Is Claude or OpenAI prompt caching cheaper in practice?

Short answer: usually OpenAI is cheaper in practice if you compare similar-capability models and you get decent cache hits; Claude can be cheaper in high-reuse, stable-prefix workflows, especially when its explicit caching fits your request pattern well.

Why:

• The cache discount itself is broadly similar on current flagship-style offerings: cached reads are around 90% cheaper than normal input for both, and cache writes are typically 1.25× input for short-lived caching; Claude also offers a 1-hour cache at 2× input.[1][2]
• So the real cost difference usually comes from the base model price and how often you actually hit the cache, not the caching mechanism alone.[1]
• In many cases, OpenAI’s underlying models are a bit cheaper, so the total bill can be lower even with similar caching terms.[1]
• Claude can win when you have large, stable prefixes reused many times within the cache window, because the 90% read discount amortizes the write cost quickly.[1][2]

Practical rule of thumb:

• Choose OpenAI if you want lower risk, simpler automatic caching, or a cheaper comparable model.[1][2]
• Choose Claude if you have very repetitive prompts, stable system/tool/doc prefixes, and high cache hit rates—especially in agentic or document-heavy workflows.[1][2][3]

One important caveat: if your prompts are dynamic or cache hits are unreliable, the cheaper option is often the one with the lower base model price, because cache savings may not materialize.[1][2]

So, in practice: OpenAI is often cheaper by default; Claude can be cheaper when caching is highly optimized and frequently reused.