DeepSeek V4 Pro API Pricing

DeepSeek V4 Pro API pricing now changes by time of day. Direct DeepSeek API usage costs $0.022 per million cache-hit input tokens, $0.66 per million cache-miss input tokens, and $1.98 per million output tokens off-peak. During peak hours, those rates double to $0.044, $1.32, and $3.96.

The August 13 release is DeepSeek-V4-Pro-0813, but the API model ID remains deepseek-v4-pro. You do not put 0813 in the API request.

DeepSeek V4 Pro 0813 at a glance

All prices are USD per one million tokens. Every time outside the two listed windows is off-peak and costs half the peak rate.

Official DeepSeek V4 Pro API price table

The general calculation is:

cost = cache_hit_input_millions x cache_hit_rate
     + cache_miss_input_millions x cache_miss_rate
     + output_millions x output_rate

DeepSeek defines the windows in UTC but does not state on this price page how a request crossing a boundary is split. Keep budgets and logs in UTC, use the provider's billed usage as the source of truth, and confirm the edge case before scheduling very long requests around a boundary.

The DeepSeek V4 Pro model page owns the current OmniAKey quote, availability, and request reference. The official table above describes direct DeepSeek billing. A gateway price and DeepSeek's time-based direct price are separate billing relationships, so verify the live model catalog instead of assuming they match.

What “0813” changes

DeepSeek-V4-Pro-0813 is the downloadable GA snapshot that supersedes V4 Pro Preview. DeepSeek kept the hosted API alias unchanged:

{
  "model": "deepseek-v4-pro"
}

The release keeps the preview architecture and adds a DSpark speculative decoding module. DeepSeek also publishes the model weights under the MIT license. Those facts matter if you self-host; they do not turn local inference into a free service because hardware, serving, storage, and operations still have costs.

For the hosted API, the practical changes are:

• stronger vendor-reported agent benchmark results;
• three explicit reasoning levels: low, high, and max;
• native Responses API support with Codex optimization;
• a 1M context window and recommended maximum 384K output for high and max;
• the new peak and off-peak rate card.

The alias behavior is useful for a quick upgrade, but a moving alias is less reproducible than a downloadable snapshot. Record the request date, API model ID, reasoning level, harness version, and acceptance test when comparing results.

Cost example 1: cache-miss input

Assume one coding task uses 100,000 cache-miss input tokens and 20,000 output tokens.

Moving the identical token workload outside the peak windows saves $0.1056. It does not make the task more token-efficient; it only changes the rate applied to those tokens.

Cost example 2: cache-hit input

Now assume the same 100,000 input tokens qualify as cache hits and the model still produces 20,000 output tokens.

Here output dominates the bill. A warm input prefix helps, but verbose reasoning, long final answers, repeated tool loops, and retries can still outweigh the cache saving.

How DeepSeek context caching works

DeepSeek enables disk context caching by default. A later request receives the cache-hit rate only when its prefix fully matches a persisted cache-prefix unit. Similar wording or a partially reused file is not automatically a hit.

Inspect the response usage fields:

• prompt_cache_hit_tokens reports input tokens billed at the cache-hit rate;
• prompt_cache_miss_tokens reports input tokens billed at the cache-miss rate.

Caching is best effort. Construction takes time, and unused cache data is normally cleared after hours or days. Budget from the reported fields, not from the number of tokens you expected to reuse.

Stable system instructions, tool schemas, repository guidance, and large reference documents are good cache candidates. Put changing instructions after the stable prefix when the client lets you control request layout.

Thinking tokens are output tokens

Thinking mode is enabled by default at high. DeepSeek maps OpenAI-style medium and xhigh requests to high; the model's actual supported levels are low, high, and max.

On Chat Completions, set the reasoning_effort parameter to one of those three supported values.

In the Responses API, reasoning usage appears under usage.output_tokens_details.reasoning_tokens. It is part of output usage, so do not estimate cost from visible answer text alone. A short final answer can still carry substantial reasoning output.

Use low for straightforward extraction or transformation, high for normal agent work, and test max only when the higher accepted-task rate justifies more output, latency, and cost. The 384K maximum is capacity, not a recommended default.

For workflow-level context control, the token-usage reduction guide explains how to remove stale context and unbounded tool output before lowering model quality.

V4 Pro 0813 benchmark changes

DeepSeek reports the following gains over V4 Pro Preview:

Source: DeepSeek (vendor-reported; not independently reproduced). For public code-agent benchmarks, DeepSeek used the minimal mode of DeepSeek Harness, max reasoning effort, temperature = 1.0, and top_p = 0.95. Two other results in its table, DSBench-FullStack and DSBench-Hard, are internal test sets.

These numbers support the claim that 0813 improved over Preview under DeepSeek's evaluation setup. They do not prove that it beats every model, client, repository, or production workload. Reproduce a representative task with the same inputs, permissions, tools, and pass/fail check.

The DeepSeek Harness setup guide owns installation and custom-provider configuration. This article owns the price and 0813 cost decision.

DeepSeek V4 Flash costs about one-third as much

DeepSeek lists V4 Flash at about one-third of the Pro rates. Cache-miss input and output are exactly one-third in the displayed table; the rounded cache-hit figure is slightly lower than one-third.

Its API ID is deepseek-v4-flash, and the current snapshot is DeepSeek-V4-Flash-0731. Flash also has a higher listed concurrency limit: 2,500 versus 500 for Pro.

That makes Flash the cheaper first test for high-volume, well-specified work. It does not establish equal quality. Compare accepted-task cost, including retries and correction time, before routing an agent workload permanently.

Call DeepSeek V4 Pro through OmniAKey

Create a scoped key in the API Keys dashboard, confirm that deepseek-v4-pro is present in the live catalog, and call the OpenAI-compatible route:

curl https://api.omniakey.com/v1/chat/completions \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [
      {"role": "user", "content": "Find the highest-risk assumption in this migration plan."}
    ],
    "stream": true
  }'

Use the exact gateway model ID shown in the catalog. The API quick start documents the endpoint and authentication pattern.

After a representative run, inspect the usage dashboard. Do not apply DeepSeek's direct peak or off-peak table to an OmniAKey usage record unless the live OmniAKey quote explicitly says that billing rule applies.

The transparent billing guide explains how model, token, cache, and cost line items should be reconciled.

A practical scheduling and budget policy

Use this sequence:

1. Keep user-facing and urgent work latency-driven; do not delay it solely for a discount.
2. Move batch evaluation, indexing, nightly analysis, and retryable jobs outside 01:00-04:00 and 06:00-10:00 UTC when using DeepSeek direct.
3. Record cache-hit, cache-miss, reasoning, and output tokens separately.
4. Set one spending cap per person, environment, or automation role.
5. Compare Pro with Flash on the same accepted-task check.
6. Recheck the official price page before changing a production schedule.

A spending cap limits exposure but does not reduce token use. Optimize the request first, then set a cap with enough room for normal task variance.

Common DeepSeek V4 Pro pricing mistakes

Sending DeepSeek-V4-Pro-0813 as the API model

That is the snapshot name. The hosted API model ID remains deepseek-v4-pro.

Treating every hour as off-peak

Peak pricing applies from 01:00-04:00 and 06:00-10:00 UTC. The other 17 hours use the half-price table.

Assuming repeated text is automatically cached

The request must fully match a persisted cache-prefix unit. Trust prompt_cache_hit_tokens, not visual similarity.

Treating 1M context or 384K output as included credit

Both are capacity limits. Tokens placed into those windows are still metered.

Ignoring hidden reasoning output

Reasoning tokens count inside output usage. A short visible response is not necessarily a low-output request.

Calling vendor benchmarks independent evidence

The 0813 gains are vendor-reported under a disclosed Harness and max-effort setup. They are a useful upgrade signal, not a universal production ranking.

Frequently asked questions

How much does the DeepSeek V4 Pro API cost?

Off-peak, DeepSeek lists $0.022 per million cache-hit input tokens, $0.66 per million cache-miss input tokens, and $1.98 per million output tokens. Peak rates are $0.044, $1.32, and $3.96.

When is DeepSeek V4 Pro off-peak?

Every UTC hour outside 01:00-04:00 and 06:00-10:00 is off-peak. Off-peak rates are 50% below peak rates.

What is the DeepSeek V4 Pro 0813 API model ID?

Use deepseek-v4-pro. DeepSeek-V4-Pro-0813 names the released weights and snapshot, not the hosted API request ID.

Does DeepSeek V4 Pro support one million tokens?

Yes. DeepSeek lists a 1M context window and a 384K maximum output. These are limits, not free allowances.

Is DeepSeek V4 Pro cheaper than V4 Flash?

No. Flash is about one-third of Pro's listed rates in both time windows. Pro should justify its premium through a higher accepted-task rate on your workload.

Did 0813 improve coding-agent performance?

DeepSeek reports large gains over its Pro Preview on Terminal Bench 2.1, DeepSWE, and Toolathlon-Verified. The public code-agent runs used DeepSeek Harness minimal mode and max effort. Treat them as vendor-reported until reproduced on your own tasks.

Sources checked

• DeepSeek V4 Pro GA release
• DeepSeek models and pricing
• DeepSeek API model aliases
• DeepSeek thinking mode
• DeepSeek context caching
• DeepSeek-V4-Pro-0813 model card and weights
• OmniAKey DeepSeek V4 Pro model page
• OmniAKey API quick start

Fact-checked August 18, 2026. Provider prices, peak windows, model aliases, gateway availability, and supported features can change. Verify the linked first-party documentation and live catalog before committing a production budget.