DeepSeek Cache Discount: 50x Cheaper Input Tokens, Tested

Cached input tokens on DeepSeek V4-Flash cost $0.0028 per million versus $0.14 for a cache miss, a 50× discount applied automatically at the official api.deepseek.com endpoint. We confirmed it on a live call: a repeated 557-token prompt prefix served 512 of those tokens (92%) from cache on the second request, cutting that call's input cost roughly 10× without a single line of code change. For chat applications with stable system prompts, the cache hit rate is the single largest lever on your DeepSeek bill.

DeepSeek cache hit pricing (verified 2026-05)

According to DeepSeek API Docs, the V4-Flash cache-hit input rate is $0.0028 per million tokens, 50× cheaper than the $0.14 cache-miss rate. A common misreading: the docs note that on 2026-04-26 the cache-hit price was "reduced to 1/10 of the launch price." That 1/10 describes the price cut on that date, not the cache-hit-to-cache-miss ratio. The ratio you actually pay against is 50×, and we verified it on the live pricing page on 2026-05-13.

Model	Input cache miss ($/1M)	Input cache hit ($/1M)	Ratio
DeepSeek V4-Flash	$0.14	$0.0028	50× cheaper

The same 2026-04-26 cache price cut applied across all model tiers, so V4-Pro and the deepseek-reasoner alias carry a comparable structural cache discount; we quote the V4-Flash row here because it is the model the deepseek-chat alias resolves to and the one we verified end to end on a live call.

Sources: DeepSeek API Docs – Pricing and DeepSeek API Docs – Pricing Details USD, verified 2026-05-13.

We tested the cache on the live API (first-hand evidence)

Documentation tells you the discount exists. We called the endpoint twice to confirm it actually fires and to measure how much it saves. The full run is in our evidence pack; here is the cache test.

We sent an identical 557-token system-prompt prefix twice, about one second apart, to deepseek-chat:

Call	prompt_tokens	cache_hit_tokens	cache_miss_tokens	Input cost
#1 (cold)	557	0	557	$0.0000780
#2 (warm)	557	512	45	$0.0000077

On the second call, 512 of 557 prefix tokens (92%) were served from cache and billed at $0.0028 per million instead of $0.14 per million. The input portion of the bill fell from $0.0000780 to $0.0000077, a 10× reduction on that call from cache alone. No configuration, no flag, no separate endpoint. The cache fired because the prefix was byte-identical to the previous request.

One detail the docs do not advertise: when we called the deepseek-chat alias, the API returned deepseek-v4-flash as the actual model. The legacy V3 figures some articles still quote for deepseek-chat are stale; the alias now resolves to V4-Flash at $0.14 / $0.28. We document this so you budget against the model you actually get, not the one the name implies.

Why the cache hit rate matters more than the headline price

Cache hit is the event of an input token being served from DeepSeek's prompt-prefix cache rather than re-processed from scratch, and it triggers the discount automatically. The cache key is the exact leading prefix of the input prompt, so workloads with long stable system prompts and short variable user messages benefit most.

According to DeepSeek API Docs, DeepSeek's cache is a server-side optimization that detects repeated input prefixes and serves them at the $0.0028/M cache rate, distinguishing it from third-party providers that pass through the official rate without their own caching layer. Our live test makes this concrete: the 512-token hit on call #2 is exactly the prefix shared with call #1, while the 45 cache-miss tokens are the small tail that differed.

The mechanism rewards a specific prompt architecture. Put everything stable at the front of the prompt: system instructions, few-shot examples, retrieved context that does not change per user. Put the variable user message last. The longer and more stable your shared prefix, the larger the fraction of every request that bills at the cache rate. According to DeepSeek API Docs, the prefix-cache discount fires automatically on exact byte-identical leading prefixes, with no API flag or endpoint change required.

What the discount does to a real monthly bill

Take a chat workload running a 3,000-token system prompt plus 200 tokens of user input per call, generating 500 tokens of output, at 1 million calls per month with a 70% prefix cache hit rate.

Input tokens per month total 3.2 billion (3,200 × 1,000,000). At a 70% hit rate, 2.24 billion bill at the cache rate ($0.0028/M) and 0.96 billion at the cache-miss rate ($0.14/M): $6.27 + $134.40 = $140.67. Output tokens total 500 million (500 × 1,000,000) at $0.28/M = $140. Total monthly input-plus-output lands at roughly $281.

Without the cache, the same input volume would bill entirely at $0.14/M: 3.2 billion × $0.14/M = $448 on input alone, pushing the total to $588. Our own arithmetic shows why the prefix cache is the dominant cost lever for production chat on DeepSeek: at a 70% hit rate it cut this workload's input bill by roughly 69%, from $448 to $141, while the output bill stayed flat at $140.

Choose this if you have a stable system prompt

Design your prompt schema with the cache in mind. Keep the system prompt and any few-shot examples at the very start, never the end, so the cache prefix is maximized. Avoid embedding per-user variables such as names, timestamps, or session IDs in the prefix portion, because any change to the leading bytes invalidates the cache for the whole prefix. The cache discount is automatic but only triggers on exact prefix match, so consistent prompt structure is the prerequisite, not a runtime setting.

If your workload has no stable prefix, for example a one-shot classification service where every prompt is unique, the cache will rarely fire and you should budget at the full cache-miss rate.

FAQ

Does DeepSeek charge less for cached input tokens? Yes. On V4-Flash, cached input costs $0.0028 per million versus $0.14 for a cache miss, a 50× difference. We confirmed it on a live call where 512 of 557 repeated prefix tokens billed at the cache rate on the second request.

How much does the cache actually save on a real call? On our test, the input cost of an identical 557-token prefix dropped from $0.0000780 (cold) to $0.0000077 (warm), a 10× reduction on the input side of that single call. At scale, a 70% hit rate cut a sample chat workload's monthly input bill by roughly 69%.

When did DeepSeek cut the cache price? DeepSeek reduced the cache-hit input price to 1/10 of its launch-day cache price on 2026-04-26 at 12:15 UTC, which is what brought the current V4-Flash cache rate to $0.0028/M, 50× below the cache-miss rate.

Does the cache discount apply to output tokens? No. The discount only applies to input tokens served from the prompt-prefix cache. Output tokens are billed at the standard rate for the model, since they are generated freshly per request.

This is part of the DeepSeek API pricing hub. For the underlying rate card, see the 1M-token cost article.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-05-13.