DeepSeek V4 Pro and V4.1 Flash explained: peak vs off-peak pricing

Updated November 2, 2026. Researched and drafted with AI help; every figure is checked against the linked source.

DeepSeek sells two models through its API: DeepSeek V4 Pro and DeepSeek V4.1 Flash. Both are cheap by list price, but DeepSeek is also the only provider in our price tables that charges by the clock: the same request costs twice as much during weekday peak hours as it does at night or on the weekend. This guide covers what the two models are, the exact peak hours, and what that does to real costs.

The short version

V4 Pro (deepseek-v4-pro): $1.32 input / $3.96 output per million tokens at peak, $0.66 / $1.98 off-peak. Cache hits: $0.044 peak, $0.022 off-peak.
V4.1 Flash (deepseek-flash): $0.30 / $1.20 at peak, $0.15 / $0.60 off-peak. Cache hits: $0.006 peak, $0.003 off-peak.
Peak hours: 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour is off-peak, including weekends and Chinese public holidays in full. Off-peak prices are half the peak prices.
Context and output: DeepSeek’s pricing page lists a 1M-token context and up to 384K output tokens.
Vision: V4.1 Flash accepts images. V4 Pro does not.

The two models

V4 Pro is the larger model. DeepSeek’s changelog says V4 Pro reached general availability on August 13, 2026 (the pricing page calls this version DeepSeek-V4-Pro-0813), with three thinking-effort levels (low, high and max), and that peak/off-peak pricing took effect on August 16. The pricing page lists a concurrency limit of 500 for it.

V4.1 Flash arrived on September 10, 2026. DeepSeek calls it the smallest model in a new architecture family, with native visual understanding, and says multiple parties’ tests put it ahead of V4 Pro on “performance, cost, speed & total runtime”. That is DeepSeek’s claim; the release note does not name the testers or give the numbers in text. Its concurrency limit is 2,500.

Both models support thinking and non-thinking modes (thinking is the default), JSON output, tool calls, and both an OpenAI-compatible and an Anthropic-compatible API. Fill-in-the-middle completion (beta) works in non-thinking mode only.

Model names after the September change

The V4.1 Flash release retired V4 Flash and V4 Flash Vision Exp. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work for now (DeepSeek calls the routing temporary), but DeepSeek’s pricing page says those requests are served by V4.1 Flash and billed at the Flash price. Use deepseek-flash in new code.

The September 10 release note also said deepseek-v4-pro would route to V4.1 Flash from September 14 until a V4.1 Pro launches. DeepSeek’s changelog says it then decided to keep serving V4 Pro after September 14, and the current pricing page maps deepseek-v4-pro to DeepSeek-V4-Pro-0813 at the Pro prices above. If you saw the earlier note, check which model your requests are actually billed as.

When the cheap hours are

Peak time is seven hours per weekday: 01:00–04:00 and 06:00–10:00 UTC. In local time that is:

Time zone	Peak hours (Mon–Fri UTC)
UTC	01:00–04:00 and 06:00–10:00
Beijing (UTC+8)	09:00–12:00 and 14:00–18:00
New York in summer time (UTC−4)	21:00–00:00 (Sun–Thu evenings) and 02:00–06:00

So peak is the Chinese working day. For users in the Americas it falls mostly overnight, which means a business running during US office hours already pays off-peak prices on weekdays.

Seven hours a day over five days is 35 of the week’s 168 hours, about 21%. If your traffic were spread evenly around the clock, about 79% of it would land off-peak, and the blended price would be roughly $0.80 input / $2.39 output for V4 Pro and $0.18 / $0.73 for V4.1 Flash. Chinese public holidays push that share a little further toward off-peak.

What it costs on real workloads

We use the same workloads as our cost per 1,000 requests guide, plus one output-heavy job. All figures are for 1,000 requests at official list prices, standard tier.

Chat: 2,000 input + 500 output tokens.
Agent step: 2,000 fresh input + 18,000 cached input + 1,000 output tokens.
Summary: 10,000 input + 300 output tokens.
Long writing: 2,000 input + 8,000 output tokens.

Model	Price in / out	Cache read	1,000 chats	1,000 agent steps	1,000 summaries	1,000 long writes
GPT-6 Luna	$0.10 / $0.50	$0.01	$0.45	$0.88	$1.15	$4.20
Claude Haiku 5.5*	$0.10 / $0.50	$0.01	$0.45	$0.88	$1.15	$4.20
DeepSeek V4.1 Flash, off-peak	$0.15 / $0.60	$0.003	$0.60	$0.95	$1.68	$5.10
DeepSeek V4.1 Flash, peak	$0.30 / $1.20	$0.006	$1.20	$1.91	$3.36	$10.20
DeepSeek V4 Pro, off-peak	$0.66 / $1.98	$0.022	$2.31	$3.70	$7.19	$17.16
Mistral Large 4 (preview)†	$0.70 / $2.10	$0.07	$2.45	$4.76	$7.63	$18.20
Gemini 3.8 Flash‡	$0.75 / $3.75	$0.075	$3.38	$6.60	$8.63	$31.50
DeepSeek V4 Pro, peak	$1.32 / $3.96	$0.044	$4.62	$7.39	$14.39	$34.32
Grok 4.7	$2 / $6	$0.50	$7.00	$19.00	$21.80	$52.00
Claude Sonnet 5.5	$2 / $10	$0.10	$9.00	$15.80	$23.00	$84.00
GPT-6.1 Sol	$2 / $10	$0.10	$9.00	$15.80	$23.00	$84.00

* Claude Haiku 5.5 prices apply to prompts up to 100K tokens; all workloads here are under that.

† Mistral Large 4 is in public preview. Mistral cut its price from $1.36 / $4.18 ($0.14 cache reads) to $0.70 / $2.10 ($0.07) when it released the open weights in late October 2026.

‡ Google lists this Gemini 3.8 Flash price through December 31, 2026. From January 1, 2027 it lists $1.50 / $7.50 with $0.15 cache reads.

What the numbers say

Off-peak halves everything, including the cache. DeepSeek applies the 50% to cache hits too, so the off-peak column is exactly half the peak column for every workload. For a pipeline running one million agent steps, that is $7,392 on V4 Pro at peak against $3,696 off-peak, and $1,908 against $954 on V4.1 Flash.

V4.1 Flash is cheap, but not the cheapest small model any more. Even off-peak, V4.1 Flash costs a little more than GPT-6 Luna and Claude Haiku 5.5 on every workload here ($0.60 vs $0.45 per 1,000 chats, $0.95 vs $0.88 per 1,000 agent steps). Its advantages are elsewhere: a 1M context with a flat price and image input in the same model.

Mistral Large 4 now sits between V4 Pro’s off-peak and peak prices. After Mistral cut Large 4 to $0.70 / $2.10 in late October, V4 Pro off-peak is only slightly cheaper ($2.31 vs $2.45 per 1,000 chats), with a wider gap on cache-heavy agent steps ($3.70 vs $4.76) because its cache read is $0.022 against $0.07. At peak, V4 Pro costs more than Mistral Large 4 on every workload here ($4.62 vs $2.45 per 1,000 chats, $34.32 vs $18.20 per 1,000 long writes). At peak, V4 Pro also lands near Gemini 3.8 Flash on most workloads (summaries are the exception: $14.39 vs $8.63).

Against $2-input models, V4 Pro is far cheaper at any hour. At peak, V4 Pro costs about half as much as Claude Sonnet 5.5 or GPT-6.1 Sol per agent step ($7.39 vs $15.80) and about 40% as much on long writing ($34.32 vs $84.00). Off-peak it is roughly a quarter to a third. Whether it is good enough for your task is a separate question that only your own tests can answer.

How to use the off-peak window

Schedule batch-like jobs. DeepSeek’s pricing page does not list a separate Batch API discount, so off-peak hours are its version of one: no waiting for a batch queue, just a clock. Nightly evaluations, document backfills and embedding-style preprocessing are natural fits, as long as they finish outside 01:00–04:00 and 06:00–10:00 UTC on weekdays.
Check your own time zone. If your users are in Europe, the 06:00–10:00 UTC block overlaps the European morning, so interactive traffic there will partly pay peak rates.
Budget for the worst case. DeepSeek’s pricing page says it “reserves the right to adjust” prices, and time-of-day pricing itself only started in August 2026. Plan with the peak price and treat off-peak as a saving.

Where to go next

Put your own token counts into the cost calculator, or compare V4 Pro head to head with Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash or Grok 4.7. Our prompt caching guide shows how far DeepSeek’s cache prices go on repeated-document work. The DeepSeek V4 Pro page shows today’s community votes on whether it feels dumber than usual, next to official incidents, and you can embed a live rating with a badge.