DeepSeek API Pricing: Flash & V4 Pro Peak, Off-Peak & Cache Rates
Short answer: at peak rates deepseek-flash costs $0.30 per 1M input (cache miss) and $1.20 per 1M output, while deepseek-v4-pro costs $1.32 / $3.96. Every rate is exactly half outside peak hours, and a cache hit cuts the input line by 98.0% on flash. Both models take a 1,000,000-token context and can write 384,000 tokens in one reply.
The DeepSeek rate card, both ways
Official USD rates per 1,000,000 tokens, read from DeepSeek's Models & Pricing page on 2026-11-04. Each cell shows peak / off-peak.
| Model | Cache hit / 1M | Cache miss / 1M | Output / 1M | Context | Max output |
|---|---|---|---|---|---|
deepseek-flashLowest DeepSeek-V4.1-Flash-1017 · 2,500 concurrent | $0.006 / $0.003 | $0.30 / $0.15 | $1.20 / $0.60 | 1,000,000 | 384,000 |
deepseek-v4-pro DeepSeek-V4-Pro-0813 · 500 concurrent | $0.044 / $0.022 | $1.32 / $0.66 | $3.96 / $1.98 | 1,000,000 | 384,000 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours — including weekends and Chinese public holidays in full — are off-peak, and off-peak rates are exactly half of peak rates. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by V4.1-Flash-1017 and billed at the flash price. Prices are informational estimates in USD; caching tier, batch, tools and taxes follow the rules below. The DeepSeek invoice is the source of truth. Official source: DeepSeek Models & Pricing.
The clock decides half your bill.
DeepSeek is the only rate card on this site with a time-of-day price. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; every other hour is off-peak at exactly half the rate. There is no threshold, no minimum spend and no opt-in — the same 2,000-token prompt with a 500-token answer simply costs half as much at 12:00 UTC as it does at 07:00 UTC on a Tuesday.
On the reference workload — 2,000 input and 500 output tokens, 5,000 requests a day — the two models together bill $873.00 a month at peak rates and $436.50 a month if every request runs off-peak. That is a $436.50 (50.0%) saving for changing when you send the traffic, not what you send. Peak hours are only 7 of the 24 weekday hours, so most batch, ingestion and evaluation jobs already fall outside them by default. Model by model, deepseek-flash drops from $180.00 to $90.00 and deepseek-v4-pro from $693.00 to $346.50 — a $90.00 (50.0%) and $346.50 (50.0%) cut respectively.
A cache hit almost zeros the input line.
DeepSeek splits input two ways before the clock is even applied: tokens served from the prompt cache are billed at a separate cache-hit rate, and everything else at the cache-miss rate. On deepseek-flash a hit costs $0.006 per 1M input against $0.30 for a miss — 50.0x cheaper, a 98.0% cut. On deepseek-v4-pro it is $0.044 against $1.32 — 30.0x cheaper, a 96.7% cut. Output is never cached and is billed at the same rate either way.
If every input token hit cache at peak rates, the reference workload's two-model total falls from $873.00 to $402.00 a month — a $471.00 (54.0%) saving. Model by model: deepseek-flash goes $180.00 → $91.80 (49.0%) and deepseek-v4-pro goes $693.00 → $310.20 (55.2%). The gap is wider on V4 Pro despite its smaller 30.0x multiple, because output is a smaller share of its bill. A stable system prompt or a repeated retrieved prefix is what earns the hit; measure how much of your prompt is re-sent with the Conversation Cost Growth Calculator.
Price your workload on DeepSeek
Five inputs: token volumes, request volume, the share of input tokens that hit cache, and the share of requests that land in peak hours. Every figure is computed in your browser from the rate card above.
| Model | Per request | Per day | 30-day estimate |
|---|
Where DeepSeek sits beside the eight weekly-verified cards.
These comparisons are arithmetic between separate rate cards, not a merge into the eight-model panel. At peak rates deepseek-flash ($0.30/$1.20) has 1.5x the input rate of GPT-5.6 Luna ($0.20/$1.20) but the same output rate. At off-peak rates ($0.15/$0.60) it undercuts Luna by 25.0% on input and 50.0% on output — the cheapest input and output rate on this site at that hour.
deepseek-v4-pro at peak ($1.32/$3.96) is 34.0% cheaper on input and 60.4% cheaper on output than Claude Sonnet 5 ($2.00/$10.00), and off-peak halves again to $0.66/$1.98. Both DeepSeek models also accept a 1,000,000-token context with a 384,000-token maximum output — 3.0x the 128,000 cap on GPT-5.6 and Claude, and 5.86x the 65,536 cap on Gemini. That output cap, not the window, is what makes single-pass long generation possible here.
One more line on the invoice: the standalone Search and Fetch APIs are billed at $0.75 per 1,000 successful requests each, from 00:00 Beijing time on November 1, 2026. Server-side tool calls from DeepSeek Harness, Claude Code or Codex hit those APIs and were free before that date. No other card on this site charges by the request rather than by the token.
Where these numbers come from.
Every rate on this page was read directly from DeepSeek's official Models & Pricing documentation on 2026-11-04 and stored in a separate window.DEEPSEEK rate card. It is deliberately kept out of window.MODELS: the eight-model panel is defined as one flat input rate and one flat output rate per card, re-verified every week, and the pricing index is a fixed basket of exactly those eight cards. DeepSeek's cache-hit / cache-miss split and its peak / off-peak clock do not fit that shape, and adding a ninth card mid-series would break the index rather than extend it.
Off-peak rates are derived as exactly half of peak rates — that is the published rule, not an estimate — so only the peak rates are stored. The calculator multiplies your token volumes by the effective input rate (cache miss and cache hit weighted by your hit share, then scaled by your peak-hour share) and by the effective output rate; it assumes 30 days in a month and does not estimate tokenization. DeepSeek's official documentation and invoice are the source of truth; this page is an estimate for budgeting.
DeepSeek API pricing questions
How much does the DeepSeek API cost?
At peak rates, deepseek-flash costs $0.30 per 1M input tokens (cache miss) or $0.006 on a cache hit, and $1.20 per 1M output tokens; deepseek-v4-pro costs $1.32 per 1M input (cache miss) or $0.044 on a cache hit, and $3.96 per 1M output. Off-peak rates are exactly half: $0.15/$0.003/$0.60 for flash and $0.66/$0.022/$1.98 for V4 Pro.
What are DeepSeek's off-peak hours?
Off-peak rates are half of peak rates during every hour outside 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Weekends and Chinese public holidays are off-peak in full. Shifting the same workload out of peak hours cuts the bill exactly 50.0% — at 2,000 input and 500 output tokens on 5,000 requests a day, the two-model total drops from $873.00 to $436.50 a month.
How much does prompt caching save on DeepSeek?
A cache hit is 50.0x cheaper than a cache miss on deepseek-flash ($0.006 against $0.30 per 1M input) and 30.0x cheaper on deepseek-v4-pro ($0.044 against $1.32). If every input token hit cache at peak rates, the two-model total on the reference workload falls from $873.00 to $402.00 a month — a $471.00 (54.0%) saving. Caching only reduces input tokens; output is billed at the same rate.
What is DeepSeek's context window and maximum output?
Both DeepSeek models accept a 1,000,000-token context and allow up to 384,000 output tokens in one response. That maximum output is 3.0x the 128,000-token cap on GPT-5.6 and Claude models, and 5.86x the 65,536-token cap on Gemini — the largest of any rate card published on this site.
Why is DeepSeek not in the eight-model pricing comparison?
Because the eight-model panel in window.MODELS is defined as one flat input rate and one flat output rate per model, re-verified weekly, and the LLM Pricing Index is a fixed basket of exactly those eight cards. DeepSeek prices input three ways — cache hit and cache miss, each with a peak and an off-peak rate — so it is published as a separate rate card, verified against DeepSeek's official pricing page on 2026-11-04, and is not part of the weekly-verified panel or the index. See the LLM API Pricing Comparison for the eight-card panel.
How does DeepSeek pricing compare with OpenAI, Claude and Gemini?
At peak rates deepseek-flash ($0.30/$1.20) has 1.5x the input rate of GPT-5.6 Luna ($0.20/$1.20) but the same output rate; at off-peak rates ($0.15/$0.60) it undercuts Luna by 25.0% on input and 50.0% on output, making it the cheapest rate card on this site on both lines. deepseek-v4-pro at peak ($1.32/$3.96) is 34.0% cheaper on input and 60.4% cheaper on output than Claude Sonnet 5 ($2.00/$10.00). These are separate rate cards, not part of the weekly panel.
Does DeepSeek charge for anything other than tokens?
Yes. The standalone Search and Fetch APIs are billed per successful request at $0.75 per 1,000 requests each, starting 00:00 Beijing time on November 1, 2026. Server-side tool calls made through DeepSeek Harness, Claude Code or Codex hit those APIs; they were free before that date. The token calculator on this page prices tokens only — Search and Fetch are excluded.