Prompt Caching Pricing Comparison: 8 APIs Measured

On the same repeated ~1800-token prefix, only three of the eight endpoints we priced on 2026-07-28 got cheaper on the second call: DeepSeek native fell 88% on computed input cost, GLM-5 fell 75% and MiniMax-M2.5 fell 62% on billed cost.

Prompt caching pricing looks tidy on a rate card and behaves messily in production. Whether a published discount reaches your invoice depends on two things the rate card never tells you: which path you call through, and whether the cache engages by default or only when you mark it.

Prompt caching pricing comparison (verified 2026-07)

Prompt caching is an input-side billing mechanism that charges a repeated prompt prefix at a reduced rate instead of the full input price, because the provider reuses computed state rather than recomputing it. It moves cost and latency. It does not touch output quality, and no vendor here claims that it does.

Disclosure on every number above: DeepSeek native figures come from api.deepseek.com, and every other row was measured through OpenRouter. Routed rows are billed cost from the router's own usage field. The DeepSeek native row is computed from measured hit and miss tokens at official published rates, because the native API does not return a cost field. This cluster carries needs_native_reverify=true.

What the run actually measured, first-hand

We sent the same fixed ~1800-token prefix, a fictional API changelog, two or three times to each endpoint on 2026-07-28, two seconds apart, changing only a one-line question, at temperature 0 with max_tokens 40. Cache behavior was read from each API's own usage fields rather than inferred. All 17 calls succeeded and total routed spend was $0.051.

DeepSeek native was the clearest result. Call A reported 3403 prompt tokens with 3403 counted as cache misses and took 1.43 seconds. Call B reported 3328 cache-hit tokens against 75 misses and returned in 0.90 seconds. Call C hit 3328 tokens again. Priced at official rates, that is $0.000476 of input on the first call and $0.0000571 on the repeat.

The sharpest comparison is DeepSeek against itself. The same model family through the router reported zero cached tokens on both calls and billed $0.000698 then $0.000696, full price twice. On this pair, the native repeat call cost roughly one twelfth of the routed repeat call.

This design measures implicit caching only. Providers whose caching requires explicit markers are expected to report zero cached tokens here, and that is a finding about default behavior, not about capability. The zero rows for Anthropic, OpenAI, Kimi and Qwen say what happens when you send an ordinary request, nothing more.

The published rate structures behind the discounts

According to DeepSeek's API documentation, cache-hit input tokens are billed at $0.014 per million against $0.14 per million on a miss, applied automatically with no request-side flag. That ten-to-one structure is what turns our measured 98% hit rate into an 88% input-cost cut rather than a rounding error.

OpenAI and Anthropic take opposite design routes to a similar goal. According to OpenAI's prompt caching guide, caching is enabled automatically for prompts of 1024 tokens or longer and matching prefixes bill at the cached-input rate.

Anthropic charges for the privilege of control. According to Anthropic's prompt caching documentation, five-minute cache writes cost 1.25 times base input, one-hour writes cost 2 times base input, and cache reads cost 0.1 times base input. Nothing caches until a cache_control breakpoint says so.

That write premium is why Claude's row shows no change in our table. An implicit-only test cannot trigger an opt-in cache, so Anthropic billed $0.01207 and $0.01201 for two full-price passes over the same prefix. A workload that sets breakpoints correctly would land on the 0.1x read multiplier instead. The pricing lever exists; our request simply never pulled it.

Alibaba sits in the middle. According to Alibaba Cloud's Model Studio documentation, Qwen offers context caching on its native endpoint, which we could not demonstrate through the routed path.

The Qwen row is a pricing artifact, not a cache

Qwen3-235B billed $0.000726 on call A and $0.000331 on call B with zero cached tokens reported on both. That 54% drop is not cache evidence. Call B matches Qwen's list price for the tokens involved, roughly $0.00032 computed, while call A billed about twice list, so the delta tracks provider-price variance on the routed marketplace, where different upstream hosts serve the same model at different rates.

We flag this because the lazy version of this article would print the number as a fourth caching win. Any prompt caching pricing comparison built purely from billed deltas on a multi-provider router will manufacture savings that never existed.

Routing is the variable, not the vendor

"Routers lose caching" is false as a blanket claim, and our own data is what disproves it. GLM-5 returned 3200 cached tokens through the router and its billed cost fell 75%. MiniMax-M2.5 returned 3296 cached tokens and fell 62%. Both discounts appeared in routed billing without any native access. Meanwhile DeepSeek, the vendor with the strongest native cache in the set, showed nothing through the same router on the same day.

Cache preservation is a per-integration property, not a property of routers in general. Two providers on the same aggregator behaved in opposite ways inside one 17-call run. If cache economics matter to your budget, check the usage fields on your own repeat calls rather than trusting a feature matrix.

Choose native access if your workload replays a large fixed prefix, such as a long system prompt or a retrieval context block, thousands of times per day. Choose a router if your prefixes are short or your traffic is spiky. For DeepSeek's cache-hit pricing in isolation, see our DeepSeek cache discount breakdown.

FAQ

Which API had the cheapest cached repeat call in this test? DeepSeek native at $0.0000571 computed at official rates; the cheapest routed billed figure was MiniMax-M2.5 at $0.000204. The two are not strictly comparable because one is a billed routed figure and one is computed from native hit and miss tokens.

Does prompt caching make the model's answers better? No. Caching affects input cost and latency only. Our measurement records token accounting, billed cost and response time, and we make no quality or accuracy claim anywhere in this cluster.

Why did Claude and GPT-4o show no saving? Because the test sent ordinary requests with no cache markers, through a router. Anthropic's caching is opt-in via cache_control, and OpenAI documents automatic caching on its own API, which we did not test natively. Both zeros describe default routed behavior.

Is the 88% DeepSeek saving a billed number? No. It is computed from the measured 3328-hit and 75-miss split at DeepSeek's official rates of $0.014 and $0.14 per million input tokens. The native API returns token fields, not a cost field, so the arithmetic is ours and the tokens are theirs.

How long do these figures stay valid? Treat them as a 2026-07-28 snapshot from a single region. Rate cards and router integrations both change, and this cluster is flagged for native re-verification before 2026-09-26.

This page is one micro in the prompt caching cluster, which collects the full 17-call run, the latency findings and the native-versus-routed analysis in one place.

Author: Kevin Fan, Customer Success Manager. Last verified: 2026-07-28.