Cheapest DeepSeek API model in 2026: deepseek-flash vs Pro

Your DeepSeek bill depends on three things the price per token doesn't show: which model you call, what hour you call it, and how much of your prompt the cache remembers. The short answer: of the two official models, deepseek-flash is the cheaper one in every column. Off-peak it costs $0.15 per 1M input tokens (cache miss) and $0.60 per 1M output tokens. At peak, both double.

The prices below come from DeepSeek's pricing page, fetched on 29 October 2026. The page carries no publication date, so treat the fetch date as the price date.

Two official models, one rename#

The official API sells two models: deepseek-flash, served by DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, served by DeepSeek-V4-Pro-0813. Both list a 1M context and up to 384K tokens of output.

If your code still says deepseek-v4-flash or deepseek-v4-flash-vision-exp, nothing breaks. DeepSeek's 17 October 2026 release note says those models are retired and their names now route to DeepSeek-V4.1-Flash-1017 at the Flash price. What changed inside the model is covered in the write-up of the release.

Here are the prices per 1M tokens, in USD:

"Cache hit" means DeepSeek recognises the beginning of your prompt from an earlier request and bills that part at the cheap rate. Prompt caching explained has the mechanics. Everything that isn't a hit is a miss.

Peak hours: same request, double the price#

Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, except Chinese public holidays. Outside those windows you pay the off-peak price, which is half.

Seven hours a day sounds like little. It is: 7 × 5 = 35 of the week's 168 hours, about a fifth. But those windows cover most of the Beijing working day and the European morning, so a bot that talks to people in those time zones lives inside them. A job that runs at 22:00 UTC never sees peak pricing.

A support bot: 1,000 requests a day#

Say you run a support bot. Each request carries about 1,500 input tokens (the instructions plus the customer's message) and gets back about 400 tokens. That's 1.5M input and 0.4M output tokens a day, and for simplicity none of it hits the cache.

On deepseek-flash at peak, input is 1.5 × $0.30 = $0.45. Output is 0.4 × $1.20 = $0.48. A day comes to $0.45 + $0.48 = $0.93. Off-peak, input is 1.5 × $0.15 = $0.225 and output is 0.4 × $0.60 = $0.24, so a day is $0.225 + $0.24 = $0.465.

On deepseek-v4-pro at peak, input is 1.5 × $1.32 = $1.98. Output is 0.4 × $3.96 = $1.584. A day comes to $1.98 + $1.584 = $3.564.

Look at where the money goes. On Flash the 400-token answers cost slightly more than the 1,500-token prompts, because output is the expensive side. And Pro is the model you pick after a test on your own tickets shows Flash's answers falling short, not before: that test is worth $3.564 a day at peak.

An agent loop: where the cache pays#

A coding agent is the opposite of a chatbot. Every step re-sends the same big prefix: instructions, tool definitions, the files it has already read. Say a run is 100 steps with 30,000 input tokens and 2,000 output tokens each, and 80% of every prompt is that repeated prefix. That's 3M input tokens (2.4M hits, 0.6M misses) and 0.2M output tokens.

On deepseek-flash at peak, with no cache at all: 3 × $0.30 = $0.90 for input plus 0.2 × $1.20 = $0.24 for output, $0.90 + $0.24 = $1.14.

With the 80% hit rate: 2.4 × $0.006 = $0.0144 for the hits, 0.6 × $0.30 = $0.18 for the misses, and the same $0.24 for output, $0.0144 + $0.18 + $0.24 = $0.4344.

The cache took input from $0.90 to $0.1944 and made output the biggest line on the bill. That's the pattern with agents: once the prefix is cached, the answers cost more than the reading. Keep the start of your prompt identical from step to step, and put what changes at the end.

Change the numbers to your own loop:

The calculator uses live prices, so it moves when the price list does. Official figures above stay as of 29 October 2026.

Which setup fits your job#

The same loads on CheapTokens#

For the support bot and the agent loop, CheapTokens has DeepSeek-V4-Flash: id deepseek-ai/DeepSeek-V4-Flash-0731, 380K context, up to 32,768 output tokens per answer. It's one price at any hour, with no peak windows, one balance for all models, and an OpenAI- and Anthropic-compatible API at https://api.cheaptokens.dev/v1. In total, with input and output counted together, it is cheaper than OpenRouter.

One support-bot request (1,500 in, 400 out) costs $0.000063. The 100-step agent run above (3M in, 0.2M out) costs $0.065. The table below is the bot's day, 1,000 requests, across all our models with the OpenRouter column next to them; set it beside the official figures above:

To send the bot's request yourself (use the id exactly as written; short ids from price labels aren't accepted):

curl https://api.cheaptokens.dev/v1/chat/completions \
  -H "Authorization: Bearer $CHEAPTOKENS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "max_tokens": 400,
    "messages": [{"role": "user", "content": "Where is my order #1042?"}]
  }'

The response has a usage block with the real input and output token counts. Multiply them by the live prices, or paste them into the table above, and you have the cost of your own request instead of a guess.

For the connection details see the CheapTokens docs, the DeepSeek page and the live pricing table. If you're weighing DeepSeek against another family, MiniMax-M2.7 vs DeepSeek-V4-Flash puts them side by side on the same tasks.