CALL IT deepseek-flashREPLACES V4 FLASHNATIVE VISIONPEAK / OFF-PEAK BILLING1M CONTEXT
DeepSeek V4.1 Flash API Pricing
DeepSeek's new architecture line starts here, and it arrived cheaper than the model it replaces: $0.30/M input on a cache miss, $0.006/M on a hit and $1.20/M output at peak, exactly half of each off-peak. The API id is deepseek-flash, not deepseek-v4.1-flash — DeepSeek dropped the version from the name on September 10, 2026 and retired both V4 Flash and V4 Flash Vision Exp the same day, folding vision into this one model. Against the retired V4 Flash that is 32% off a cache miss, 57% off a cache hit and 9% off output, or 21% off our blended agentic rate. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays; weekends and Chinese public holidays bill off-peak in full. The weights are on Hugging Face under the MIT licence, like the rest of the V4 line.
Input - per 1M tokens
$0.30/M
Off-peak $0.15/M peak
Output - per 1M tokens
$1.20/M
Off-peak $0.60/M peak
Cached input
$0.006/M
Off-peak $0.003/M - 50x below a miss peak
Effective - agentic blend
$0.15/M
92/8 split - 82% cache
§ 01 / TERMINAL
Run the numbers.
Calculator pre-loaded with DeepSeek V4.1 Flash PEAK rates — halve every result for an off-peak run, which covers 17 of every 24 hours on a weekday and all of every weekend. Thinking is on by default and its tokens bill at the output rate, so a reasoning-heavy task costs more than its visible answer suggests.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Real-world presets.
AGENT LOOP
Repo-wide bug fix
$0.019/task
LONG DOC
Reading 100-page contracts
$0.020/doc
RAG SUPPORT
Support agent ticket triage
$0.00167/ticket
ASSISTANT TURN
Research planning turn
$0.00346/turn
Calibrated · measured on the vendor's tokenizer · 2026-06-10 Auto-counts as you type
Counts use a chars-per-token calibration measured on the vendor's own published tokenizer (deepseek-ai/DeepSeek-V3.2-Exp, 2026-06-10). English prose is typically within a few percent; code and non-Latin scripts tokenize heavier. For billing-exact counts use the vendor's count-tokens API.
Characters 485
Words 90
Tokens (estimated) 92 tokens
Cost as input · uncached $0.00003 USD
Cost as output · uncached $0.00011 USD
Cost as cached input $0.000001 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash Current | $0.30 cache $0.006 | $1.20 | $0.15 agentic 92/8 | 1M | Cheap 1M-context agent loops with images |
| GLM-5.3-Flash | $0.15 cache $0.03 | $0.50 | $0.0875 −42% | 1M | The cheaper multimodal flash tier |
| DeepSeek V4 Flash | $0.44 cache $0.014 | $1.32 | $0.189 deprecated · +26% | 1M | What this model replaced, for the record |
| DeepSeek V4 Pro | $1.32 cache $0.044 | $3.96 | $0.569 +3.8× | 1M | Harder reasoning, same 1M window |
| Gemini 3.8 Flash | $0.75 cache $0.075 | $3.75 | $0.481 +3.2× · promo | 1M | Google stack, multimodal input |
| GPT-5.4 mini | $0.75 cache $0.075 | $4.50 | $0.541 +3.6× | 400K | Teams already on OpenAI tooling |
Frequently asked.
What changed on September 10, what the id is called now, and what an image costs.
Q · 01 What is the model name — deepseek-flash or deepseek-v4.1-flash? +
deepseek-flash. DeepSeek dropped the version number from the API id with this release, and the pricing table's MODEL row reads deepseek-flash while the MODEL VERSION row underneath reads DeepSeek-V4.1-Flash. The older ids deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but DeepSeek calls that routing temporary — they now serve this model and bill at this model's price.Q · 02 How much does DeepSeek V4.1 Flash cost? +
At peak — 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday — it is $0.30/M on a cache miss, $0.006/M on a cache hit and $1.20/M output. Every other hour is off-peak at exactly half:
$0.15 / $0.003 / $0.60. Weekends and Chinese public holidays are off-peak in full. Under our 92/8 agentic blend with 82% cache hits the planning figure is $0.150/M at peak and $0.075/M off-peak.Q · 03 Is this cheaper than the V4 Flash it replaces? +
Yes, on every line. Cache miss fell from
$0.44 to $0.30 (-32%), cache hit from $0.014 to $0.006 (-57%) and output from $1.32 to $1.20 (-9%). On the blended agentic rate that is $0.189/M down to $0.150/M, about 21% less, and the model gained vision on the way. Price cuts with a capability gain are the normal direction of travel on this vendor — worth remembering when a page tells you to budget upward.Q · 04 What happened to V4 Flash and V4 Flash Vision Exp? +
Both were retired on September 10, 2026 in the same change-log entry that shipped this model. Their ids still answer, routed here and billed at the Flash price, but DeepSeek describes that as temporary compatibility rather than a supported path — so treat any code still naming them as work to do, not as working. We keep their pricing pages as a record of what they cost.
Q · 05 What does an image cost? +
Images are billed as ordinary input tokens after a resize. Anything under roughly 544x544 is scaled up; anything larger is scaled down to about a 1300x1300 pixel budget, which puts a ceiling of 1024 tokens on any single image. At peak that is $0.000307 per image and half that off-peak, whether the original was 2000x2000 or 5000x5000. You can send up to 600 images per request, and
detail: "low" downscales to 512x512 for less. The retired Vision Exp capped at 384 tokens per image, so the ceiling rose with this release while the rate fell.Q · 06 Is it the cheapest model of its kind? +
No. GLM-5.3-Flash is multimodal with the same 1M context at
$0.15/$0.50, which works out 42% below this model on the same blend — its cache hit is dearer ($0.03 against $0.006), so the gap narrows as your hit rate rises and closes entirely for cache-heavy agent loops. DeepSeek's argument is capability per dollar rather than the lowest sticker.Q · 07 Does thinking mode cost extra? +
No. One rate card covers both thinking and non-thinking modes, and thinking is the default. The mode is free; the tokens it writes are billed at the output rate like any others, so a reasoning-heavy request still costs more than a terse one.
Q · 08 Are the weights open, and what does self-hosting actually need? +
Yes — MIT licensed on Hugging Face, the same licence as the rest of the V4 line. The model card describes a 552B-parameter backbone that activates 8B parameters per token on prefill and 16B on decode, with a separate DeepSeek-ViT encoder on the image path, so the hardware you provision tracks the active count rather than the 552B headline. The card also credits KV-cache compression for a global cache roughly 4x smaller than V4 Flash's. That is the plainest available explanation for the thing that makes this release unusual: a larger model arriving at a LOWER price than the one it replaced, with no promotional framing and no end date to watch.
Q · 09 What are the limits? +
A 1M-token context with a 384K maximum output, and an account-level concurrency limit of 2500 — five times V4 Pro's 500. FIM completion is supported in non-thinking mode, which the retired Vision Exp never supported at all. Cache storage itself is free; you pay only the per-token rates above.