ScaleDown targets AI inference costs with task-specific small models

Why it matters

ScaleDown is selling a narrower bet than frontier-model labs: many production AI tasks are repetitive enough that a cheaper specialized model can win on unit economics.

Neal Patel (@neal_k_patel) introduced ScaleDown in a 22-post thread on X, pitching task-specific small language models for AI workloads he says do not require a frontier model.

https://x.com/neal_k_patel/status/2062534030638141695

Patel's headline claim is aggressive: ScaleDown is "15x cheaper," "63x faster" and "5.1% more accurate than GPT-5.4 Mini." Those numbers come from Patel's launch post, not an independently published benchmark in the materials provided. He also says 70% to 80% of AI workloads do not need a frontier model, a framing that puts the startup in the cost-cutting lane as companies try to reduce inference bills without ripping out AI features.

ScaleDown's own materials position the company as an applied AI research lab building purpose-built small language models for four repeatable jobs: compression, summarization, extraction and classification. The pitch is not another general-purpose assistant. It is that narrow models, trained to do one task, can deliver frontier-quality outputs with lower cost and latency than general-purpose models such as GPT-4 or Claude.

The first product, COMPRESS, is described as lossless, query-aware compression for long prompts and documents. ScaleDown says it strips noise while preserving facts relevant to a query, with a typical 50% to 70% compression ratio. The company frames the model for RAG pipelines, document analysis, long conversation management, code review, batch workflows and reducing overhead in reasoning-model traces.

SUMMARIZE is aimed at abstractive summaries of long documents, transcripts and threads without truncation. ScaleDown lists legal contracts, research papers, customer support tickets, surveys, meeting transcripts, earnings calls, financial filings and content pipelines as target use cases. EXTRACT is pitched as semantic named-entity recognition driven by natural-language instructions, returning structured JSON for contracts, resumes, medical records, financial documents and e-commerce data. CLASSIFY is built for high-throughput tagging, triage and filtering, including support routing, moderation, intent recognition, domain tagging and lead scoring.

The current product focus is narrower than a general chatbot and more directly tied to developer pipelines. The company offers the four models through a unified REST API, with documentation at docs.scaledown.ai. Its API base URL is https://api.scaledown.xyz, with endpoints for POST /compress, POST /summarize, POST /extract and POST /classify, authenticated through an x-api-key header. ScaleDown is offering API keys through its sign-up page.

In replies, Patel said ScaleDown charges $0.05 per 1 million input tokens and "never" charges for output tokens. ScaleDown's public materials describe the same public API price as a flat $0.05 per 1 million tokens, with 50 million free tokens to start and no credit card required. Enterprise and self-hosted plans are described as custom-priced, with deployment in a customer's VPC, fine-tuning on customer data, dedicated support and SLAs.

Patel listed peak throughput at 20,000 tokens per second for compression, 5,000 for summarization, and 12,000 for classification and extraction. ScaleDown does not have a streaming API, Patel wrote. The company's materials separately claim its models are 10x cheaper and 2x faster than models like GPT-4 or Claude, and cite self-reported benchmark claims including 70% compression on FinanceBench versus GPT-5 direct, summarization that is 14x cheaper than GPT-5.4, and classification that is 113x cheaper than GPT-5.4 Mini at the same accuracy.

Patel also gave limits that matter for buyers: unlimited context for compression, 1 million tokens for summarization, classification and extraction, and best results in English and European languages, with other scripts supported but variable. ScaleDown supports OCR for document extraction and summarization, according to Patel, though it is not a vision-model company. Patel linked a public SLM agent repo meant to help users choose smaller models for tasks.

For enterprise buyers, ScaleDown is also making a security argument. Its materials say prompts, outputs and API keys are never logged or stored, and that customer inputs are not used to fine-tune or improve its models. The company says GDPR, HIPAA and SOC 2 compliance are in progress, and that enterprise plans can include SSO, RBAC and audit logs. It lists AWS, Google Cloud, NVIDIA, Intel and Nutanix as infrastructure partners.