Zumik
Sign inSign up

Rankings that account for reuse

A model that looks cheap on paper can be expensive in practice, and vice versa. These boards rank by what agent workloads actually pay once caching is working - plus the latency, quality, and capture lenses that matter.

Cheapest models for cached agent workloads

$/1M tokens

Which model is cheapest once a typical agent prefix is served from cache?

  1. OpenAI gpt-oss-120b · $0.21
  2. GPT-5 Mini · $0.59
  3. Grok 4.3 · $1.13
Full ranking

Fastest time-to-first-token (warm prefix)

ms

Which model responds fastest when the prefix is already cached?

  1. Claude Haiku 4.5 · 110 ms
  2. OpenAI gpt-oss-120b · 120 ms
  3. DeepSeek-V4-Pro · 130 ms
Full ranking

Best models for coding agents

index

Which models balance code quality with reuse economics?

  1. Claude Fable 5 · 84.0
  2. Claude Opus 4.8 · 83.0
  3. GPT-5.5 · 79.2
Full ranking

Best intelligence per dollar

index per $

Which model gives the most quality per reuse-adjusted dollar?

  1. OpenAI gpt-oss-120b · 372.9
  2. GPT-5 Mini · 136.0
  3. Grok 4.3 · 78.7
Full ranking

Highest measured cache capture

%

Which models convert reuse opportunity into billed savings most reliably?

  1. Claude Fable 5 · 93%
  2. Claude Opus 4.8 · 93%
  3. Claude Opus 4.7 · 92%
Full ranking

Rankings are a starting point. Measure your own.

Your traffic determines your real reuse. Run a diagnostic to rank models for your workload specifically.

Run a diagnosticModel catalog