Skip to main content
ScaleDown models are small and built for one job each: Classify, Extract, Summarize and Compress. We benchmark each against leading general-purpose models on public datasets and compare accuracy, cost and latency side by side.

Methodology

The four tasks, their datasets and metrics.

ScaleDown vs OpenAI

Accuracy and latency against GPT-5.5, Mini and Nano.

​
Each comparison reports

  • Accuracy: ScaleDown versus every model in a provider’s lineup.
  • Cost: price per 1K requests for each model.
  • Latency: response time per task.

​
Head-to-head

OpenAI

GPT-5.5, GPT-5.4 Mini, GPT-5.4 Nano8.72% higher accuracy, 2.4× faster.

Anthropic

Fable 5, Opus 4.8, Sonnet 4.6, Haiku 4.58.0% higher accuracy, 3.8× faster.

Gemini

3.1 Pro, 3.5 Flash, 3.1 Flash-Lite9.0% higher accuracy, 8.3× faster.

Grok

Grok 4.3, Grok 4.27.93% higher accuracy, 2.2× faster.

DeepSeek

V4 Pro, V4 Flash19.5% higher accuracy, 4.1× faster.

Qwen

Qwen3.8 Max, 27B, Qwen3.5 4B16.71% higher accuracy, 13.9× faster.
All benchmarks: Full charts, cost tables and latency on scaledown.ai.