Long context is useful only if you can afford the tokens. On the 17 September 2026 pricing page, deepseek-flash and deepseek-v4-pro both list 1M context and 384K max output. Filling a million tokens at peak rates gets expensive fast—use retrieval, trimming, and cached input. Always recheck platform.deepseek.com for the live limits.
What the card actually says
Per the 17 September 2026 pricing page, Flash and Pro both advertise 1M context and 384K max output. Flash adds vision and concurrency cited at 2,500; Pro skips vision with concurrency 500. Long context does not change the per-million rates: Flash peak $0.30/$1.20, Pro $1.32/$3.96, off-peak half, cached input much cheaper. Recheck platform.deepseek.com—marketing ceilings can be revised. Treat every rate and limit as a dated snapshot from DeepSeek or the other vendor docs linked in sources, not as a permanent contract price. Recheck platform.deepseek.com for Flash and Pro cards, and confirm peak UTC windows on
Cost control patterns
Do not paste entire monorepos on every turn. Prefer search, summaries, and file slices. Put stable instructions in a cacheable prefix. Schedule fat jobs in off-peak when your UTC math allows—India daytime often sits outside the described peak windows (weekdays 01:00–04:00 and 06:00–10:00 UTC), but confirm the clock on DeepSeek’s pricing page. Cap max tokens so a runaway completion cannot emit toward the 384K ceiling unnoticed. Treat every rate and limit as a dated snapshot from DeepSeek or the other vendor docs linked in sources, not as a permanent contract price. Recheck platform.deepseek.com for Flash
When long context is the wrong tool
If the task is a 40-line bugfix, 1M context is irrelevant—use a tight prompt. If policy forbids sending whole codebases to a third party, long context is a compliance problem, not a feature. In those cases, local RAG with redaction beats dumping files. Pair DeepSeek API jobs with editor agents carefully so you do not send the same large payload twice through Cursor pools and DeepSeek keys. Treat every rate and limit as a dated snapshot from DeepSeek or the other vendor docs linked in sources, not as a permanent contract price. Recheck platform.deepseek.com for
Related on this site
Frequently asked questions
-
Do Flash and Pro both have 1M context?
Yes on the 17 September 2026 pricing page for both. Confirm live on platform.deepseek.com before you design around that number. Confirm live details on the vendor docs linked in this brief before you quote a client or ship production traffic. Attribute DeepSeek Flash and Pro figures to the 17 September 2026 pricing page when you reuse them, then recheck platform.deepseek.com.
-
Is 384K max output free?
No. Output tokens bill at the model’s output rate. The cap is a limit, not a free allotment. Set lower max_tokens for normal work. Confirm live details on the vendor docs linked in this brief before you quote a client or ship production traffic. Attribute DeepSeek Flash and Pro figures to the 17 September 2026 pricing page when you reuse them, then recheck platform.deepseek.com.
-
Does free chat offer 1M context?
This brief’s 1M figure is from the API model cards. Do not assume the consumer chat UI matchesu2014check the product you use. Confirm live details on the vendor docs linked in this brief before you quote a client or ship production traffic. Attribute DeepSeek Flash and Pro figures to the 17 September 2026 pricing page when you reuse them, then recheck platform.deepseek.com.
-
How does cache help long prompts?
Cached input is priced much lower than fresh input (Flash $0.006/$0.003, Pro $0.044/$0.022 peak/off-peak per million on 17 Sep 2026). Stable prefixes pay off. Recheck live rates. Confirm live details on the vendor docs linked in this brief before you quote a client or ship production traffic. Attribute DeepSeek Flash and Pro figures to the 17 September 2026 pricing page when you reuse them, then recheck platform.deepseek.com.
Sources
Bdeb Technology builds websites, WordPress systems, and tools on top of models like these. A written quote comes back within 24 hours.