What "Half Price Off-Peak" Concealed: DeepSeek Switches V4's API Pricing to Time-of-Day Billing and Hands Out a Sticker Price That's Actually Higher Than Today's

The First Word: "Off-Peak Is Half Price"

On August 16, DeepSeek is significantly restructuring its API pricing. Timed to the general availability of the V4 series, it is abandoning its previous flat-rate pricing in favor of a two-tier "peak/off-peak" structure where the unit price shifts by time of day. The change takes effect at 16:00 Coordinated Universal Time (UTC) on August 16.

The official announcement said that "off-peak is 50% cheaper than peak." Read on the numbers alone, it looks like a discount notice. But the basis for that comparison is not "cheaper than now"—it's "cheaper than the new peak rate." That one-word difference is what determines the real cost you carry.

The New Sticker Price That Changes by Time of Day

The peak windows fall at UTC 01:00–04:00 and 06:00–10:00. Every other hour is treated as off-peak. Converted to Japan Standard Time (UTC+9), the peaks land at 10:00–13:00 and 15:00–19:00, which means much of the domestic workday sits on the peak side.

Here is the per-million-token pricing for the main models, laid out alongside the current rates (sources: DeepSeek's official pricing page, plus a reconstruction from various reports).

Even "Half Price" Is Higher Than Today

The key point is that even off-peak is more expensive than the current rate. Flash output goes from $0.28 to $0.66 off-peak (about 2.4x) and $1.32 at peak (about 4.7x). Pro output climbs by roughly the same multiple, from $0.87 to $1.98–$3.96.

Prompt-cache reads are even more striking. Pro's cache-hit input jumps from $0.003625 to $0.044 at peak—a more than tenfold increase. The more your usage leans on repeatedly reading long contexts, the harder the price hike will hit.

It Hits the Coding Agent's Wallet

This change lands directly on teams that use DeepSeek through a coding agent. Codex and Cline can both select DeepSeek V4 over an API connection. Because agents re-read long contexts over and over, the increases on the input and cache sides feed straight into execution cost.

The time-of-day design can't be ignored either. Since Japan's daytime overlaps with the peak windows, interactive, hands-on work tends to land in the expensive hours. Whether you can shift unattended batches and overnight bulk jobs to off-peak is what will move your monthly bill.

Pause on the Word "Discount"

This episode is also a reminder to watch how a pricing change is framed. "50% off at off-peak" is factually true, but it's a phrasing that anchors the comparison to the new peak. If you take the first social-media report at face value as a "price cut," you may overlook the real cost increase. The numbers always need to be read against the current rates.

That said, time-of-day billing has an upside, too. If load can be flattened, the provider uses its infrastructure more efficiently, and users can hold down their unit cost by shifting work. In the end, it all comes down to whether you know which hours your own workload is concentrated in.

What to Check Right Now

• Measure which UTC hours your organization's DeepSeek usage is concentrated in.
• Move any shiftable batch processing to off-peak (roughly UTC 10:00 onward).
• Re-run the math on cache-heavy workflows, factoring in that the increase there is large.
• For cost-sensitive use cases, take fresh unit-price comparisons against other models.

Reprinting the sticker price is a matter of operational design, not of performance. Now that the "it's cheap" reputation you chose it for has shifted, it's time to measure that cost all over again.

References: DeepSeek API official pricing page / DeepSeek official announcement (X) / TechApple Global / AI Pricing Guru: DeepSeek V4 Pricing Guide