Comparison · Chat & code
DeepSeek V4 Flash vs DeepSeek V4 Pro
DeepSeek V4 Flash vs V4 Pro: rates, size, thinking modes and the shared 1M context. See when the cheaper Flash is enough and when Pro is worth paying for.
What each one accepts
| DeepSeek V4 Flash | DeepSeek V4 Pro | |
|---|---|---|
| Vendor | DeepSeek | DeepSeek |
| Type | Chat & code | Chat & code |
| Released | 2026-04 | 2026-08 |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Max output | 384,000 tokens | 384,000 tokens |
| Endpoints | /v1/chat/completions | /v1/chat/completions |
Token prices
| DeepSeek V4 Flash | DeepSeek V4 Pro | |
|---|---|---|
| Input / 1M tokens | $0.135 | $0.54 |
| Output / 1M tokens | $0.27 | $1.22 |
| Input from cache / 1M | $0.0378 | same as input |
| 1,000 replies, 1,500 in / 400 out | $0.31 | $1.30 |
Charged from a prepaid USD balance, no taxes added. One key runs both, through /v1/chat/completions in the OpenAI format. A failed generation is $0.
I start every DeepSeek job on V4 Flash. Within the V4 family it is the compact option, and its token rates on input and output are a fraction of what V4 Pro charges. V4 Pro is the one to use when your own data shows the bigger model winning, or when an existing pipeline was built around its answers.
When to pick DeepSeek V4 Flash
- A coding agent pulls big chunks of a codebase into each request, so input volume is huge.
- You classify, extract or summarize long documents in bulk.
- You run support chat or other high-traffic flows.
- You are prototyping and expect thousands of test calls.
When to pick DeepSeek V4 Pro
- On your own hard examples, Pro's answers come out clearly stronger.
- Regression tests in production were written against V4 Pro output.
- You analyze long context and Pro handles your prompts better.
- You want an A/B baseline before deciding to move traffic to Flash.
Differences you will notice
- Rates. DeepSeek V4 Flash is billed at $0.135 / $0.27 per 1M tokens and DeepSeek V4 Pro at $0.54 / $1.22 per 1M tokens. Flash is cheaper on input and output, and the cost of 1,000 replies above puts a number on the gap.
- Size. V4 Pro uses a MoE design: 1.6T parameters in total, of which 49B work on each token. Flash is the lighter sibling, built for cheap volume.
- History. Pro spent April to August 2026 in preview; general availability came in August.
- Reasoning. Each model reasons first unless told otherwise, with three effort levels: low, high and max. Reasoning is billed as output, so high effort raises the output share of the bill.
- Shared ground. A 1,000,000-token window, answers as long as 384,000 tokens, published weights, function calling, JSON output and the same OpenAI-style endpoint.
How to decide
Price the task first. Output is the pricier side on both models, so long answers with high effort cost the most. When I measured tokenizers on a Polish sample (write-up in Polish on promptowy.com), V4 Pro came out roughly 5% above OpenAI models; I did not include Flash, so I have no figure for it. If Flash fails on one task type, run that type on Pro in parallel and compare. Switching is one string in model. For more budget picks see Cheapest LLM models, and other pairs are on the comparisons page.
Questions
Will V4 Pro stay available?
DeepSeek once announced that V4 Pro requests would be rerouted to a newer model, then changed its mind and kept Pro online. A forced move may still happen, so store the model id as a setting, not a constant in code.
What do I change in code to go from Pro to Flash?
Just the model value, deepseek-v4-pro becomes deepseek-v4-flash. URL, key and JSON format do not change. What you need to verify is answer quality on your own tasks.
Do I need a separate DeepSeek account?
No. Your Kadraris key reaches both models from one prepaid balance, and your prompts and outputs are not used to train models.