Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Developers and teams putting AI into practical use
⚡ TL;DR — the 30-second version
- Account for writes, reads, uncached input and output separately.
- Use current prices and workload-specific reuse.
- A stable cache must not override data freshness requirements.
Separate stable context from per-request state. A reusable prefix might include instructions, tool definitions and reference material; user questions and changing business state usually belong after it. Confirm each provider's actual cache key and retention semantics before relying on this pattern.
Current prices illustrate why the details matter: Opus 5.5 lists $0.20 per million cache-read tokens versus $4 for uncached input. GPT-6 Sol lists $0.20 cached input versus $2 uncached input, and cache writes have their own rate. DeepSeek has both cache-hit pricing and a peak/off-peak schedule. These are billing units, not measured savings percentages.
“Cache hit rate is only half the story. A cache is useful when its writes, reads and expiry pattern reduce your real invoice.”
Estimate the bill from separate token buckets: uncached input, writes, reads and output. Then compare the estimate with your usage record. Include tools, long-context multipliers and any storage or service-tier costs. Cache reuse should also be tested against the application's freshness policy.
OpenAI's September changelog introduces Prompt Cache Diagnostics and documents its cache dashboard. Use actual miss reasons instead of assuming every miss is a prompt-ordering problem. Treat a lower input bill as one component of the full workload cost, not a reason to stop measuring output quality.
Field guide / 01
A worked cache example, with the denominator visible
Suppose a reusable prefix is 10,000 tokens, its ordinary input rate is $2/M, its write multiplier is 1.25 and its read multiplier is 0.10. Across ten requests with one write and nine complete hits, the prefix costs $0.043 rather than $0.20 uncached. This is arithmetic under stated assumptions, not a measured saving for an application.
The changing suffix, outputs, tools and retries still cost money. If the prefix changes or expires, another write changes the result. Use the same accounting for a cold request, a warm request and a request after compaction.
input_cost = (ordinary_tokens × ordinary_rate
+ write_tokens × write_rate
+ read_tokens × read_rate) / 1,000,000
Each input token belongs to one billing bucket, not all three.Field guide / 02
GPT-6 caching is not the older retention interface
The current OpenAI guide documents prompt_cache_options, explicit breakpoints and a 30m TTL for GPT-5.6 and later. It distinguishes those models from earlier prompt_cache_retention behavior. Read the current compatibility table before copying a caching recipe written for an older model.
Measure cache-write and cached-token fields and inspect diagnostics for misses. Preserving a logical conversation does not guarantee an identical rendered prefix; tool ordering, schemas and configuration can change it.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Does a cache return the same answer?
No. Prompt caching reuses prefix computation. The model still generates a new response; it is not a response cache.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01Introducing Claude Opus 5.5Anthropic↗
- 02GPT-6 Sol API model and pricingOpenAI↗
- 03DeepSeek models and pricingDeepSeek↗
- 04OpenAI API changelogOpenAI↗
- 05OpenAI — current prompt-caching behaviorOpenAI↗
- 06OpenAI — model migration and cache changesOpenAI↗