DeepSeek API Pricing: V4.1 Flash, V4 Pro and Cache Costs

A DeepSeek budget can be wrong even when its token total is correct. Old model names, expired promotions and a cache estimate that ignores output can all distort the expected bill.

This matters if your budget still uses the older V4 Flash launch price or the expired V4 Pro promotion. Those are historical prices. This guide replaces the older rate table with the current official schedule and shows how to budget traffic with different output and cache shares. Sources were checked on October 2, 2026; the calculations below are estimates, not measured invoices or fresh latency tests.

Current official DeepSeek API rates

All amounts below are USD per one million tokens, on the official DeepSeek endpoint. They are not rates from OpenRouter, SiliconFlow or another intermediary.

Source: DeepSeek Models & Pricing.

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays. In China Standard Time, the weekday peak windows are 09:00–12:00 and 14:00–18:00.

Use the period in which the provider bills the request. If you can schedule non-urgent work, off-peak pricing changes the estimate substantially. Interactive traffic should be budgeted against its actual time distribution; assuming every request qualifies for the cheaper row will understate spending.

Which API name should you use?

DeepSeek released V4.1 Flash on September 10. Use deepseek-flash for that model. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporary compatibility routes to V4.1 Flash, billed at its price. They no longer identify the retired original models.

The official change log also says V4 Pro API service continues after September 14 in response to user demand. An earlier announcement about routing Pro requests to Flash is therefore not the current integration rule. Pin the intended API name and check the change log when revisiting an integration.

The pricing page lists 1M context/384K maximum output for both current options, and image input for Flash only. Match supported inputs to the task before comparing price.

See the directory entries for V4.1 Flash and V4 Pro. Those pages separate model identity from individual hosting offers.

How to calculate a workload budget

For each billing period, calculate:

cost = uncached_input_tokens / 1,000,000 × uncached_input_rate
     + cached_input_tokens / 1,000,000 × cached_input_rate
     + output_tokens / 1,000,000 × output_rate

Count cached and uncached input separately. Do not bill the entire input total again after adding the cache-hit portion. Use the output usage reported by the provider, rather than counting only the visible answer when a request also produces billable reasoning output.

For a monthly workload of 7.5M uncached input tokens and 2.5M output tokens, with no cache hits:

The Flash peak estimate is 7.5 × $0.30 + 2.5 × $1.20. If half of those tokens are billed in each period, the same token mix costs $3.9375 for Flash or $14.85 for Pro. This example assumes the input/output mix is identical in both periods. When it differs, calculate each period independently and add the subtotals.

These are arithmetic scenarios. They exclude taxes, currency conversion, third-party markup, retries and any workload-specific additional charges.

Cache savings have a limit: output still costs money

Suppose the same 7.5M-input/2.5M-output Flash workload gets 80% input cache hits. That means 6M cached input tokens and 1.5M uncached input tokens. At peak rates:

1.5 × $0.30 + 6 × $0.006 + 2.5 × $1.20 = $3.486

The input bill falls sharply, but output still contributes $3.00. Improving cache reuse cannot remove output cost. A coding agent that generates long responses may gain more from reducing avoidable output and repeated failed attempts than from another small increase in cache hits.

Cache-hit eligibility is determined by the provider; repeating a prompt is not evidence that every token was cached. Track the API's cache usage fields and compare estimated cost with actual account usage. Set a ceiling on retries and review failed requests separately so that the successful-answer average does not hide them.

Comparing the same model across hosting providers

Compare the exact version, supported inputs, context limits, region and billing unit before comparing numbers. A third-party row for original V4 Flash is not automatically a V4.1 Flash offer. A shared family name does not establish identical service behavior.

This revision does not republish May-era third-party prices as current offers. Use the pricing matrix and each row's source and verification date, then confirm the provider's current catalog before committing a budget. An empty or hidden rate means the directory lacks a current verified offer; it does not mean access is free or discontinued.

For a useful purchasing test, run a small, fixed set of your real tasks through the routes you can actually use. Record task success, total billable tokens, retries and elapsed time. Evaluate cost per acceptable result. The cheapest token rate is a poor choice if it produces more failed work.

Common budgeting questions

Can I assume off-peak means every night in my timezone?

No. The official schedule is expressed in UTC and also excludes Chinese public holidays. Convert the specified windows into your application timezone and use the provider's billing record to confirm the result.

Does this guide prove which model is best?

No. It documents pricing and integration distinctions. No new task-quality or latency benchmark was performed for this revision. Choose between Flash and Pro using your own acceptance criteria, then apply the rate table to the resulting token usage.

How should I budget failed attempts?

Keep successful and failed attempts in the same cost ledger. A token budget that counts only completed answers hides repeated tool calls and repairs. Record retry limits alongside your expected traffic, then compare the estimate with billed usage during a small pilot.