LLM / providers / deepseek

DeepSeek API Pricing

DeepSeek offers open-weight models at some of the lowest per-token rates in this comparison, with a large context window and a very steep cached-input discount, which suits high-volume or cost-sensitive workloads.

Prices verified June 2026 · changes logged in the changelog
Heads up: DeepSeek V4 Pro's headline rate reflects a standing promotional discount (roughly 75% off). If that promotion ends, the list price would step up several-fold — treat the current rate as promotional rather than permanent when you forecast a long-term budget.
Model$ input /1M$ output /1M$ cached /1MBatch≈ $/mo *
DeepSeek V4 ProMID $0.435$0.87$0.003625$52.71
DeepSeek V4 FlashBUDGET $0.14$0.28$0.0028$17.19

* Example workload — chatbot, 100k requests/mo, 2,000 input / 300 output tokens per request, 70% of input cached. Computed by the same engine as the calculator. Batch: no verified batch discount published for DeepSeek at our last revision — we only list discounts we've confirmed.

Prompt caching

DeepSeek publishes the cheapest cache reads we track: DeepSeek V4 Flash at $0.0028/1M (2% of input); DeepSeek V4 Pro at $0.003625/1M (0.8% of input). The calculator models this with your cache share.

Batch / async

No verified batch discount published at our last revision — we only list discounts we've confirmed, so the Batch toggle in the calculator leaves DeepSeek prices unchanged.

Context window

DeepSeek V4 Pro and DeepSeek V4 Flash run a verified 1M-token context window.

When DeepSeek is worth it

Use caseVerdict
Cost-sensitive, high-volume workloadsDeepSeek's rates are among the lowest here
Heavy reuse of cached contextThe cached-input discount is very steep
You need a stable multi-year price guaranteeAccount for the promotional rate possibly ending
Is DeepSeek the right price for your workload?
The calculator puts these two models next to the other 26 we track — at your volume, token mix and cache share.
Open calculator

Frequently asked questions

The flagship Pro tier's rate reflects a standing promotional discount in our data. It has been in place for some time, but it is a promotion — if it ends, the list price would be materially higher. Budget for that risk on long-running workloads.
Our data does not list a separate batch (async) discount for DeepSeek, so batch is not assumed in the estimate. The steep cached-input rate is the main lever for lowering cost — see the table above.
All 28 models → OpenAI pricing → Anthropic pricing → Gemini pricing → Grok pricing → Mistral pricing → DeepSeek alternatives → Cheapest LLM API → Price changelog →