DeepSeek V4 Pricing: Flash, Pro, and Cache Costs
On this page
Answer first
DeepSeek V4 pricing has three cost axes: Flash versus Pro, cache hit versus cache miss, and peak versus off-peak time. Do not copy one low cache-hit number into a general cost claim. Measure the input state, output volume, request time, retries, and accepted result.
The official page checked on August 20, 2026 lists a 1M context length, 384K maximum output, thinking and non-thinking modes, and separate deepseek-v4-flash and deepseek-v4-pro model names.
Current public rates
Prices below are US dollars per 1 million tokens from the official API documentation:
| Model and token class | Off-peak | Peak |
|---|---|---|
| V4 Flash cache-hit input | $0.007 | $0.014 |
| V4 Flash cache-miss input | $0.22 | $0.44 |
| V4 Flash output | $0.66 | $1.32 |
| V4 Pro cache-hit input | $0.022 | $0.044 |
| V4 Pro cache-miss input | $0.66 | $1.32 |
| V4 Pro output | $1.98 | $3.96 |
DeepSeek currently defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, with other hours charged at off-peak rates. Product prices and windows can change, so the live billing page and account remain authoritative.
Estimate the completed-task cost
Use this model for each accepted task:
cost = cache-hit input × hit rate + cache-miss input × miss rate + output × output rate + retries + external tools + review time
Run the same representative tasks through Flash and Pro. Record:
- exact model version and thinking mode;
- peak or off-peak timestamp;
- billed cache-hit and cache-miss tokens;
- output and reasoning volume;
- retries, tool calls, latency, and errors;
- acceptance result and human correction time.
Flash is the lower token-price candidate. Pro is justified only when its measured quality or reliability reduces enough retries or review work to offset the higher rate.
Cache and scheduling boundaries
A repeated prompt does not prove a billed cache hit. Keep stable prefixes where the provider contract supports them, then verify the usage record. Do not move latency-sensitive or regulated work into an off-peak batch merely to reach a lower rate without checking completion deadlines and data controls.
The API offers OpenAI-format and Anthropic-format base paths. Compatibility does not make model behavior, error semantics, tools, or billing identical to those providers. Pin the DeepSeek model and test the exact request contract.
Related decisions
- Use the DeepSeek V4 model guide for model identity, capability, and access boundaries.
- Compare another current model family with the GPT-5.6 guide.
- Use the AI content brief generator when the real job is planning a reviewed writing workflow rather than choosing an API.
Frequently asked questions
How much does DeepSeek V4 Flash cost?
At the source check, peak rates were $0.014 cache-hit input, $0.44 cache-miss input, and $1.32 output per million tokens. Off-peak rates were half those amounts.
How much does DeepSeek V4 Pro cost?
Peak rates were $0.044 cache-hit input, $1.32 cache-miss input, and $3.96 output per million tokens. Off-peak rates were half those amounts.
Which DeepSeek V4 model is cheaper?
Flash has lower token prices. The cheaper completed task still depends on success, total tokens, real cache hits, retries, latency, tools, and review effort.
Official source
Source check: August 20, 2026. Recheck model versions, token rates, peak windows, context, output limit, concurrency, balance rules, and account availability before use.