We use cookies for analytics and personalized ads. We use cookies for basic analytics and, where we show ads, to personalize them. Privacy PolicyRead our Privacy Policy.
See what DeepSeek costs per 1M tokens, then estimate your own requests with your real cache-hit rate and peak-hour billing before you commit traffic to it.
Quick answer: DeepSeek V4 Flash costs $0.22 per 1M input tokens and $0.66 per 1M output tokens off-peak, and cache hits cost only $0.007. Peak-hour rates are roughly double. The API is pay-as-you-go, not free; only the web chat is free.
Rates last updated 2026-09-29
Count Tokens LocallyRequest cost = cache-miss input + cache-hit input + output — and since 16 August 2026, which hour you send it in. DeepSeek now bills roughly double during two daily UTC windows (01:00–04:00 and 06:00–10:00 UTC, weekdays). This calculator lets you toggle peak vs. off-peak directly, rather than quoting one rate and hoping it's the one you'll actually pay.
DeepSeek charges separately for new input (cache miss), repeated input (cache hit), and output. Off-peak rates apply at all times except the peak windows (01:00–04:00 and 06:00–10:00 UTC, weekdays), when every rate is roughly double.
| Model | Window | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|---|
| DeepSeek V4 Flash | off-peak | $0.22 | $0.007 | $0.66 |
| DeepSeek V4 Flash | peak | $0.44 | $0.014 | $1.32 |
| DeepSeek V4 Pro | off-peak | $0.66 | $0.022 | $1.98 |
| DeepSeek V4 Pro | peak | $1.32 | $0.044 | $3.96 |
USD per 1M tokens. Last updated 2026-09-29. DeepSeek changes prices often, so confirm on DeepSeek's official pricing page before budgeting production spend.
Cost of 1,000 requests for four common workloads, using the token shapes from the example workloads in the calculator below.
| Workload | DeepSeek V4 Flash off-peak | DeepSeek V4 Flash peak | DeepSeek V4 Pro off-peak |
|---|---|---|---|
| Support chat reply | $0.3420 | $0.6840 | $1.03 |
| Document extraction | $1.33 | $2.66 | $3.99 |
| Coding agent step | $1.23 | $2.46 | $3.7 |
| RAG answer | $0.5260 | $1.05 | $1.58 |
Set your request shape once — both models and both pricing windows update instantly below.
How much of your input is a repeated prefix (system prompt, chat history, retrieved docs) versus brand-new text each call.
At this request shape, V4 Flash is about 67% cheaper than V4 Pro ($17.1 vs $51.34 per month). Worth paying the Pro premium if your evaluations show Flash genuinely failing on this task — not simply because Pro is the newer or larger option.
DeepSeek collapsed its lineup down to two models. Both share a 1M-token context window and support a thinking and a non-thinking mode under the same model ID, so the real decision is which tier, not which mode — reasoning depth is a request-time setting, not a separate model to route to.
Same request shape as above, sorted cheapest to priciest. Rates per 1M tokens.
Covers standard text input, eligible cached reads, and generated text only — not image inputs, tool calls, or other separately billed features.
The message a user types is often the smallest part of what actually gets billed. System instructions, conversation history, retrieved documents, and tool schemas are all assembled into the final request before it ever reaches the model — count that assembled request, not the text box.
New prompt content the API hasn't seen as a matching prefix before. Every retrieved passage, every fresh instruction, and the first turn of any conversation lands here at the standard input rate.
Example: A RAG pipeline that stuffs three new documents into context each call can rack up cache-miss tokens far faster than the visible chat message suggests.
Trap: Assuming a prompt 'looks the same' each time is enough to guarantee a cache hit — a single changed character near the start of the prefix breaks the match.
The portion of input tokens matching a previously seen prefix, billed at a steep discount. DeepSeek's automatic caching rewards stable system prompts and consistent ordering — reusable content first, variable content last.
Example: A support bot with a fixed system prompt and growing chat history can push a large share of later-turn input into cache-hit pricing.
Trap: Putting the variable, per-user part of a prompt before the stable instructions, which prevents the stable part from ever matching a cached prefix.
Output is priced well above input on both tiers, and thinking mode's visible reasoning tokens count toward this total the same as any other output token.
Example: Leaving thinking mode on for a simple classification task can multiply output tokens for no quality benefit on that task.
Trap: Not setting a realistic output ceiling, so a small share of unusually long responses dominates the average monthly bill.
Since 16 August 2026, the same request costs roughly double during DeepSeek's peak UTC hours compared to off-peak — a dimension no other major provider currently prices this way.
Example: A nightly batch job scheduled at 02:00 UTC lands inside a peak window and pays double what the same job would cost run two hours later.
Trap: Forecasting a monthly bill using only the off-peak rate because that's the number most blog posts lead with.
A docs assistant sends a fixed 800-token system prompt plus 2,400 tokens of retrieved documentation and chat history on every call, and generates a 300-token answer with V4 Flash. The system prompt and the first half of the retrieved context repeat often enough to hit the cache about 55% of the time; the rest is new each call. It runs 40,000 times a month, entirely outside peak hours.
Without the cache hit, that same request would cost roughly $0.000902, or about $36.08/month — the caching discount alone accounts for roughly 42% of this bot's monthly spend, which is why keeping the stable part of a prompt first and consistent matters more on DeepSeek than the raw per-token rate suggests.
The same open DeepSeek weights are served by several third-party hosts alongside DeepSeek's own API, often at broadly similar per-token rates but with different uptime guarantees, regional latency, and rate limits.
DeepSeek publishes open model weights, so a team with enough GPU capacity can run inference in-house instead of paying per token at all — at the cost of owning the serving stack, scaling, and reliability engineering.
Both V4 tiers expose thinking and non-thinking mode on the same model ID, so reasoning depth is a request-time parameter rather than a model choice that locks you into a price tier.
Any job that doesn't need to run the moment it's triggered can be scheduled outside DeepSeek's peak UTC windows to pay the off-peak rate automatically, with zero change to the model or prompt.
The numbers above are only useful once you actually have a key and somewhere to send requests. Here's exactly how that goes, including the coding-tool setup people ask about most.
Create an account on DeepSeek's own platform to generate a key billed at the rates modeled above, or get one through a third-party host (OpenRouter, Together, Fireworks) if you'd rather consolidate billing with other models you already call. Either way, treat the key like any other API secret — load it from an environment variable, never commit it to a repo.
Coding assistants that support custom or OpenAI-compatible endpoints can point at DeepSeek directly — set the base URL and API key in the tool's model settings, pick V4 Flash for routine edits, and reserve V4 Pro for the harder refactors and multi-file agentic steps where Flash's answers start slipping.
Agentic coding sessions run more input tokens than a simple chat turn — file context, tool output, and prior steps all get resent on every turn. Load the Coding agent step preset in the calculator above, set your monthly volume to a realistic number of turns for how you'd actually use it (not just one test session), and compare V4 Flash against V4 Pro before deciding which one becomes your default in the editor.
Drop the DeepSeek calculator into your own docs or internal wiki with a single iframe. The embedded version keeps a small credit link back to Apex Tech Journal.
<iframe
src="https://apextechjournal.com/embed/deepseek-token-calculator?ref=widget"
width="100%"
height="440"
style="border:0;border-radius:12px;"
loading="lazy"
title="DeepSeek Token Calculator by Apex Tech Journal"
></iframe>