DeepSeek V4 pricing is not one number. Flash and Pro have different rates, cached input costs far less than fresh input, and every rate halves off-peak. On top of that, Search and Fetch calls now have their own per-request charge.
A lot changed since this post first went up. The default Flash model is now called deepseek-flash, and it runs DeepSeek-V4.1-Flash-1017. The experimental vision model is retired. Flash prices dropped, and the Pro cache-hit rate changed.
DeepSeek still charges by usage. There’s no flat coding plan like GLM’s or Alibaba’s, and that suits light or bursty use, because you pay only for the traffic you generate.
The model tiers
DeepSeek’s pricing page lists two models right now. All prices are in US dollars per million tokens.
| deepseek-flash | deepseek-v4-pro | |
|---|---|---|
| Model version | DeepSeek-V4.1-Flash-1017 | DeepSeek-V4-Pro-0813 |
| Input, cache hit (peak / off-peak) | $0.006 / $0.003 | $0.044 / $0.022 |
| Input, cache miss (peak / off-peak) | $0.30 / $0.15 | $1.32 / $0.66 |
| Output (peak / off-peak) | $1.20 / $0.60 | $3.96 / $1.98 |
| Vision | Supported | Not supported |
| Concurrency limit | 2500 | 500 |
Both models support JSON output, tool calls, the Responses API and the Anthropic API. Chat Prefix Completion and FIM Completion are in beta. FIM works in non-thinking mode only. Thinking mode is the default, and the page links a Thinking Mode guide for switching.
Pro’s peak cache-miss input and output prices are the same as in our August snapshot. Its cache-hit rate is now $0.044 per million at peak. Flash is cheaper across the board than the old deepseek-v4-flash rates.
Peak and off-peak hours
DeepSeek introduced peak and off-peak pricing on August 16, 2026. The pricing page now gives the exact schedule:
- Peak: 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays.
- Off-peak: all other hours, including weekends and Chinese public holidays in full.
Off-peak rates are half the peak rates, and that goes for both input and output.
Peak windows add up to seven hours on a weekday. Do you run batch jobs, such as nightly code review, bulk refactors or test generation? Schedule them outside those windows and you pay half. In Kuwait time (UTC+3), peak is 04:00 to 07:00 and 09:00 to 13:00 on weekdays.
A common mistake is assuming off-peak means “night.” It doesn’t. It means outside the UTC windows above, so a 14:00 UTC job on a Tuesday is already off-peak, while 08:00 UTC is peak.
If you schedule with cron on a server set to UTC, a simple pattern is to run heavy jobs at a fixed off-peak hour. Check your server’s time zone first, because a cron entry on a local-time server will shift relative to UTC.
Cache-hit discounts
Input tokens that hit DeepSeek’s cache are billed far cheaper than fresh input. When an agent resends the same system prompt, file contents or conversation history, those repeated tokens can be billed at the cache-hit rate.
That matters for coding agents like Claude Code and OpenCode. Both of them send the same context on almost every turn. At peak, deepseek-flash cache-hit input is $0.006 per million tokens against $0.30 for a cache miss, so a cached token costs 2% of a fresh one. On Pro the ratio is $0.044 against $1.32.
Two habits help your hit rate:
- Keep the stable part of the prompt, such as system instructions and project rules, at the start, and keep it the same between calls.
- Avoid tools or wrappers that inject timestamps or random IDs at the top of every request, since that changes the prefix.
Base URLs and model names
Use the base URL that matches your tool:
# OpenAI-compatible tools such as Cline, OpenCode, Aider, Codex custom providers
https://api.deepseek.com
# Anthropic-compatible tools such as Claude Code
https://api.deepseek.com/anthropic
Model names:
- Use
deepseek-flashas the default model. DeepSeek’s pricing page says to use this name. - Use
deepseek-v4-profor harder work where the extra cost is justified. deepseek-v4-flashanddeepseek-v4-flash-vision-expare legacy names. They’re still accepted, but the models behind them are retired. Requests are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price.
For image input, you no longer need a separate model. deepseek-flash supports vision. Pro doesn’t, so don’t route image tasks to deepseek-v4-pro.
The legacy names keep old configurations working. That’s convenient, but it also hides the fact that the model changed underneath you. Update your configs to deepseek-flash so logs and cost reports show what you’re actually running.
Context, output limits and concurrency
The pricing page lists a 1M-token context length and a maximum output of 384K tokens. Concurrency limits differ by model: 2500 for deepseek-flash and 500 for deepseek-v4-pro. DeepSeek points to its Rate Limit & Isolation page for details.
For agent setups, plan around the concurrency gap. If you fan out many parallel subagents, Pro will hit its limit five times sooner than Flash. Keep the parallel workers on Flash, and send only the hard steps to Pro.
Search and Fetch pricing
Search and Fetch are standalone DeepSeek APIs. Each is billed per request:
| API | Unit | Price |
|---|---|---|
| Search | Per 1,000 successful requests | $0.75 |
| Fetch | Per 1,000 successful requests | $0.75 |
This hits Claude Code users directly. DeepSeek says that when you use its API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit the Search and Fetch APIs. Those calls were free. Per-request billing starts at 00:00 Beijing time on November 1, 2026, which is 16:00 UTC on October 31.
Only successful requests are charged. If your agent does a lot of web lookups, this is a new line on the bill that token estimates won’t show. Check your usage page after a few days of normal work to see how many requests your sessions actually make.
How to estimate your cost
DeepSeek’s deduction rule is simple. Model cost is tokens times price, and Search and Fetch are charged per successful request. Fees come out of your granted balance first, then your topped-up balance.
To estimate:
- Estimate tokens per task, split into cache-hit input, cache-miss input and output.
- Multiply each by the rate for your model and time window.
- Add Search and Fetch requests at $0.75 per 1,000.
- Multiply by tasks per day and working days per month.
Here’s a worked example with illustrative numbers, not measurements. Say one agent session uses 2M cache-hit input tokens, 300K cache-miss input tokens and 100K output tokens.
Example session cost (illustrative token counts)
| deepseek-flash, peak | 2 × $0.006 + 0.3 × $0.30 + 0.1 × $1.20 = $0.222 |
|---|---|
| deepseek-flash, off-peak | $0.111 |
| deepseek-v4-pro, peak | 2 × $0.044 + 0.3 × $1.32 + 0.1 × $3.96 = $0.880 |
| deepseek-v4-pro, off-peak | $0.440 |
If the same session also makes 200 Search or Fetch requests, add 200 ÷ 1,000 × $0.75 = $0.15. On Flash at peak, that’s a big share of the session’s total, so web-heavy agents should watch it.
Two takeaways. Output and cache-miss input drive the bill, not cached context. And moving the same work off-peak halves the token cost without changing anything else.
Pro vs Flash: which to use
Use deepseek-flash for routine coding, agent loops, parallel subagents and anything with images. Move to deepseek-v4-pro for hard reasoning, debugging or architecture work, where a better answer saves you a retry.
Default to the cheap tier, and escalate only when a task needs it. Pro costs about four times Flash for cache-miss input at peak, and over three times for output. Routing everything to Pro by habit is the fastest way to inflate a bill.
Before you budget
- Check current rates on the official pricing page
- Switch configs from deepseek-v4-flash to deepseek-flash
- Remove deepseek-v4-flash-vision-exp and send image tasks to deepseek-flash
- Use deepseek-v4-pro only for hard tasks, and never for image input
- Schedule batch jobs outside the 01:00-04:00 and 06:00-10:00 UTC weekday peak windows
- Account for cache-hit discounts on repeated context
- Budget for Search and Fetch requests from Claude Code, Codex or DeepSeek Harness
- Use the OpenAI base URL for OpenAI-compatible tools and /anthropic for Claude Code
What to use
Start with deepseek-flash. It’s cheaper than the old Flash, handles images, and has the higher concurrency limit. Keep deepseek-v4-pro for the hard cases, and move batch work off-peak to halve its cost. If you use Claude Code or Codex with DeepSeek, add Search and Fetch requests to your estimate. And check the official pricing page before you rely on these figures in a budget.
To put it to work, see run DeepSeek with Claude Code. To compare against flat-rate options, see coding plans vs pay-per-token or our roundup of the cheapest AI coding APIs.