This DeepSeek API pricing update takes effect at 16:00 UTC on August 16, 2026, and it's not a minor tweak: DeepSeek is moving from one flat rate to peak/off-peak billing, and output tokens on some models get up to 4.7x more expensive during busy hours. If you're calling the API today at the old flat rate, tomorrow's bill will look different no matter what time zone you're in.
Short answer: DeepSeek's new peak/off-peak pricing starts August 16, 2026, at 16:00 UTC. Off-peak rates (most of the day) run roughly 1.6–2.4x today's flat price; peak rates (01:00–04:00 and 06:00–10:00 UTC) run up to 4.7x higher. DeepSeek-V4-Flash output goes from $0.28 to $0.66–$1.32 per million tokens. Direct API, OpenRouter, and Together AI now land at different price points for the same models.
Disclosure: AISagely may earn a commission if you sign up through some links on this page, at no extra cost to you. DeepSeek itself has no affiliate program — I have no financial stake in whether you keep using it after this price change.
I've had a DeepSeek API key running in a side project since V4 launched, so this update landed in my own invoice, not just my reading list. In my testing this week I pulled the live numbers straight from DeepSeek's pricing docs, then checked what the same models cost on two third-party routers to see whether switching actually saves money or just moves the bill somewhere else.
What this DeepSeek API pricing update actually changes
Right now, through 15:59 UTC on August 16, 2026, DeepSeek bills a single flat rate no matter when you call the API. After that, every request gets billed at either a peak or off-peak rate, and peak hours are specific: 01:00–04:00 UTC and 06:00–10:00 UTC, which is roughly 7 hours a day. Everything else — about 17 hours — counts as off-peak, but off-peak is still pricier than today's flat rate.
Here's what changes for the two current models, in dollars per million tokens:
| Now (flat) | Off-peak (from Aug 16) | Peak (from Aug 16) | |
|---|---|---|---|
| V4-Flash input (cache miss) | $0.14 | $0.22 | $0.44 |
| V4-Flash output | $0.28 | $0.66 | $1.32 |
| V4-Pro input (cache miss) | $0.435 | $0.66 | $1.32 |
| V4-Pro output | $0.87 | $1.98 | $3.96 |
Cache-hit pricing barely moves in dollar terms — V4-Flash cache hits go from $0.0028 to $0.007 (off-peak) or $0.014 (peak) per million tokens, still a fraction of a cent. That matters more than it sounds: if your app reuses the same system prompt or long context across requests, caching absorbs most of this increase. If most of your traffic is fresh, uncached prompts, you'll feel the full jump. DeepSeek has framed the change as a way to spread developer traffic away from its busiest hours rather than a straightforward price hike, but the off-peak floor is still higher than what anyone pays today.
How I checked these numbers
I pulled every price in this piece directly from DeepSeek’s own pricing page on August 15, 2026, the day before the change lands, then cross-checked the peak/off-peak figures against TechNode’s coverage of the same announcement. For the alternatives below, I checked each router's live pricing page the same day rather than trusting a cached number, because routers reprice models on their own schedule, independent of DeepSeek's.
I didn't rank these three by "cheapest possible number," since that number depends on your traffic pattern. I ranked them by which one actually saves money for a specific, realistic way of using the API.
Top picks at a glance
| Option | Best for | V4-Flash price (in/out per 1M) | Tied to DeepSeek's clock? |
|---|---|---|---|
| DeepSeek direct API | Lowest possible cost if you can batch off-peak | $0.22 / $0.66 (off-peak), $0.44 / $1.32 (peak) | Yes |
| Together AI | Flat, predictable billing | $0.14 / $0.28 (cached input $0.03) | No |
| OpenRouter free tier | Testing and low-volume personal use | $0 on :free model variants |
No |
Prices change fast in this category — confirm current numbers on DeepSeek’s pricing page and each router's own page before you commit spend.
Best overall: DeepSeek's direct API
If your workload can shift into off-peak hours, DeepSeek's own API is still the cheapest way to run V4-Flash or V4-Pro, and it's the only place you get the models the day they ship. In my testing, a batch summarization job I run overnight already lands entirely inside the 17-hour off-peak window without any scheduling changes on my end — off-peak covers most of the clock, so "cheapest" mostly just means "not one of the two peak blocks."
The real advantage is caching. DeepSeek's cache-hit price barely increases in absolute terms, so an app that reuses a long system prompt or a shared document context across many calls will see a much smaller bill increase than the headline numbers suggest. Where direct API stops being the obvious choice is unpredictability: if you can't control when your users' requests fire, you'll eat some peak-hour billing whether you planned for it or not. Get an API key at platform.deepseek.com if you want the newest models straight from the source.
Best for value: Together AI
This is the pick that actually got cheaper, relatively speaking, once DeepSeek's own price moved. Together AI runs DeepSeek-V4-Flash at a flat $0.14 per million input tokens ($0.03 cached) and $0.28 per million output tokens — no peak or off-peak math, and no relation to DeepSeek's billing clock at all. After August 16, that flat rate is cheaper than even DeepSeek's own off-peak rate for Flash ($0.22/$0.66), and it stays that way at any hour.
It's not a universal win — Together's V4-Pro pricing ($1.74 input / $3.48 output) sits above DeepSeek's own off-peak Pro rate ($0.66/$1.98), so for Pro specifically, direct API off-peak still wins on price. Together's real value is predictability: if your team needs to forecast a monthly bill without accounting for time-of-day surcharges, a flat rate that doesn't change is worth more than chasing the theoretical lowest number. Try Together AI here: Together Ai.
Best free option: OpenRouter's free tier
For testing, prototyping, or genuinely low-volume personal projects, OpenRouter hosts free :free variants of several DeepSeek models — I confirmed DeepSeek-V3.2 currently has one at $0 per token. It's rate-limited: 20 requests per minute either way, and 50 requests per day if you've never bought credits, rising to 1,000 requests per day once you've purchased at least $10 in credits over the account's lifetime.
That's not enough for production traffic, but it's genuinely useful for trying a prompt or building a demo without a payment method on file, and it sidesteps DeepSeek's peak/off-peak billing entirely since OpenRouter sets its own terms for the models it hosts. Paid DeepSeek models on OpenRouter exist too, but pricing varies by which upstream provider handles the request, so check the specific model page before assuming a number. Start free at Openrouter.
How to choose the right option
Start with your traffic pattern, not the sticker price. If your jobs are batchable — nightly processing, scheduled reports, anything that doesn't need to run the instant a user clicks something — DeepSeek's direct API off-peak rate is hard to beat, and you get the newest models first. Our guide to DeepSeek V4 Flash walks through getting an API key and making your first call if you haven't set this up yet.
If your traffic is real-time and unpredictable — a live chat feature, a customer-facing tool — you can't reliably dodge DeepSeek's peak windows, so a flat-rate router like Together AI removes the guesswork even if it's occasionally not the single cheapest option. And if you're still deciding whether DeepSeek is the right model family at all, how to use DeepSeek and how to use DeepSeek for coding cover the basics, while our wider best AI models comparison and DeepSeek vs ChatGPT look at whether a different model family fits your use case better than chasing DeepSeek's price changes at all.
One more thing worth checking before you commit spend anywhere: if you're building on DeepSeek's newer agent tooling rather than calling the API directly, our DeepSeek Harness developer preview review covers what that layer costs and where it broke in testing.
My verdict
DeepSeek is still cheap by frontier-model standards, but this update quietly removes its "obviously cheapest" claim for anyone running peak-hour, uncached traffic. V4-Flash's new peak output price, $1.32 per million tokens, is over three times what OpenAI charges for GPT-5-nano output ($0.40 per million, per OpenAI's pricing page) — a comparison that would have been unthinkable for DeepSeek a year ago. Off-peak, DeepSeek is still competitive; at peak, it isn't automatically the budget option anymore.
My recommendation: if you can batch your workload into off-peak hours, stay on DeepSeek's direct API — it's still the best deal there. If you can't control timing, or you just want one predictable number on your invoice, Together AI's flat rate is now the safer default for Flash specifically. Skip switching entirely if your volume is small enough that the dollar difference is a rounding error either way — that's most side projects, and it's not worth the migration effort. For anyone just starting out, our list of free AI tools has other no-cost options worth trying alongside DeepSeek's free tier.
Frequently Asked Questions
What is the DeepSeek API pricing update?
It's a shift from a single flat rate to peak/off-peak billing, effective August 16, 2026, at 16:00 UTC. Peak hours (01:00–04:00 and 06:00–10:00 UTC) cost more than off-peak hours, and both are priced higher than today's flat rate.
When does the new DeepSeek pricing take effect?
16:00 UTC on August 16, 2026. Any request made before that time bills at the old flat rate; anything after bills at the new peak or off-peak rate depending on when it lands.
How much more expensive is DeepSeek's API now?
It depends on the model and the hour. V4-Flash output goes from $0.28 to $0.66 (off-peak) or $1.32 (peak) per million tokens. V4-Pro output goes from $0.87 to $1.98 (off-peak) or $3.96 (peak). Cached input tokens barely change in dollar terms.
Is there a free way to use DeepSeek's models?
Yes. DeepSeek's consumer chat app remains free with no paid tier, and OpenRouter hosts free, rate-limited :free variants of some DeepSeek models for API-style testing. Neither is meant for production-volume traffic.
Should I switch away from DeepSeek because of this price increase?
Not automatically. If your traffic can run off-peak, DeepSeek's direct API is still the cheapest option I found. If your traffic is real-time and you want predictable billing, a flat-rate router like Together AI is worth checking — but for most low-volume users, the actual dollar difference is small enough not to bother.