What DeepSeek V4.1 Flash Costs Per Million Tokens
DeepSeek V4.1 Flash costs $0.15 per million cache-miss input tokens off-peak and $0.30 at peak. A cached input token costs $0.003 off-peak, which is 50x below Flash's own cache-miss rate (derived).
Flash arrived around 10 September 2026 and appears in the official documentation as DeepSeek-Flash. The flagship DeepSeek-V4-Pro-0813 has been generally available since 13 August 2026.
Every figure below is USD per million tokens, taken from api-docs.deepseek.com and confirmed in September 2026.
| Model and window | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| DeepSeek-V4-Pro-0813, off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek-V4-Pro-0813, peak | $0.044 | $1.32 | $3.96 |
| DeepSeek-V4.1-Flash, off-peak | $0.003 | $0.15 | Not published |
| DeepSeek-V4.1-Flash, peak | $0.006 | $0.30 | Not published |
DeepSeek has not published an output price for V4.1 Flash that can be confirmed on the official page. If you see one quoted elsewhere, treat it as unverified until it appears in the official table.
DeepSeek is unusual in a second way that matters more to founders than the per-token number: it runs no startup credit program at all. That gap is covered below, and it is why AI Perks tracks the providers that do run one.
Why DeepSeek Charges Two Different Prices for the Same Token
DeepSeek is the only major model vendor that bills by time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and everything outside those windows costs exactly half.
Read that window list twice, because there is a gap in the middle of it. The stretch from 04:00 to 06:00 UTC sits between the two peak blocks and is billed at the off-peak rate. It is not one continuous nine-hour peak.
Derived from the published windows: peak covers seven hours a day, on weekdays only, which is 35 hours out of the 168 in a week. Roughly 79 percent of the calendar is already off-peak.
That reframes the problem. You are not trying to find the cheap window, you are trying to avoid the expensive one. Three workloads where that is nearly free to do:
- Nightly batch jobs. A job scheduled at 08:00 UTC sits inside peak. Moving it to 12:00 UTC halves the bill and changes nothing else.
- Evaluation and regression runs. These have no user waiting on them, so they belong outside the peak blocks by default.
- Embedding and re-indexing passes. Large one-off input burns are the single most peak-sensitive thing most teams run.
Interactive traffic is the part you cannot move, and it should be priced at peak rates in any forecast you build.
The Cache Hit Is the Bigger Multiplier
On DeepSeek-V4-Pro, a cached input token costs $0.022 per million against $0.66 for a miss. That is a 30x difference (derived from the table above), and it dwarfs the 2x you get from timing.
The two multipliers stack, and that stacking is the whole story of DeepSeek's cost profile. A peak cache-miss input token at $1.32 against an off-peak cache-hit token at $0.022 is a 60x spread on the same model (derived).
On V4.1 Flash the cache ratio is wider still: $0.15 for a miss against $0.003 for a hit is 50x (derived).
What this means in practice is that prompt structure beats model choice. Moving from V4-Pro to Flash cuts cache-miss input by 4.4x (derived, $0.66 against $0.15). Getting your cache to hit on V4-Pro cuts it by 30x. The cheaper model is the smaller lever.
Cache hits come from a stable prefix, so the practical rules are boring and effective: keep the system prompt byte-identical across calls, put tool definitions and retrieved context before anything user-specific, and never interpolate a timestamp or a session ID near the top of a prompt.
The Endpoint Names DeepSeek V4.1 Flash Retired
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Code that still sends either string will fail.
| Retired model name | Status |
|---|---|
deepseek-v4-flash | Retired with V4.1 Flash |
deepseek-v4-flash-vision-exp | Retired with V4.1 Flash |
This was a quiet change rather than a headline deprecation, which is exactly what makes it expensive. The model string is usually the least reviewed line in an AI integration: it sits in an environment variable, a config file, a Terraform template or a vendored SDK wrapper, and nobody touches it between releases.
The wider problem is the long tail of content. Tutorials, YouTube walkthroughs, forum answers and generated code samples written before September 2026 all still reference the old names, and none of them will be corrected. An assistant that learned from those pages will keep suggesting them.
If you run DeepSeek in production, grep your repository for both strings, then check your infrastructure config and any secrets manager separately. A model name set six months ago in a deployment variable will not show up in a code search.
What Changed Since the Earlier DeepSeek Pricing Guide
Three things have invalidated parts of every DeepSeek pricing page written before mid-August 2026: the peak and off-peak split, the V4-Pro and Flash split, and the retired endpoint names.
| Date | Release or change | Confidence |
|---|---|---|
| 24 April 2026 | V4-Pro and V4-Flash preview: 1M-token context, up to 384K output, switchable thinking and non-thinking modes | Medium |
| 31 July 2026 | DeepSeek-V4-Flash-0731, stable release of the Flash line | Medium |
| 13 August 2026 | DeepSeek-V4-Pro-0813 general availability | High |
| 16 August 2026 | Peak and off-peak pricing introduced | Medium |
| Around 10 September 2026 | DeepSeek-V4.1-Flash, old Flash endpoint names retired | High |
The historical figure worth knowing is the last known Flash pricing before the 16 August change: $0.14 per million cache-miss input and $0.28 per million output. That is the number most third-party comparison tables still show. It is not current pricing, and it predates the peak multiplier entirely, so any forecast built on it understates a peak-heavy workload by roughly 2x (derived: $0.14 against the current $0.30 peak cache-miss rate is 2.1x).
A single flat input price for DeepSeek is now meaningless without two qualifiers attached: which side of the cache, and which side of the clock.
DeepSeek Has No Free API Tier and No Startup Credit Program
The DeepSeek chat app is free to use. The API is not, and there is no stated free tier on the official pricing page. DeepSeek also runs no dedicated startup credit program.
Claims that new API accounts receive free credits circulate widely and are inconsistently reported. They are not confirmed on the official pricing page, so plan your first month of spend as if they do not exist.
Third-party hosts are a partial route around this, with the usual caveat that their prices and their terms are their own. Baseten lists DeepSeek V4.1 Flash input at $0.30 per million and carries a standing $30 new-workspace trial credit (verify both before you budget against them). OpenRouter, Together and Fireworks also host DeepSeek models, and the choice between them is usually about routing, region and rate limits rather than headline price.
The real gap for founders is structural. Most of the model and infrastructure vendors a startup runs on operate some form of credit program, and the current amounts and terms for each one are what AI Perks tracks. DeepSeek operates none, which means DeepSeek is the one line of your AI bill you always pay in cash.
That is the argument for treating it as the cheap default tier inside a stack rather than the whole stack. Route the high-volume, cache-friendly, off-peak work to DeepSeek, and route the work that needs frontier reasoning to a provider whose bill somebody else is covering. AI Perks tracks $7.7M in credits across 194 companies, which is where that second half comes from.
See which providers run credit programs at getaiperks.com →
Frequently Asked Questions
How much does DeepSeek V4.1 Flash cost?
DeepSeek V4.1 Flash costs $0.15 per million cache-miss input tokens off-peak and $0.30 at peak. Cached input is $0.003 off-peak and $0.006 at peak. DeepSeek has not published a confirmable output price for Flash, so treat any output figure you find elsewhere as unverified.
What are DeepSeek's peak and off-peak hours?
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else, including the 04:00 to 06:00 UTC gap between the two blocks and all weekend traffic, is off-peak at half price. Derived from those windows, about 79 percent of the week is already off-peak.
Why does my deepseek-v4-flash API call suddenly fail?
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Any code, tutorial or config still sending those strings will fail. Check your environment variables and deployment config as well as your source, since model names often live outside the repository. The change shipped quietly around 10 September 2026.
Does DeepSeek give free API credits to new accounts?
There is no stated API free tier on DeepSeek's official pricing page. Reports of new-account credits are inconsistent and unconfirmed, so budget as if none exist. The chat app is free. For providers that do publish credit programs, see AI Perks.
Does DeepSeek have a startup credit program?
No. DeepSeek runs no dedicated startup credit program, which makes it one of the few major model APIs you always pay for in cash. Most other providers do run one, and AI Perks tracks $7.7M in credits across 194 companies covering them.
Is DeepSeek V4-Pro or V4.1 Flash cheaper?
Flash is cheaper on every published comparable: $0.15 against $0.66 per million cache-miss input tokens off-peak, a 4.4x gap (derived). But fixing your prompt caching on V4-Pro is a 30x lever, so prompt structure matters more than model choice. Neither has a credit program behind it, which is the part AI Perks tracks for the providers that do.
DeepSeek is the cheapest line on your AI bill. Get the other lines covered at getaiperks.com.