DeepSeek V4 Flash API Pricing
¥2.00 input · ¥8.00 output per 1M tokens at peak — 50% off off-peak. The most-used coding model on Earth.
DeepSeek V4 Flash has quietly become the most-used coding model on Earth — on OpenCode Go it holds a 69% token share. It combines a 1M-token context window, a 96% real-world cache-hit rate, and a reasoning score that rivals flagship models, at prices that make it the default workhorse for high-volume AI workloads.
Price per 1M tokens
DeepSeek V4 Flash · CNY official list price
| Tier | Rate |
|---|---|
| Input | Peak ¥2.00 · Off-peak ¥1.00 |
| Output | Peak ¥8.00 · Off-peak ¥4.00 |
| Cache hit | Peak ¥0.04 · Off-peak ¥0.02 |
DeepSeek lists official prices in CNY; USD uses ≈ 6.9 exchange rate
Best for
96% cache hit rate
Real-world cache-hit rates on OpenCode Go sit at 96%, so most input rides the cached lane at ¥0.04/1M peak.
1M token context
A 1M-token context window lets whole codebases and long conversations fit in a single request.
100/100 reasoning
Normalized reasoning score of 100/100 — it reasons like a flagship model at a fraction of the price.
High-volume coding agents, batch summarization, RAG pipelines, and any repetitive workflow where cost per token matters more than absolute ceiling. Still ~4× cheaper than V4 Pro on input (and ~3.4× on output) at every hour, it's the best everyday agent model on the market.
Peak / off-peak billing
Peak hours are Beijing time 09:00–12:00 and 14:00–18:00 and bill at full price; everything else is 50% off. Cached input is billed at 1/50 of the uncached rate (¥0.04 vs ¥2.00 at peak).
V4 Flash vs the competition
At off-peak rates, V4 Flash output costs ¥4.00 ($0.58) per million tokens — lower than GPT-5.6 Luna's $1.20 and a fraction of V4 Pro's ¥8.00 peak. It wins on cached input economics and context length against both OpenAI and Anthropic.
Frequently Asked Questions
What is the price of DeepSeek V4 Flash per million tokens?
DeepSeek V4 Flash bills ¥2.00 per million input tokens and ¥8.00 per million output tokens at peak hours (Beijing 09:00–12:00 and 14:00–18:00). Off-peak hours are 50% off: ¥1.00 input / ¥4.00 output. Cached input is charged at ¥0.04 peak / ¥0.02 off-peak — about 1/50 of the uncached rate.
Is DeepSeek V4 Flash actually cheap enough for production workloads?
Yes. With a 1M-token context window, a 96% real-world cache-hit rate, and reasoning scores that rival flagship models, V4 Flash has become the most-used coding model on Earth (69% of OpenCode Go's token share). For high-volume routine work it sets the floor for cost planning.
How do DeepSeek peak and off-peak prices work for V4 Flash?
DeepSeek splits the day in half: peak hours are Beijing time 09:00–12:00 and 14:00–18:00 and bill at full price; everything else is billed at 50% off. A nightly batch job generating 100M output tokens costs ¥800 at peak but only ¥400 after 8 PM — same tokens, zero code changes.
Related model pricing
Compare this model against the rest of the lineup.
Estimate your exact cost
Model your own workload — input/output tokens, cache hit rate, and peak-hour share — against any model mix with the interactive pricing calculator.
Open the Pricing CalculatorRecommended Tools We ARE USING
We are using these tools ourselves for Development / Deployment. Check out for more details.
Opencode Go
You will receive $5 when subscribed, which directly offsets the 1st month's Go fee.
Could be the cheapest option for calling DeepSeek V4 Flash model in the world.
Vultr
RReceive $300 to test out VVultr platform
Referred user must be active 30+ days and spend $10–$25
Railway
Deploy anything without the complexity
Connect your repo, Railway handles the rest.
腾讯云 Tencent Cloud
Cloud servers, cloud databases, COS, CDN, SMS and other cloud products are on special offer now.
AWS high-quality alternative, we use COS as a replacement for S3
硅基流动 SiliconFlow
¥16 universal voucher for all platforms, valid for 180 days from the date of receipt
A comprehensive product matrix supporting the full-process implementation of AI applications.
Warp
One of the best AI terminal tools in the world
Interestingly, we can directly use various top-tier models in Warp without leaving the terminal.