DeepSeek Pricing Calculator for V4-Flash and V4-Pro API Costs

You opened a DeepSeek pricing calculator because the rate card alone does not answer the question you have: what will the monthly bill look like at this volume? A real estimate splits input into cache hits and cache misses, adds output separately, and uses the right model and pricing period — peak or off-peak. Get any of those wrong and the forecast can be off by a factor of two or more.

This guide walks through DeepSeek’s official rate card as of 22 October 2026, gives a copy-paste calculation template for deepseek-flash and deepseek-v4-pro, and prices one illustrative chatbot workload on each, at peak and off-peak rates. Looking for V4-Flash? It was retired on 10 September 2026, when V4.1-Flash launched; its successor on the API is deepseek-flash, served by DeepSeek-V4.1-Flash-1017.

The DeepSeek rate card you are calculating against

The calculator has two tabs. Single model takes a model — deepseek-flash (DeepSeek-V4.1-Flash-1017) or deepseek-v4-pro (DeepSeek-V4-Pro-0813) — input tokens (default 100,000), output tokens (default 20,000) and a cache-hit ratio slider (default 50%), and prices Input · cache hit, Input · cache miss, Output and Total at Peak and Off-peak side by side. Compare models prices the workload on both models, says how many times as much deepseek-v4-pro costs and adds verified competitor rows. A line shows whether the current UTC time is in a published peak window; it cannot detect Chinese public holidays and says so.

Non-USD amounts (EUR, GBP, JPY, INR, CNY, AUD, CAD, BRL) are approximate conversions of the USD amounts at the European Central Bank’s daily euro reference rates (date shown), fetched and cross-converted to USD by our server. A daily job re-reads DeepSeek’s pricing page and keeps the last good prices whenever a fetch fails validation; the footnote shows the date of the last successful fetch.

The API serves two models: deepseek-flash runs DeepSeek-V4.1-Flash-1017, the 17 October 2026 build of V4.1-Flash, and deepseek-v4-pro runs DeepSeek-V4-Pro-0813, the 13 August 2026 general-availability release. Both follow DeepSeek V4, an open-weight Preview launched on 24 April 2026. Our DeepSeek V4-Pro page covers the Pro line; DeepSeek V4-Flash covers the predecessor of today’s Flash model. Pro costs 3.3× Flash on output, 4.4× on cache-miss input and about 7.3× on cache-hit input, in both periods.

Per-million-token rates as of 22 October 2026 (peak and off-peak)

Source: DeepSeek’s official pricing page (last modified 19 October 2026, checked 22 October 2026). Both models have a 1M-token context window and up to 384K output tokens; only deepseek-flash takes images.

Two points that are easy to get wrong:

• Off-peak is half price. Since 16:00 UTC on 16 August 2026 (Change Log, entry of 13 August 2026), off-peak rates have been 50% of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak, including weekends and those holidays in full. An earlier off-peak discount ended on 5 September 2025.
• Thinking mode does not change the rate. The model and the time of the request set the rate, not thinking mode (the default) versus non-thinking mode. Thinking does add tokens: its chain of thought, returned in reasoning_content alongside content, is billed as output.

How a correct cost calculation is structured

Every accurate estimate enumerates three token buckets and names one model and one pricing period — or shows both periods side by side, as the calculator does.

Rates: one model, one period (peak or off-peak)

Input, cache hit  :  X tokens ÷ 1,000,000 × hit_rate   = $A
Input, cache miss :  Y tokens ÷ 1,000,000 × miss_rate  = $B
Output            :  Z tokens ÷ 1,000,000 × out_rate   = $C
                                                         ------
Total             :                                      $(A + B + C)

The cache-hit bucket is the input billed at the cache-hit rate — 1/50 of the miss rate on deepseek-flash, 1/30 on deepseek-v4-pro. Caching is on by default (Context Caching guide, checked 22 October 2026). A request hits only where its prefix fully matches a stored cache prefix unit, so a reworded system prompt misses. Caching is best-effort, and user_id isolates the KV cache between your users (Rate Limit & Isolation).

The cache-miss bucket is the rest of the input — typically the user’s new message plus any per-call context.

For prefix matching, see our notes on DeepSeek context caching; pair this with a DeepSeek token counter if you lack token counts.

Worked example A: chatbot on deepseek-flash

A hypothetical workload, priced as illustrative arithmetic rather than a measured bill: one million calls a month, each with a 2,000-token system prompt assumed to hit the cache every time (optimistic — first requests with a new prefix miss), a 200-token user message that misses, and a 300-token response that includes any reasoning tokens.

Each cell is tokens ÷ 1,000,000 × rate (e.g. 2,000,000,000 ÷ 1,000,000 × $0.006 = $12.00), exact to the cent. The month totals $432.00 at peak or $216.00 off-peak ($0.000432 or $0.000216 per call). Without the cache hits, the prefix would cost 2,000,000,000 ÷ 1,000,000 × ($0.30 − $0.006) = $588.00 more at peak, or half that, $294.00, off-peak.

In the calculator, 2,200,000,000 input and 300,000,000 output tokens at a 91% cache-hit ratio (nearest 1% step) give $431.412 at peak (2,002 × $0.006 + 198 × $0.30 + 300 × $1.20, tokens in millions) and half that, $215.706, off-peak, shown rounded to $431.41 and $215.71.

The Compare models tab also prices five competitor models at each vendor’s standard (non-batch) rates, with no cache hits on either side, so each total is exactly 2,200,000,000 ÷ 1,000,000 × input rate + 300,000,000 ÷ 1,000,000 × output rate: the deepseek-flash baseline ($0.30 / $1.20 per 1M in / out) $1,020.00 at peak, or ($0.15 / $0.60) $510.00 off-peak; GPT-4o mini ($0.15 / $0.60) $510.00, Claude Haiku 4.5 ($1.00 / $5.00) $3,700.00, Gemini 3.1 Pro Preview ($2.00 / $12.00) $8,000.00, GPT-4o ($2.50 / $10.00) $8,500.00, Claude Sonnet 4.6 ($3.00 / $15.00) $11,100.00. Uncached, GPT-4o mini is cheaper than deepseek-flash at peak and equal off-peak. Rates as of 22 October 2026 from OpenAI (shutdown of both announced for 14 January 2027), Anthropic (both listed as legacy) and Google (prompts up to 200K tokens); vendors’ cache discounts and tokenizer differences are not modelled.

Worked example B: the same workload on deepseek-v4-pro

Same hypothetical traffic and assumptions, at deepseek-v4-pro rates:

That is $1,540.00 ÷ $432.00 ≈ 3.56× the deepseek-flash bill at peak, and the same off-peak ($770.00 ÷ $216.00); the Compare models tab rounds it to 3.6×.

On quality, DeepSeek’s 10 September 2026 release note claims V4.1-Flash benchmarks ahead of flagship models including V4-Pro (a vendor claim, not an independent result); the Change Log says V4-Pro stayed available after 14 September 2026 in response to user demand. Default to deepseek-flash; keep deepseek-v4-pro to compare against, or for workflows already validated on it.

Using the calculator’s formula in your code

Chat requests go to POST /chat/completions at https://api.deepseek.com. The API is stateless, so each call resends the conversation history (Multi-round Conversation guide). This sample makes one thinking-mode call as the official docs show it, then prices it:

import os
from datetime import datetime, timezone
from openai import OpenAI

client = OpenAI(api_key=os.environ.get("DEEPSEEK_API_KEY"),
                base_url="https://api.deepseek.com")
resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[
        {"role": "system", "content": "You are a concise code reviewer."},
        {"role": "user", "content": "Summarise this PR."},
    ],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

# deepseek-flash, USD per 1M tokens (hit, miss, output), 22 October 2026
RATES = {"peak": (0.006, 0.30, 1.20), "off_peak": (0.003, 0.15, 0.60)}
t = datetime.now(timezone.utc)  # Chinese public holidays not handled
peak = t.weekday() < 5 and (1 <= t.hour < 4 or 6 <= t.hour < 10)
u = resp.usage
tokens = (u.prompt_cache_hit_tokens, u.prompt_cache_miss_tokens, u.completion_tokens)
rates = RATES["peak" if peak else "off_peak"]
print(tokens, sum(n / 1_000_000 * r for n, r in zip(tokens, rates)))

In usage, prompt_cache_hit_tokens plus prompt_cache_miss_tokens equals prompt_tokens, and completion_tokens_details.reasoning_tokens gives the reasoning share of completion_tokens (API reference, checked 22 October 2026). Price each request at the rate for the period in which it ran; the sample uses the response’s arrival time instead, so its result is an estimate.

Thinking is on by default (effort high) on both models. The official syntax is reasoning_effort="high" with extra_body={"thinking": {"type": "enabled"}}; {"type": "disabled"} turns it off, and efforts are low, high and max (Thinking Mode guide). Thinking mode ignores temperature without an error.

Legacy names: deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but retired, and served by DeepSeek-V4.1-Flash-1017 at the Flash price. deepseek-chat and deepseek-reasoner were retired (scheduled for 24 July 2026, 15:59 UTC) and are no longer listed. To migrate, switch model= to deepseek-flash (same base_url); for deepseek-chat’s non-thinking behaviour, add "thinking": {"type": "disabled"} (via extra_body in the Python SDK). The DeepSeek API pricing reference has the full rate card.

Common calculator mistakes that produce wrong numbers

1. Mixing models or periods. Pairing deepseek-flash’s $0.30 peak cache-miss input with deepseek-v4-pro’s $3.96 peak output, or peak and off-peak rates for one request, is a math error.
2. Skipping the uncached-input line. A cached system prompt does not absorb the user’s new message.
3. Counting reasoning tokens as input. Thinking is the default, and reasoning tokens bill as output.
4. Assuming a higher cache-hit ratio than your usage data shows. Set the slider from the prompt_cache_hit_tokens your responses report.
5. Ignoring the pricing period. One rate around the clock misprices traffic by up to a factor of two. In a normal week, 35 of 168 hours (≈20.8%) are peak.

JSON mode, output caps and the budget headroom you need

With response_format={"type": "json_object"}, set max_tokens high enough that the response cannot truncate mid-object: a truncated response is invalid JSON, and you have still paid for it. DeepSeek’s JSON Output guide (checked 22 October 2026) asks for the word “json” in the prompt plus an example of the format, and warns that the API may occasionally return empty content; the API reference adds that without such an instruction the model may emit whitespace up to the token limit. Patterns are in the DeepSeek API JSON mode guide.

For broader budgeting, the DeepSeek cost estimator covers monthly and annual projections at peak, off-peak and blended rates, and the wider set of DeepSeek tools and utilities includes hardware sizing (rough estimates) for the self-hosted V4 Preview open weights.

DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Cost estimates here are arithmetic on published vendor prices; we do not run our own evaluations. Model IDs, pricing and API behaviour change; check the official documentation and pricing page before committing to a production decision.