DeepSeek to Raise Prices from 8/17: Its Priciest Item Jumps 12x, and Taiwan's Whole Workday Falls in Peak Hours

Three days from now, at 12am on August 17, the bill for that DeepSeek program you’ve had running smoothly will jump to more than triple what it is now.

This isn’t a rumor — it’s an official announcement. On August 13, alongside the launch of the official V4 Pro release, DeepSeek announced an across-the-board API price increase, and for the first time introduced “peak-valley pricing” — selling tokens the way electricity is priced: expensive during peak hours, half-price during off-peak hours.

If there’s only one thing you need to know, it’s this: peak hours are 9:00-12:00 and 14:00-18:00 Beijing time. Taiwan, Hong Kong, and Singapore share the same time zone as Beijing (UTC+8). That means for people in Taiwan, the entire workday — from turning on your computer in the morning to clocking off in the afternoon — falls squarely within the most expensive time slot.

I. Price List: Copied Directly from the Official Source

DeepSeek’s official documentation lists prices in both USD and RMB simultaneously. Below are the prices per million tokens:

DeepSeek V4 Pro

DeepSeek V4 Flash

Effective time: Beijing/Taiwan time August 17, 2026, 00:00 (the official English documentation states 16:00 UTC on August 16, which converts to the same moment).

A note on the source of these figures: The USD prices in the table above are taken directly from DeepSeek’s official API documentation pricing page — this is first-hand data. The RMB prices do appear in the official Chinese-language announcement, but this site obtained the original figures via reporting from financial media outlets (Sina Finance, NetEase, Ifeng), making it second-hand. This site divided each of the six RMB line items by its corresponding official USD price, and the resulting exchange rates all fell consistently within 6.82–6.9, so the RMB figures reported by the media are judged to be credible. This site is only relaying information and does not represent this site’s position — please still defer to DeepSeek’s official announcement as the authoritative source.

II. How Much of an Increase? Don’t Just Look at the Percentage — Run the Real Numbers

The headline “up to 11x increase” is everywhere, and it refers to V4 Pro’s peak-hour cache-hit input price: ¥0.025 → ¥0.3, a 12x price, an 11x increase.

But looking at that single line item alone is misleading, because cache-hit pricing was already cheap to the point of being nearly free. Let’s run a number that looks more like a real workload.

Suppose you run an agent task using V4 Pro, consuming 500,000 input tokens (cache miss) plus producing 100,000 output tokens:

• Now: 0.5 × 3 + 0.1 × 6 = ¥2.1
• 8/17 off-peak: 0.5 × 4.5 + 0.1 × 13.5 = ¥3.6 (1.71x)
• 8/17 peak: 0.5 × 9 + 0.1 × 27 = ¥7.2 (3.43x)

If you run this kind of task 100 times a day, the difference over a month (30 days) is: ¥6,300 now, versus ¥21,600 if run at peak hours after 8/17. That extra fifteen thousand is money you pay purely because the calendar flipped a page — without changing a single line of code.

III. “Off-Peak Is Half Price” Is a Bit of Spin

This point needs to be spelled out clearly, because it’s the easiest part of the whole announcement to misread.

The announcement states that the off-peak price is half the peak price. That’s true, but that “half” is relative to the new peak price, not relative to what you’re paying now.

V4 Pro output: old price ¥6 → off-peak ¥13.5. The off-peak price, supposedly “already halved,” is still 2.25x what you’re paying today.

This is one of the most common traps in price-checking: conflating a post-discount promotional price with the pre-adjustment list price. When you see “half price,” don’t celebrate — ask “half of what number?” In this case, the answer is: half of a number that was just raised.

IV. This Site’s Take

(The following is Ask Cat editorial opinion, not news reporting)

First, the people actually being punished here are the ones who did things right.

The steepest increases are on cache-hit pricing — 12x for V4 Pro, 5x for Flash. Who are the people using prompt caching? Teams who pulled out their system prompts, pinned down long documents, and did serious engineering optimization. Everyone used to be taught that caching was the money-saving trick — now the unit price of that trick has jumped by an order of magnitude. By comparison, sloppy usage that resends everything in full every time only went up 3x. That signal is a bit hard to swallow.

Second, time zones mean Chinese-speaking developers are being singled out for an extra tax.

Peak-valley pricing itself makes sense — servers sit idle at night, so selling that capacity cheap is a win-win. But this scheme is tailor-made for mainland China’s daily rhythm, and Taiwan, Hong Kong, Macau, Singapore, and Malaysia happen to share the same UTC+8 zone. The result: a developer in San Francisco has their entire workday (which falls in the middle of the night to early morning in Taiwan time) land entirely in off-peak hours, automatically getting half price; a developer in Taipei has their entire workday land entirely in peak hours, paying double. Same program, same model — the price differs twofold purely based on where you are and when you hit run.

Your actual off-peak window is now just: lunch break 12:00-14:00, plus 18:00 in the evening through 9:00 the next morning.

Third, but don’t rush to switch providers just yet.

Emotionally, it’s easy to be upset — but rationally, look at the numbers. Even after the increase, V4 Pro’s peak output price is US$3.96 per million tokens. Compare that against the official price lists: Kimi K3 output is ¥100 (official documentation), Claude Opus 5 output is US$25, Claude Sonnet 5 output is US$10 (Anthropic’s official documentation). Even after the hike, DeepSeek remains among the cheapest in its tier.

So the accurate takeaway isn’t “DeepSeek is unusable now” — it’s “the era of using AI without thinking about the bill, like tap water or electricity, is over.” The bargain-basement prices everyone got used to over the past two years were, at their core, subsidies from companies burning cash to grab market share. Subsidies always end eventually. This is just the first bill.

Fourth, this isn’t just a DeepSeek thing.

At the same time, AI-related products from Zhipu, Tencent Cloud Hunyuan, Alibaba Cloud, and Baidu AI Cloud are all raising prices too. DeepSeek is getting singled out in headlines because of its fame, but it’s the tide, not the exception. If you’re currently building a product that only pencils out because of cheap tokens, this should be a trigger to recalculate your business model from scratch — not just a reason to switch providers.

V. Four Things to Do Before 8/17

1. Run your monthly bill numbers today. Pull up last month’s token usage and recalculate it against the new price list above. Don’t go by gut feeling — open a spreadsheet.
2. Reschedule non-urgent tasks to off-peak hours. Batch processing, data cleaning, daily reports, crawler summaries — anything nobody’s waiting on in real time should be scheduled for after 18:00 Taiwan time or during lunch break. This single move cuts your cost in half, with no model changes required.
3. Consider pre-loading your balance. In its August 6 advance-notice announcement, DeepSeek explicitly reminded users to “plan your usage needs and prepayment.” Whether an existing balance will still be billed at the old price isn’t clearly stated by the official announcement, and this site cannot guarantee that for you — please check the terms on the DeepSeek platform yourself before deciding whether to top up.
4. Re-evaluate your tiered model usage. After the increase, V4 Flash’s peak output price is ¥9 — still less than 70% of V4 Pro’s off-peak price of ¥13.5. When the price gap was small, “everything’s cheap anyway, just use Pro for everything” made sense. Now that the gap has widened, downgrading simple tasks to Flash delivers real, tangible savings.

One last word: the point here isn’t “prices went up, that’s expensive” — it’s that “the cost structure has changed.” AI billing used to be a flat line; now it has time-of-day, caching strategy, and model-tier dimensions to it. People who do this math can keep their cost increase to around 70%; people who don’t will simply pay more than triple.

The difference comes down to whether you do your homework in the next three days.

Sources

All links verified as of: August 14, 2026

First-hand (official)

• DeepSeek official API pricing documentation (USD prices, old/new price lists, peak-hour UTC ranges): https://api-docs.deepseek.com/quick_start/pricing
• Kimi official K3 pricing documentation (input cache hit ¥2 / cache miss ¥20 / output ¥100): https://platform.kimi.com/docs/pricing/chat-k3
• Anthropic official API pricing documentation (Claude Opus 5: input US$5 / output US$25; Sonnet 5: US$2 / US$10): https://platform.claude.com/docs/en/about-claude/pricing

Second-hand (media reporting on the content of official announcements)

• Sina Finance, “DeepSeek No Longer Bargain-Priced? API Prices Rise Up to 11x”: https://finance.sina.com.cn/stock/t/2026-08-13/doc-inineuqm9882414.shtml
• NetEase, “DeepSeek Announces API Price Increase, Peak Prices to Rise as Much as ¥27 Starting August 17”: https://www.163.com/dy/article/L485A7RT05198NMR.html
• cnBeta, “DeepSeek V4 Pro Official Release Goes Live: Agent Capabilities Upgraded, API Enables Peak-Valley Pricing”: https://www.cnbeta.com.tw/articles/tech/1573060.htm
• Sina Technology, “DeepSeek-V4-Pro Official Release Goes Live, API Pricing Adjusted”: https://finance.sina.com.cn/tech/roll/2026-08-13/doc-inineuqm9894249.shtml
• 21 Jingji, “DeepSeek Announces Advance Notice of API Price Increase, End of the AI Low-Price Bonus Era”: https://www.21jingji.com/article/20260807/herald/10ac18e0ab73de1dbadba766561e702f.html

This site’s disclaimer: The RMB prices and the content of the August 6 advance-notice announcement were not obtained directly from DeepSeek’s official website, but are cited from the reporting of the financial media outlets listed above (links as shown). This site is only relaying information and does not represent this site’s position; actual terms and prices are subject to DeepSeek’s official announcements.