ScaleDown has launched three small AI models—Classify, Extract, and Summarize—designed for routine in-app tasks. The models support 1M-token multimodal input and are available via ScaleDown's API and platforms like OpenRouter, Vercel AI Gateway, LiteLLM, and LangChain. They are positioned as faster and cheaper alternatives to frontier LLMs for production workloads.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
ScaleDown has released three small models—Classify, Extract, and Summarize—aimed at routine in-app tasks, touting 1 million token input, text and image support, and pricing of $0.05 per million input tokens ($0.04 during the launch promotion) with output tokens not billed. However, the performance and cost advantages currently come mainly from the vendor's own claims, and the sole source is labeled a paid partnership.
ScaleDown has released three small AI models for routine in-app tasks: Classify, Extract, and Summarize. All three models have a 1 million token input limit and support text or image input.
Functionally, Classify routes text by assigning user-provided labels and gives a probability for each label; Extract can read files, collect the needed data, and output it as JSON fields with high confidence scores; Summarize can read long transcripts or reports and generate a short summary of up to 1,000 tokens.
In terms of distribution channels, the three models are accessible via ScaleDown's API, and are also available through OpenRouter, Vercel AI Gateway, LiteLLM, and LangChain.
Input is currently billed at $0.05 per million tokens, with a launch promotional price of $0.04 on OpenRouter, and output tokens are not billed. This billing structure is the vendor's own claim, and the source does not state whether it can be sustained long term.
On the public intent detection benchmark CLINC OOS, ScaleDown claims Classify achieves 95% accuracy, while the average for three OpenAI models is 89.6%.
This comparison is a vendor claim, and the source does not disclose the specific evaluation setup, samples, or the versions of the comparison models for CLINC OOS, so it remains unclear whether the 95% vs. 89.6% was measured under equivalent conditions.
ScaleDown co-founder and CEO Neal Patel proposes a new AI model architecture called "Constrained Decoding Models" (CDMs), arguing that most production workloads only need fast, consistent, and affordable answers.
The source does not state whether the CDMs architecture has actually been used in these three models, or whether it is only presented as a concept.
The source mentions metrics such as 77.2% on DecisionBench, 200x faster, and 467x cheaper, but provides no verifiable evaluation details. The baselines and measurement methods for these numbers are unclear, so they can only be treated as mentioned by the source, not as confirmed performance or cost conclusions.
A comment from Fireply.ai argues that Classify's "one probability per label" is useful, as a threshold can be set to escalate only uncertain samples to a larger model. This is a third-party comment and does not represent independent verification of the model's overall performance.
All the above information comes from a single post by TestingCatalog on X and its paraphrases, lacking independent corroboration from ScaleDown's official first-party sources. The post is labeled a paid partnership, indicating a commercial promotional nature.
Therefore, the performance and cost comparison data are vendor claims and have not been independently verified by third parties; whether the paid partnership affected the post's content and data presentation cannot be determined. Key specifications such as model parameter scale, latency, and context output limits are also not disclosed, and the performance of these models under real production workloads lacks independent evaluation.
All information in this article comes from a single post by TestingCatalog on X and its paraphrases; the post is labeled a paid partnership and lacks independent corroboration from ScaleDown's official first-party sources. The performance and cost comparison data are vendor claims and have not been independently verified by third parties; the specific evaluation setup, samples, and comparison model versions for CLINC OOS and DecisionBench are not disclosed.
ScaleDown's three small models make clear claims regarding input length, distribution channels, and pricing structure, but the performance and cost advantages currently remain at the level of vendor claims. In the absence of independent evaluation and official first-party materials, these numbers are better treated as leads to be verified rather than conclusions that can be directly accepted.
Primary report
Primary source