As checked 2026-10-25: Prompt caching is not a promise that identical text always receives a discount. This is a dated economics guide, not a permanent price guarantee or a personal hands-on benchmark.
Scope: global English reference; USD list prices where shown; local eligibility, currency, taxes, negotiated terms and model release dates require separate confirmation. Dynamic source-page prices have a check date, not a verified initial effective date.
Why the first request may cost more
Prompt caching is not a promise that identical text always receives a discount. The initial request may create a reusable cache entry with a dedicated write rate, while a later eligible request can pay a separate hit rate. Anthropic’s dated May 27, 2026 list explicitly distinguishes five-minute cache writes, one-hour cache writes and hits or refreshes. Google’s Gemini pricing can add cached-token storage charges by token-hour. OpenAI model pages publish cached-input rates. The cache mechanism, minimum prefix size and lifetime are vendor- and model-specific; do not silently treat these features as equivalent.
The break-even equation
For N otherwise identical requests each carrying T reusable input tokens, an uncached scenario costs N × T × base_input_rate / 1,000,000. A simple paid-write model costs T × write_rate / 1,000,000 plus (N − 1) × T × hit_rate / 1,000,000, before storage and cache invalidation. Solve for the smallest N at which the second expression is lower. If the platform instead creates cache entries automatically without a separate write SKU, use that platform’s rate schedule, not this paid-write formula.
Cache-hits are an operational metric
High theoretical reuse can fail in practice when a tool adds timestamps, search results or user data ahead of the stable prefix. Place policy and static reference material before frequently changing details when the API’s cache rules favor prefix matching. Instrument actual cached token counts and expiration, not just request counts. For an article-production pipeline, a shared style guide or taxonomy might be reusable; a newly fetched price page or personal record should not be assumed identical or retained indefinitely.
When caching is a poor fit
Caching creates limited value for distinct one-off prompts, low repetition, or material that changes every request. Retention has security and privacy implications: check the provider’s current data handling terms and residency configuration. Counting cached input at both the full uncached rate and the discounted hit rate would double-count it. Storage charges are especially easy to omit when caches persist long after active traffic has fallen.
Responsible comparison
Compare total observed cost and latency under the same workload and quality acceptance criteria, not just the cache-hit rate shown on a price page. Document whether the rate applies to five-minute or one-hour retention, the relevant region and subscription tier, and the exact document date. No universal cache savings percentage is defensible without knowing the reuse distribution.
Design a reusable prompt safely
For a knowledge assistant, put stable policy text and non-sensitive instructions in a versioned template. Append the account-specific question and fresh document passages after the static portion only if the provider’s published matching rules support that architecture. Rebuild the cache when a policy revision is released and log the version hash. This prevents an apparently cheap hit from serving outdated compliance instructions. Caching should never be used to bypass data-handling restrictions or to preserve personal information beyond an authorized retention period; the organization’s privacy rules still take priority.
Two practical break-even checks
First calculate the write premium above an ordinary input call and divide it by the savings on each subsequent hit. Next include the probability that repeated requests arrive before expiry, because a nominal break-even of three hits is irrelevant if users return the next day to a five-minute cache. Storage-priced caches require a third component: the token-hours retained between hits. In a trial, compare cache-hit counters, response correctness and latency against a no-cache baseline, then identify whether differences are due to price rates or prompt restructuring.
Decision checklist
- Identify the specific plan or API endpoint, model version, region and currency before pricing a job.
- Measure input, output, caching, generated media and failed-job units separately.
- Check the official pricing page again before procurement, especially if the SKU has no effective-date notice.
- Keep source evidence and a timestamped price ID; do not treat unknowns as zero.
Pricing evidence boundary
This cost discussion concerns Multi-vendor. Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- Official list-price PDF with effective date 2026-05-27 and scoped Claude API rate rows — official first-party source, accessed 2026-10-25; official price-effective date 2026-05-27.
- Gemini Developer API shows tier/model/context/modality-specific rates and cache storage — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- gpt-5.4 text-token standard list rates and threshold above 272k input tokens — official first-party source, accessed 2026-10-25; original effective date not specified in source.