DeepSeek API costs: peak, off-peak and cache hits on one workload
Here's the quick version: ten million uncached input tokens and two million output tokens cost $5.40 at DeepSeek Flash peak rates, or $2.70 off-peak. Cached input has its own lower rate, so when you work out your bill, move those tokens out of the uncached bucket instead of counting the same input twice.
Use the current model identifier
Start with a small housekeeping step. The current API table lists deepseek-flash, served by DeepSeek-V4.1-Flash-1017, alongside deepseek-v4-pro. An older alias can still resolve, but that doesn't mean you get its old model or its old price. My advice is to record the model that was actually served, plus its current tariff, in your logs. Future you, staring at an invoice, will be grateful.
One workload in two time bands
| Flash charge per million tokens | Peak | Off-peak |
|---|---|---|
| Cache-hit input | $0.006 | $0.003 |
| Uncached input | $0.30 | $0.15 |
| Output | $1.20 | $0.60 |
Here's how that plays out. Ten million uncached input tokens cost $3 at peak, and two million output tokens add $2.40. Off-peak halves those lines to $1.50 and $1.20. That gives you the $5.40 and $2.70 totals, which don't include separately charged tools and other services.
What to keep in mind: the provider defines peak windows in UTC on weekdays, with Chinese public-holiday exceptions. When you schedule work, use the current schedule on its pricing page and your actual billing records. An offline batch can sometimes wait for the cheaper window, but an interactive customer request often can't. A lower time-band price only helps if the delay fits the job.
Move tokens between cache buckets, not into both
This is the spot where spreadsheets tend to go sideways. Take an alternative ten-million-input workload where eight million tokens are cache hits and two million are misses. With the same two million output tokens, the peak bill is 8 × $0.006 + 2 × $0.30 + 2 × $1.20 = $3.048. Off-peak it is $1.524. If you charged all ten million at the miss rate and then added eight million hits on top, you'd be counting that input twice.
The practical catch: repeated instructions can make context reusable, but similar-looking prompts don't prove that a request received the cached rate. Check the cache-hit and cache-miss token counts in the usage response body, alongside output tokens and the time band. Those measured categories are what you want to reconcile your bill against.
Search and fetch need their own line
DeepSeek’s rate card separately lists Search and Fetch at $0.75 per 1,000 successful requests each. It also describes a temporary free period for specified server-side tool calls through supported agent integrations, with billing starting on 1 November 2026 Beijing time. Before you lean on that promotion for a future month, check which route you use and the effective date.
It's easy to underestimate this one. Two successful tool requests per customer job can make a material difference when the text-model bill is very small. So count the complete agent run, including repeated searches and separately billed retries, rather than treating one user question as one model request.
API or local inference?
Running a checkpoint yourself swaps the token bill for hardware and operating costs. Look closely at the specific downloadable model, its license, and the memory, context and throughput you need. API model names and local checkpoints aren't automatically identical products, and you shouldn't assume equal-quality results.
The most reliable move is to pilot your actual workload and divide total operating cost by successful jobs. On the local side, include idle server time, storage and maintenance. For low or irregular volume, a small usage bill can be much easier to justify than keeping hardware available all month. If you want to see that second calculation worked through, the GPU rental guide walks through it.
Recalculate the DeepSeek Flash workload
Starts with peak Flash rates. For off-peak, set uncached input to $0.15, output to $0.60 and cache-hit input to $0.003. Input buckets are independent: move cache hits out of the uncached count. Add tools at the rate and effective date for your route.
Put the API bill into your application margin →
Frequently asked questions
Is DeepSeek Flash cheaper via API or self-hosting on cloud GPUs?
At current API pricing of $0.15 to $0.30 per 1M input tokens and $0.60 to $1.20 per 1M output tokens (with prompt cache hits dropping to $0.003 to $0.006 per 1M), the official DeepSeek API is drastically cheaper than renting dedicated H100 GPUs until you process hundreds of millions of tokens daily with high sustained utilization.
What hardware is required to run DeepSeek locally?
Running full unquantized DeepSeek models requires multiple enterprise GPUs with 80GB VRAM each. For local workstations or developer servers, running quantized variants (like GGUF or AWQ) still demands at least 48GB to 96GB of unified memory or multiple high-end consumer cards.
Does DeepSeek offer prompt caching discounts?
Yes. DeepSeek's API natively supports context caching, reducing input costs down to $0.003 to $0.006 per 1M tokens on repeated system prompts and document context that stay warm in the cache.