What happened

On August 16, 2026, DeepSeek quietly replaced its flat pricing for the V4 model family with a two-tier structure: a cheaper off-peak rate and a peak rate exactly 2x higher, with no in-between steps. Within hours, screenshots calling this an "1100% price hike" started circulating, based on one line in the pricing table: cached input tokens on the Pro model jumping from $0.003625 to $0.044 per million tokens in peak hours, a 12.14x multiplier.

That number is technically accurate, but it describes one cell out of six in the pricing table, and it's the cell that carries the least weight in a real invoice. Run the other five line items, flash cache-hit input, flash cache-miss input, flash output, pro cache-miss input, and pro output, through the same math, and the increase ranges from 3.0x to 5.0x during peak hours and 1.5x to 2.4x off-peak. Model actual workload profiles (a chat assistant, a RAG pipeline, a batch summarization job) against the full table, and the blended increase lands between 2.3x and 2.9x, not 12x.

The second discovery matters just as much: DeepSeek defines peak hours in UTC as 01:00 to 04:00 and 06:00 to 10:00. Convert that into your local time zone and the expensive window often turns out to be a small, movable slice of the business day rather than an all-day tax.

Why it matters

### The 12x figure is a statistical outlier, not a bill forecast

The $0.003625 rate wasn't a normal price, it was a promotional rate DeepSeek introduced on April 26, 2026, when it cut cached-input pricing by 90% from its starting point. Compare the new peak rate of $0.044 against the pre-promo price of $0.03625 from four months earlier and the real increase is +21%, not +1100%. The viral number measures the distance from the bottom of a temporary discount, on the cheapest line in the entire price sheet.

In practice, cached input tokens barely move the needle on a typical bill. For a chat assistant with roughly 50% cache hit rate, that line item accounted for 0.59% of the old invoice and 2.04% of the new one. Even in a cache-heavy RAG setup, where cached tokens make up 80% of input, the share only grows from 2.79% to 9.71%. To actually hit something close to the 12x multiplier on a full invoice, a workload would need to be almost entirely cache hits with almost no output tokens, a scenario tested at 100,000 tokens of cached context, a 1-token response, and one cache miss per 1,000 requests, which still only produced an 11.15x blended increase. No production workload looks like that.

### Peak hours are a scheduling problem, not a flat surcharge

Because peak pricing is exactly double off-peak pricing and applies to specific UTC hours, the real lever isn't negotiating a lower rate, it's shifting movable workloads (batch jobs, embeddings, nightly summarization, non-urgent agent runs) outside the two peak windows.

How to use it today

Start by mapping DeepSeek's UTC peak windows onto your own time zone, since the two peak blocks (01:00 to 04:00 and 06:00 to 10:00 UTC) land very differently depending on where your team or your automation runs.

Time zone reference for peak windows: UTC/GMT runs peak 01:00-04:00 and 06:00-10:00. EST (UTC-5) runs peak 20:00-23:00 the previous day and 01:00-05:00. PST (UTC-8) runs peak 17:00-20:00 the previous day and 22:00-02:00. IST (UTC+5:30) runs peak 06:30-09:30 and 11:30-15:30. MSK (UTC+3) runs peak 04:00-07:00 and 09:00-13:00.

Best AI tools saved weekly in our channel — @aigobySubscribe →

For a Moscow-based team, this means only the 10:00 to 13:00 slice of a normal working day falls inside peak pricing, everything after 13:00 runs off-peak until 04:00 the next morning. For a US East Coast team, the second peak window (01:00-05:00 EST) sits entirely inside off-hours, so daytime API calls are rarely affected at all.

Once you know your windows, the fix is mechanical:

1. Pull your last 30 days of DeepSeek usage logs and tag each request with a timestamp.

2. Separate latency-sensitive traffic (live chat, real-time agents) from movable traffic (batch classification, embeddings, report generation, scheduled content runs).

3. Reschedule the movable share with a cron rule that delays dispatch until your local peak window closes.

4. Recalculate your blended multiplier using your own traffic mix, not the headline number, the six-line pricing table above is enough to do this by hand in a spreadsheet.

If you want to stress-test prompts, compare model outputs, or prototype a scheduling script before rolling it into production, a free AI tool like [mykreatool.com](https://mykreatool.com) is a fast way to sanity-check the logic without burning paid API credits first.

Who benefits

Teams running batch or asynchronous AI workloads, RAG indexing, content generation pipelines, data labeling, nightly summarization, stand to gain the most, since these jobs can be shifted into off-peak windows with a scheduler change and no product impact. Solo builders and small agencies running cost-sensitive tools also benefit, because the real 2.3x-2.9x increase is manageable when averaged against overall margins, especially once movable traffic is rescheduled. Finance and DevOps leads tracking AI spend get a cleaner way to model costs: use the six-line pricing table against your own traffic mix instead of trusting aggregate headlines.

Risks

The most obvious risk is budgeting off the viral 12x figure instead of your actual usage profile, teams that panic-migrate providers or over-provision budget based on the outlier line item will overcorrect for a cost increase that, in most real workloads, is closer to 2.3x-2.9x. A second risk is assuming peak windows are fixed forever; DeepSeek can change UTC boundaries or the 2x multiplier itself in a future update, so any cron-based rescheduling should be revisited periodically rather than set once and forgotten. Finally, latency-sensitive traffic shouldn't be shifted purely to chase off-peak pricing, delaying real-time chat or live agent responses to save on API costs will hurt user experience more than it saves on the bill.

Conclusion

DeepSeek V4's new pricing looks alarming at first glance, but the 1100% figure describes one narrow, low-weight line item pulled against a temporary 2026 promo rate, not what most teams will actually pay. Run your own traffic profile against the full six-line price table, and the realistic increase is 2.3x to 2.9x. Combine that with the UTC peak-hour schedule, shift what you can into off-peak windows, and the effective cost increase shrinks even further.