Is Claude or OpenAI prompt caching cheaper in practice?

Neither is universally cheaper. In practice:

• Claude is often cheaper for high-reuse workloads—agents, tool-heavy assistants, and large stable system prompts reused many times within a few minutes.
• OpenAI is often simpler and can be cheaper for sparse or uncertain reuse, because it does not charge a cache-write premium.
• The answer also depends heavily on the specific model: OpenAI’s cached-input discount varies by model—some newer models offer around 90%, while older examples such as GPT‑4o have historically offered about 50%.

The cost trade-off

Claude generally uses:

• A cache-write charge of about 1.25× normal input pricing for a 5-minute cache, or 2× for a 1-hour cache.
• Cache reads at about 0.1× normal input pricing.

OpenAI generally uses:

• No separate write premium—the first request is billed at the normal input rate.
• Cached input discounted by the model’s applicable rate, commonly 50% or 90%.

Ignoring differences in model base prices, for a 5-minute Claude cache versus an OpenAI cache with a 50% read discount:

For Claude’s 1-hour cache, the 2× write premium means OpenAI generally wins unless the cached prefix is reused roughly three or more times during that cache period.

With an OpenAI model offering a 90% cached-input discount, OpenAI may be competitive with—or cheaper than—Claude, especially because it has no write premium. Its disadvantage is that automatic caching is generally best-effort rather than an explicitly controlled cache breakpoint.

Practical decision rule

Choose Claude when:

• The same large prefix is reused repeatedly within 5 minutes.
• You can explicitly control cache breakpoints.
• Cache hits are predictable and frequent.
• Your workload contains large system prompts, tool definitions, or stable RAG context.

Choose OpenAI when:

• Reuse is occasional, spiky, or difficult to predict.
• You want automatic caching with minimal implementation work.
• Your selected model has a 90% cached-input discount.
• Cache entries often expire before being reused enough to amortize Claude’s write premium.

One important qualification: a cache hit normally applies only to an identical eligible prefix. Put stable instructions, tools, and shared context first, and changing conversation or user data afterward. Compare the actual model prices and reported cache-hit tokens for your workload—not just the advertised discount.

Bottom line: Claude is usually the better-value option for reliably high reuse; OpenAI is usually safer for low or uncertain reuse, and can win outright on models with a 90% cached-input rate.