DeepSeek's official API charges $0.15 per million input tokens and $0.60 per million output tokens for V4.1 Flash during off-peak hours, and exactly double that at peak. V4 Pro runs $0.66 input and $1.98 output off-peak, $1.32 and $3.96 at peak. DeepSeek had planned to route all V4 Pro traffic to V4.1 Flash from September 14, then reversed that on user demand, so V4 Pro stays on the official API at the same billing. V4.1 Flash launched on September 10 and retired V4 Flash, less than a month after the August 16 repricing that raised output prices by up to 371 percent with three days' notice.
The official page is no longer the whole story. Seventeen third-party hosts serve the V4 Flash weights DeepSeek just retired, and most of them list below what DeepSeek charged for that model. On V4.1 Flash the launch-day picture is the reverse: the official API is the cheapest of the four hosts serving it. This page covers the official rates for every current DeepSeek model, the August increase, every host's price, and the workload math that decides where a DeepSeek bill should actually be paid. The new model has its own page at DeepSeek V4.1 Flash API pricing.
DeepSeek API at a glance
Base URL: https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format). Two live models: deepseek-flash (V4.1 Flash) at $0.15 in, $0.60 out per million tokens off-peak, and deepseek-v4-pro at $0.66 in, $1.98 out off-peak. Peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday) are double. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but are served by V4.1 Flash at Flash rates. Full rate table and 17 third-party hosts below.
DeepSeek official API pricing
Two models are live on the official API as of September 10, 2026, and one of them is on its way out. Both have a 1M-token context window and a 384K max output. Prices are USD per million tokens.
| Model | Input, cache miss (off-peak / peak) | Input, cache hit (off-peak / peak) | Output (off-peak / peak) |
|---|---|---|---|
| deepseek-flash (V4.1 Flash) | $0.15 / $0.30 | $0.003 / $0.006 | $0.60 / $1.20 |
| deepseek-v4-pro | $0.66 / $1.32 | $0.022 / $0.044 | $1.98 / $3.96 |
| deepseek-v4-flash (legacy name, routes to V4.1 Flash) | $0.15 / $0.30 | $0.003 / $0.006 | $0.60 / $1.20 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Everything else, including weekends, bills at the off-peak rate. That is 35 peak hours out of 168 in a week, roughly 21 percent of the clock. In US terms, peak runs 9 p.m. to midnight and 2 a.m. to 6 a.m. Eastern, which is why the schedule hits Asian and European business hours far harder than American ones. The full time-zone breakdown is in DeepSeek peak and off-peak hours.
V4.1 Flash has native image input; the separate V4-Flash-Vision-Exp model is retired and its name routes to V4.1 Flash. Cache hits apply when a request repeats a prefix the API has already processed, which is common in agent loops and long system prompts. DeepSeek planned to retire the V4 Pro row on September 14, 2026 and route its requests to V4.1 Flash, then kept V4 Pro live at the same billing after user demand. Both rows are current.
What changed on August 16
Before August 16, DeepSeek charged one flat rate around the clock. V4 Pro listed at $0.435 input and $0.87 output, a price it had held since May 23. V4 Flash listed at $0.14 input and $0.28 output.
| Token type | Old flat rate | New off-peak | New peak | Change at peak |
|---|---|---|---|---|
| V4 Flash input, cache miss | $0.14 | $0.22 | $0.44 | +214% |
| V4 Flash output | $0.28 | $0.66 | $1.32 | +371% |
| V4 Flash input, cache hit | $0.0028 | $0.007 | $0.014 | +400% |
| V4 Pro input, cache miss | $0.435 | $0.66 | $1.32 | +203% |
| V4 Pro output | $0.87 | $1.98 | $3.96 | +355% |
| V4 Pro input, cache hit | $0.003625 | $0.022 | $0.044 | +1,114% |
Even the off-peak rates, the cheapest DeepSeek now offers, are 52 to 136 percent above the old flat price on regular tokens and 2.5 to 6 times higher on cached tokens.
The timeline matters as much as the numbers. DeepSeek posted a notice on August 6 saying a significant increase was coming, without figures. The rates were published on August 13 alongside the V4 Pro general-availability release. They took effect on August 16 at 16:00 UTC. A buyer who had budgeted on the August 12 price list had three days to react once the numbers were public, and no instrument to hedge with. DeepSeek's stated reason was to "allocate resources more reasonably" and push developer workloads toward less congested hours. The mechanics of that repricing, and what it cost a typical buyer, are worked through in How to lock in token prices. Those V4 Flash rates lasted 25 days. On September 10 DeepSeek retired V4 Flash and launched V4.1 Flash on a new rate card, covered in DeepSeek V4.1 Flash API pricing.
DeepSeek V4 Flash pricing by host
V4 Flash was retired from the official API on September 10, 2026, but its weights are open and third-party hosts still serve them. Prices below are per million tokens as listed on OpenRouter on September 2, 2026, sorted cheapest first. Two hosts were running promotional discounts on that date, marked with an asterisk. The official DeepSeek API is included at both rates for comparison.
| Host | Input | Output | Cache read |
|---|---|---|---|
| DigitalOcean | $0.0679 | $0.168 | $0.0168 |
| StreamLake* | $0.0886 | $0.177 | $0.0177 |
| DeepInfra | $0.09 | $0.18 | $0.018 |
| GMI Cloud* | $0.112 | $0.224 | $0.0224 |
| SiliconFlow | $0.13 | $0.28 | $0.028 |
| Alibaba Cloud | $0.134 | $0.268 | $0.0268 |
| Venice | $0.138 | $0.275 | $0.028 |
| Novita | $0.14 | $0.28 | $0.028 |
| NextBit | $0.14 | $0.28 | $0.028 |
| AtlasCloud | $0.14 | $0.28 | $0.028 |
| Baidu Qianfan | $0.14 | $0.28 | $0.028 |
| CoreWeave | $0.14 | $0.28 | $0.07 |
| Parasail | $0.14 | $0.28 | $0.07 |
| Mancer | $0.185 | $0.50 | n/a |
| Phala | $0.20 | $0.40 | $0.07 |
| Azure (US) | $0.21 | $0.56 | $0.031 |
| DeepSeek official, off-peak (until Sept 10) | $0.22 | $0.66 | $0.007 |
| Cloudflare | $0.44 | $1.32 | $0.014 |
| DeepSeek official, peak (until Sept 10) | $0.44 | $1.32 | $0.014 |
The pattern is hard to miss. A block of six hosts sits at exactly $0.14 and $0.28, which is DeepSeek's own pre-August price list. The third-party market anchored on the old official rate and stayed there when DeepSeek moved. Only Cloudflare matched the new peak rate. The result was that the official API became the second most expensive place to buy V4 Flash output at any hour of the day, right up until DeepSeek stopped selling the model. The weights are still served; DeepSeek left the market for its own retired product.
The spread from cheapest to most expensive host is 6.5x on input and 7.9x on output, for the same weights.
DeepSeek V4 Pro pricing by host
Same source and date. The Pro spread is narrower because Pro is more expensive to serve, but the official API's position has still changed.
| Host | Input | Output | Cache read |
|---|---|---|---|
| DeepSeek official, off-peak | $0.66 | $1.98 | $0.022 |
| DigitalOcean | $0.87 | $1.74 | $0.174 |
| StreamLake | $1.027 | $2.055 | $0.086 |
| Baidu Qianfan | $1.029 | $2.058 | $0.085 |
| GMI Cloud | $1.044 | $2.088 | $0.087 |
| Ionstream | $1.131 | $2.262 | $0.094 |
| CoreWeave | $1.15 | $2.55 | $0.20 |
| DeepInfra | $1.30 | $2.60 | $0.10 |
| DeepSeek official, peak | $1.32 | $3.96 | $0.044 |
| Alibaba Cloud | $1.416 | $2.832 | $0.118 |
| SiliconFlow | $1.502 | $3.135 | $0.135 |
| Novita | $1.60 | $3.20 | $0.135 |
| Venice | $1.65 | $3.301 | $0.33 |
| AtlasCloud | $1.68 | $3.38 | $0.13 |
| NextBit | $1.72 | $3.45 | $0.14 |
| Baseten | $1.74 | $3.48 | $0.145 |
| Parasail | $1.74 | $3.48 | $0.10 |
| Azure (US) | $1.91 | $3.83 | $0.16 |
Off-peak, DeepSeek is still the cheapest V4 Pro host on input and cache reads, and only DigitalOcean beats it on output. At peak, DeepSeek drops to the middle of the table: seven hosts undercut it on input and all sixteen undercut it on output. Third-party output prices run $1.74 to $3.83, and the official peak rate of $3.96 is the highest number in the column.
The official API is no longer the default cheapest host
The tables above are list prices. What a buyer actually pays depends on the token mix and the clock. The workload below is 10 million input tokens and 1 million output tokens per day, uncached, which is a typical shape for a retrieval-heavy production app.
| Route | V4 Flash, per day | V4 Flash, per year | V4 Pro, per day | V4 Pro, per year |
|---|---|---|---|---|
| Official, old flat rate (pre-Aug 16) | $1.68 | $613 | $5.22 | $1,905 |
| Official, all off-peak | $2.86 | $1,044 | $8.58 | $3,132 |
| Official, all peak | $5.72 | $2,088 | $17.16 | $6,263 |
| Official, uniform traffic (21% peak) | $3.46 | $1,261 | $10.37 | $3,784 |
| Cheapest third party (DigitalOcean) | $0.85 | $309 | $10.44 | $3,811 |
| Most expensive third party (Azure) | $2.66 | $971 | $22.93 | $8,369 |
Three conclusions fall out of that table.
For V4 Flash, the official API lost on price. A workload that runs around the clock pays $3.46 a day on DeepSeek and $0.85 a day on the cheapest host. That is a 4x difference on the same model, and it holds even against DeepSeek's best-case off-peak rate (3.4x). The old official price, $1.68 a day, is now available from six different hosts, so a buyer who liked the pre-August deal can still have it, just not from DeepSeek.
For V4 Pro, it is a coin flip that depends on your schedule. DeepSeek off-peak is the cheapest route at $8.58 a day. Uniform traffic costs $10.37 on DeepSeek and $10.44 on DigitalOcean, a rounding error apart. A workload concentrated in peak hours should leave the official API entirely: $17.16 a day versus $10.44.
The cheapest host is not automatically the right host. DigitalOcean's V4 Pro endpoint showed 33 tokens per second and 92 percent uptime on the day these prices were pulled, against 69 tokens per second and 99.6 percent for CoreWeave at $1.15 and $2.55. Promotional discounts (StreamLake, GMI Cloud) can end. Some hosts serve quantized weights. Price is one input to the routing decision, and the reasons it varies so much between hosts are covered in Why token prices differ.
Which DeepSeek model and host to use
Use V4.1 Flash for retrieval, extraction, classification and routine agent steps. Keep V4 Pro for the work that needs it; DeepSeek kept it live after user pushback on the planned September 14 routing. On launch day the official API is the cheapest host for it. The retired V4 Flash weights from a third-party host cost less still for tasks that were fine on the old model.
Use V4 Pro where the quality gap matters; it is still served and still billed at Pro rates. The official API still serves it at $0.66 input off-peak; third-party hosts list it from $0.58 to $1.91 input. Test V4.1 Flash on the workload before then. DeepSeek says it beats Pro on performance, speed, and cost, and it bills at a quarter of Pro's input rate.
Run the blended number, not the list price. Input share, output share, cache rate, and hour of day each move the answer. A model-by-model view of the same calculation across DeepSeek, GLM, Qwen, and Kimi is in the LLM API pricing comparison.
Older DeepSeek models: V3.2 and the retired names
DeepSeek's official API no longer lists V3.2. When V4 launched on April 24, 2026, the legacy model names deepseek-chat and deepseek-reasoner were pointed at V4 Flash, and both names were retired on July 24. The V3.2 rates of $0.28 input and $0.42 output are historical for the official API. Third-party hosts still serve V3.2 weights, and for buyers with V3.2 pinned in production that market is now the only one. The provider table is in DeepSeek V3.2 API pricing.
Model-specific pages, kept current as prices move:
- DeepSeek V4.1 Flash API pricing
- DeepSeek V4 API pricing
- DeepSeek V4 Pro API pricing
- DeepSeek peak and off-peak hours
- DeepSeek V3.2 API pricing
Where DeepSeek pricing goes from here
The August 16 increase is the clearest example yet of how posted-price inference works. One seller changed its number, gave three days' notice on the specifics, and every buyer on the official API absorbed it. There was no market price to check the move against and no contract to hold the old price.
What happened next is the more interesting part. Seventeen other hosts serve the same weights, most of them held the old price, and the official API went from the cheapest V4 Flash host to one of the most expensive within a week. That is a competitive supply side doing its job. What it still lacks is a single price that reflects all of it: a number a buyer can quote, budget against, and lock. The Mercatus Token Index tracks where V4 Flash and V4 Pro actually trade across hosts, and How a token exchange works explains how that number gets set by trades rather than by a pricing page. For the broader picture on why Chinese open-weight models are the ones being repriced, see US vs China token prices.
Frequently asked questions
How much does the DeepSeek API cost?
As of September 10, 2026, the official DeepSeek API charges $0.15 to $0.30 per million input tokens and $0.60 to $1.20 per million output tokens for V4.1 Flash. V4 Pro is $0.66 to $1.32 input and $1.98 to $3.96 output. DeepSeek kept it live past the September 14 date it had set for routing it to V4.1 Flash. The lower figure applies off-peak, the higher at peak. Third-party hosts list the retired V4 Flash weights from $0.05 input and $0.16 output, and V4 Pro from $0.58 input and $1.74 output.
Did DeepSeek raise its API prices?
Yes. On August 16, 2026 DeepSeek replaced its flat rates with peak and off-peak pricing. Output tokens rose 128 to 136 percent off-peak and 355 to 371 percent at peak. Cached input tokens rose up to 12x at peak on V4 Pro. The figures were published on August 13, three days before they took effect.
What are DeepSeek's peak hours?
01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. All other hours, including weekends, are off-peak and bill at half the peak rate.
Is the official DeepSeek API the cheapest way to run V4?
For V4.1 Flash, yes on launch day: the three third-party hosts listed at 2x to 2.5x the official off-peak rate. For the retired V4 Flash weights, no: sixteen hosts list below DeepSeek's last official rate. For V4 Pro, the official API is cheapest off-peak on input, and it is still serving Pro.
Is DeepSeek V3.2 still available on the official API?
No. The deepseek-chat and deepseek-reasoner names were retired on July 24, 2026, and the official pricing page lists only V4 models. V3.2 is available through third-party hosts.
Why is the same DeepSeek model priced so differently across hosts?
Hosts run different hardware, batch sizes, quantization, and margins, and they set prices independently. DeepSeek's August increase widened the spread because most hosts did not follow it. The full breakdown is in Why token prices differ.
Methodology
Official DeepSeek rates, peak-hour definitions, and model specifications are from the DeepSeek API documentation pricing page and change log as of September 2, 2026. V4.1 Flash rates, the V4 Flash retirement, and the V4 Pro routing rule are from the same pages on September 10, 2026. Pre-August 16 rates are from DeepSeek's prior pricing page and contemporaneous coverage of the August 13 announcement. Third-party host prices, throughput, and uptime figures are from OpenRouter's model pages for DeepSeek V4 Flash and V4 Pro on September 2, 2026, and reflect list prices on that date including any promotional discounts then in effect. Workload calculations assume uncached input; the uniform-traffic case weights peak hours at 35 of 168 weekly hours. Prices change; the model-specific pages linked above are updated as they do.