The rate card
| Model | Input /1M | Output /1M | Cached read | Cache write | Batch input | Context |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash | $0.44 | $1.32 | $0.014 | — | $0.22 | 1000K |
What it costs on real workloads
Realistic monthly cost at 1,000 requests/day with the calculator’s defaults — 8% retries, 40% cache hit rate, 15% infra overhead, per-model tokenizer calibration applied. The same three workloads used across every comparison on this site, so the numbers line up.
| Workload | DeepSeek V4 Flash |
|---|---|
| Simple chatbot | $28/mo |
| Chatbot with history | $60/mo |
| Multi-step agent | $357/mo |
Where it actually lands
On the history-carrying chatbot above, DeepSeek V4 Flash runs $60/mo — rank 6 of 18 in our catalog on that workload. GPT-5.4 costs roughly 8× as much for the same work, so this is still firmly a budget-tier price. It is not, however, the cheapest row: Amazon Nova Lite, GPT-OSS 20B (Groq), Llama 3 8B Lite, Mistral Small 4, GPT-OSS 120B (Groq) all come in below it. At $0.44/1M input and $1.32/1M output, the interesting question is no longer the sticker price — it’s the caching behaviour and the data policy.
A 97% cached-read discount — the steepest here
Cached input bills at $0.014/1M, roughly 97% below standard input. Every other model in the catalog discounts cached reads by about 90%. On a workload with a large stable prefix — a long system prompt, a reused document, a growing transcript — the cached portion of your bill rounds to nothing, and that discount is now doing more of the work than the base rate is.
The 1000K context window compounds that: long prompts are both cheap to send and nearly free to re-send.
The “batch” rate is a clock, not an endpoint
DeepSeek has no batch API. What it has is off-peak pricing: every token is discounted 50% outside 01:00–04:00 and 06:00–10:00 UTC. We model that in the batch fields because the arithmetic is identical to a 50% batch tier, but the mechanism is different in a way that matters — you get the discount by scheduling work into a window, not by submitting to a queue and waiting. If your load is already off-peak-heavy, you are paying the discounted rate on real-time calls, which no batch tier gives you.
The caveat you have to price in
DeepSeek is a China-based provider, and for many teams that is a decision made outside the cost model entirely — by a procurement policy, a customer contract, a regulator, or a data residency commitment. If you handle personal data, regulated data, or anything a customer agreement covers, review the data policy before the price, because no per-token discount survives a compliance objection.
Where it fits cleanly: internal tooling, non-sensitive bulk processing, prototyping, and offline pipelines over public or synthetic data. Where it usually doesn’t: anything customer-facing under a data-processing agreement that doesn’t list it.
Endpoint naming
The legacy endpoint names deepseek-chat and deepseek-reasoner were retired on 2026-07-24; both had already been routing to V4 Flash beforehand. Use deepseek-v4-flash. If you are reading an older integration guide, that is the substitution to make.
Model DeepSeek V4 Flash against the rest of the catalog on your own workload.
Open the calculator →