The DeepSeek V4 pricing update is live: alongside taking V4-Pro to general availability on 13 August 2026, DeepSeek switched its API to peak and off-peak pricing from 16:00 UTC on 16 August 2026 — off-peak rates are half the peak rates, peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays only, and V4-Pro output now tops out at 3.96 dollars per million tokens at peak versus 1.98 dollars off-peak. In plain terms: when you run your jobs now changes what your AI costs, and for most UK and US working hours the news is good.
📺 Watch: DeepSeek V4-Pro vs Expensive AI Models: IT'S OVER?
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
DeepSeek is the budget workhorse in a lot of my automations, so a pricing model change is not trivia — it directly moves the margin on every agent that runs on it. I have been through the official changelog and pricing page so you do not have to. Here is what changed, the exact numbers, and how to schedule your workloads so the change works for you rather than against you.
DeepSeek V4 Pricing Update: What Actually Changed
Two things landed in the same week, according to DeepSeek's API changelog. First, on 13 August 2026, DeepSeek-V4-Pro went to general availability — you access it by setting the model name to deepseek-v4-pro — with the release notes highlighting agent capabilities and posting benchmark scores of 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, and 42.7 without tools or 60.0 with tools on Humanity's Last Exam.
Second, the pricing system itself changed. From 16:00 UTC on 16 August 2026, the API bills by time of day: peak and off-peak, with off-peak set at half of the peak rate. Per the official pricing page, peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Everything else — including the entire weekend — is off-peak.
The Exact Numbers, Per Million Tokens
These are the figures from the official pricing page at the time of writing, shown as off-peak first, then peak:
| Model | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| deepseek-v4-pro | 0.022 / 0.044 dollars | 0.66 / 1.32 dollars | 1.98 / 3.96 dollars |
| deepseek-v4-flash | 0.007 / 0.014 dollars | 0.22 / 0.44 dollars | 0.66 / 1.32 dollars |
| deepseek-v4-flash-vision-exp | 0.007 / 0.014 dollars | 0.22 / 0.44 dollars | 0.66 / 1.32 dollars |
All three models carry a 1 million token context window with output up to 384,000 tokens, and the pricing page lists concurrency limits of 500 simultaneous requests for V4-Pro and 2,500 for V4-Flash and the vision model. Note the cache-hit column: input you have already sent recently is billed at a tiny fraction of fresh input, which is why well-structured agents with stable system prompts run so much cheaper than naive ones.
Inside AI Profit Boardroom I share the exact scheduling and caching setup that keeps my DeepSeek agents running on the cheap half of this pricing table — plus 1,000+ done-for-you workflows to run on them → Get the full setup
What Off-Peak Actually Means for Your Working Day
Here is the part that matters practically. The peak windows — 01:00 to 04:00 and 06:00 to 10:00 UTC, weekdays — sit in the early morning for Europe and overnight for the Americas. For a UK operator in British Summer Time, peak ends at 11:00 local; the entire afternoon and evening are off-peak. For the US, almost the whole business day is off-peak. And weekends are off-peak around the clock.
That makes batch scheduling a genuine money lever. Content pipelines, research runs, bulk rewrites, overnight report generation — anything that does not need to run at a specific minute should simply be queued for off-peak hours, instantly halving its token bill against peak rates. My agents already run scheduled jobs; the only change is telling the scheduler which hours are cheap. If you are new to the model family, my DeepSeek V4 tutorial covers the basics of Pro versus Flash and the thinking modes — this page is deliberately about what V4 costs, not how to prompt it.
📺 Watch: DeepSeek V4 FULL COURSE 6 HOURS (Build & Automate Anything)
Which Model to Point Your Money At
Ranked for value, in order: V4-Flash first for the overwhelming majority of agent work — at 0.66 dollars per million output tokens off-peak it is absurdly cheap for the quality, and it shares the same 1M context as Pro. V4-Pro second, reserved for the hard reasoning and agentic jobs where its Terminal Bench and DeepSWE scores earn the threefold output premium. The experimental flash-vision model third — released 21 August 2026 with image and screenshot understanding at Flash prices, and covered properly in my V4 Flash Vision Exp guide — worth testing, but it is explicitly experimental.
This is the same conclusion I keep reaching when we put models through Goldie Bench, our own benchmark: the gap between flagship and workhorse models on real tasks is smaller than the price gap, so defaulting to the cheaper tier and escalating only when quality demands it is nearly always the profitable configuration. It is also exactly how the Agent OS we run is wired — cheap model by default, expensive model on escalation — and this pricing update just made that pattern twice as rewarding on the right schedule.
Where you run DeepSeek matters too. Through the API these prices apply directly; if you pair the model with a coding harness, my best harness for DeepSeek V4 Pro breakdown and the DeepSeek harness guide cover the options, and for a completely free local setup DeepSeek V4 on Ollama sidesteps the API bill entirely at the cost of your own hardware.
The Cache Column Is the Other Half of the Story
The peak and off-peak split gets the headlines, but look again at the gap between the input columns: on V4-Pro, a cache miss costs 0.66 dollars per million tokens off-peak while a cache hit costs 0.022 — a thirtyfold difference for sending input the API has seen recently. On Flash the ratio is the same. For agent builders that gap is a design instruction, not a footnote.
The way to collect it is boring and effective: keep your system prompts, tool definitions and reference context stable and identical across calls, put the parts that change — the user request, the fresh data — at the end, and resist the urge to rewrite your prompt scaffolding every week. Agents that hammer the API with a consistent prefix get most of their input billed at the cache-hit rate automatically; agents assembled ad hoc pay full freight on every call. Combine a disciplined prompt structure with off-peak scheduling and the two discounts stack — half-price hours on top of thirtyfold-cheaper repeated input is how a serious automation pipeline ends up costing pennies.
One operational note from the same pricing page: concurrency is capped at 500 simultaneous requests for V4-Pro and 2,500 for the Flash models. For most solo operators that is headroom you will never touch, but if you batch aggressively into the cheap hours, it is worth knowing the ceiling exists before your scheduler finds it for you.
DeepSeek V4 Pricing Update FAQs
When did the DeepSeek V4 pricing update take effect?
The peak and off-peak system took effect at 16:00 UTC on 16 August 2026, per the official changelog, three days after V4-Pro reached general availability on 13 August.
What are DeepSeek's peak hours?
According to the pricing page, peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. All other hours — and the whole weekend — bill at off-peak rates, which are half the peak rates.
How much does DeepSeek V4-Pro cost now?
Per million tokens: output is 1.98 dollars off-peak and 3.96 dollars at peak; cache-miss input is 0.66 and 1.32 dollars; cache-hit input is 0.022 and 0.044 dollars.
Is V4-Flash really the same price as the vision model?
Yes — the pricing page lists deepseek-v4-flash and deepseek-v4-flash-vision-exp at identical rates, with the vision variant adding image and screenshot understanding as an experimental capability.
Verdict: Cheap Just Got a Timetable
DeepSeek was already the value play; the DeepSeek V4 pricing update turns value into a scheduling game. Half-price tokens for most of the Western working day and all weekend is a standing invitation to move every non-urgent workload off-peak — and combined with aggressive prompt caching, the effective cost of a well-built agent on V4-Flash is now closer to a rounding error than a line item. The operators who win with AI are not the ones with the biggest budgets; they are the ones who read the pricing page.
Want your automations built cost-first — the model choices, the schedules, the caching tricks that keep margins fat? That is exactly what I teach inside AI Profit Boardroom, with live calls every week → Build profitable agents with me
