Is Claude or OpenAI prompt caching cheaper in practice?

Short answer

Claude is usually cheaper for dense, repeated reuse of a large, stable prompt prefix. OpenAI is usually cheaper for sporadic traffic or uncertain cache hits. The result also depends heavily on the specific models, because OpenAI’s cached-input discount varies by model.

Why the economics differ

• Claude uses explicit caching: a cache write costs more than normal input—typically 1.25× the base input price for a 5-minute cache, or 2× for a 1-hour cache—while cache reads are about 0.1× the base price.
• OpenAI caching is automatic and has no write surcharge: a cache miss simply incurs the normal input price. Cached-input pricing has historically ranged from roughly 0.5× to 0.1×, depending on the model.

Assuming the same base input price \(P\), a five-minute Claude cache with \(N\) subsequent hits costs approximately:

• Claude: \(1.25P + 0.10PN\)
• OpenAI at a 50% cached-input discount: \(1.00P + 0.50PN\)

Under those assumptions:

For example, with ten cache hits, the input-token cost is about 2.25× \(P\) on Claude versus 6× \(P\) on OpenAI at a 50% cached-input rate.

When OpenAI is cheaper

OpenAI is generally preferable when:

• requests are isolated or arrive outside the cache-retention window;
• cache reuse is unpredictable;
• the prompt prefix changes frequently;
• you want caching without explicit cache-breakpoint configuration;
• the OpenAI model you choose already offers a very large cached-input discount—around 0.1×, in which case Claude’s write surcharge may erase its read-cost advantage.

The important benefit is that an OpenAI cache miss has no penalty beyond paying the ordinary input price.

When Claude is cheaper

Claude tends to win when:

• the same large system prompt, document set, codebase, or conversation history is reused repeatedly within minutes;
• an agent performs many sequential tool calls or reasoning iterations;
• the cached prefix is large enough that the 90% read discount matters;
• cache hits are reliable enough to amortize the initial write fee.

Claude’s one-hour option can be useful for lower-frequency reuse, but its higher write fee means it needs more subsequent hits to beat OpenAI.

Practical conclusion

For a frequently used 10,000–100,000-token context—such as a coding-agent workspace, RAG corpus, or long-running conversation—Claude will often have the lower input-token bill, particularly when compared with an OpenAI model offering only a 50% cache discount.

For occasional requests, uncertain reuse, or an OpenAI model with approximately 90% cached-input savings, OpenAI may be cheaper and carries less cache-miss risk.

Finally, compare the complete bill, not just cached-input prices: model output prices, uncached dynamic tokens, cache hit rates, and the amount of the prompt that actually matches can outweigh the caching difference. Pricing and retention rules are model-specific and change over time, so the live pricing pages should be checked before committing to an implementation.