not a meaningful sizing factor for this model.

Verified Specifications —

Classify-1

Input modalities

Text

output modalities

Text

Context window

1000000

tokens

Maximum output tokens

900000

Reasoning

Knowledge cutoff

October 2026

Pricing (standard)

0.24

AnyTokens in

AnyTokens out

Limitations & Trade-offs

Where Classify-1 falls short

20-label ceiling. A request supports up to 20 labels in either mode, multi-label mode is capped at 16, and batches are capped at 50 chunks. Larger taxonomies have to be split across calls, which means reconciling scores from separate requests. At least one competing decision model documents up to 255 options. Flat intent sets under 20 labels fit well. Large product or topic taxonomies need a two-stage hierarchy or a different model.

Always returns a winner. classify-1 returns a top label even when none fits, and single-label scores are relative: a 0.4 top score across 10 labels can still be correct. You must add a catch-all label and write your own confidence threshold. In ScaleDown's CLINC150 test, right answers averaged a 0.88 score and wrong ones 0.60, and the reasoning sentence never flagged similar-label mix-ups. Gate automation on scores, not the explanation.

Billing scales with label count. In single-label mode every label and rubric is billed as input. Multi-label mode runs one model call per label. The optional explanation re-sends the labels at 2x input billing. ScaleDown's own test noted about 610 template tokens billed per call, which weighs most on short inputs. There is no output-token charge. Turn reasoning off in automated pipelines and use batch chunks to share the label profile.

Thin independent evidence. No public leaderboard lists classify-1, and BenchLM shows no benchmark rows for it. The 200x faster and 467x cheaper claims are vendor figures that critics say lack a stated baseline. Latency evidence also varies: one tester measured a 162 ms median, while ScaleDown's own laptop tests showed about 1.05 s over short calls. Run a labelled evaluation on your data before replacing an LLM classifier.

Best-Fit Workloads

Where this model earns its place

Intent routing and ticket triage

‍
Per-label probabilities let a router send confident cases straight to a queue or model and flag the rest. An independent test scored 97.3% on CLINC150 and 92.7% on support tickets, with each label written as one plain-English line. This fits support routing, chatbot intent detection and model-selection gates. Keep label sets under 20 and include an out-of-scope label.

Long-document and scanned-file tagging

‍
The document path takes one base64 JPEG, PNG, TIFF or PDF per request and returns the extracted text with the scores. Inputs can reach 1M tokens (vendor-stated). That suits contract-type detection, invoice-versus-report sorting and policy routing without a chunking layer. Cost and latency grow with document size and label count, and only one document is accepted per request.

Agent guardrails and loop detection

‍
OpenRouter's classify-1 documentation links a cookbook for detecting stuck and looping agents. LangChain's integration post describes running Classify on a model's output and for runtime model routing. The one-sentence reason helps explain why a run was stopped. This fits long-running agents that need a cheap decision at every step. The reason is generated after the label, so trigger on the score.

Confidence-gated automation with human review

‍
With a catch-all label and a 0.8 cutoff, ScaleDown's own CLINC150 test auto-handled 76% of messages at 99.6% accuracy and sent the rest to review. Multi-label mode adds independent tags per input, up to 16 labels. This pattern suits moderation, compliance tagging and queue prioritization. The result is vendor-reported on a simplified per-domain label set, so recalibrate the cutoff on your own data.

Pricing in anytokens via AnyAPI

Input

0.24

₳

Output

₳

Cache write

—

₳

Cache read

—

₳

Get API access

Read Docs

Integration

Access Classify-1 via AnyAPI.ai

Access Classify-1 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Classify-1 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

One API integration

Access Classify-1 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

Easy model switching

Test Classify-1 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

Flexible for production

Use Classify-1 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

Multi-model applications

Use Classify-1 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free

Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

What is classify-1?

classify-1 is ScaleDown's task-specific classification model, released October 13, 2026. It scores text, images or documents against labels you define and returns a probability for each label rather than generated prose. ScaleDown positions it for routing, tagging, moderation and intent detection. It is a closed, hosted model with no reasoning mode beyond an optional one-sentence explanation.

What is the classify-1 context window and what inputs does it accept?

classify-1 supports a 1,000,000-token input limit, as stated by ScaleDown. It accepts text, images (JPEG, PNG, TIFF) and PDFs, and returns structured decisions. This is an input limit, not a generation limit, because the model doesn't write long outputs. Third-party catalogues have not independently corroborated the 1M figure.

How is classify-1 priced and what changes the bill?

classify-1 is billed on input tokens with no output-token charge, so it is a low-cost option compared with generative LLMs. The bill rises with label count, because rubrics count as input. Multi-label mode makes one call per label, and the optional explanation re-sends labels at 2x input billing. See the AnyTokens pricing on this page for current rates.

How do I handle texts that match none of my classify-1 labels?

Add a catch-all label such as other and set your own confidence threshold, because classify-1 always returns a top label even when nothing fits. Single-label scores are relative and sum to 1.0. In multi-label mode, labels at or above your threshold are returned as matched. Use the scores to gate automation, not the explanation sentence.

What are the classify-1 benchmarks?

No public leaderboard lists classify-1 yet. One independent test measured 97.3% accuracy on 300 CLINC150 utterances and 92.7% on 110 support tickets, with a 162 ms median latency. ScaleDown's own test put it level with small LLMs, within noise. Its 200x speed and 467x cost claims are vendor figures, so test on your own data.

* Benchmark data source: Artificial Analysis artificialanalysis.ai