DeepSeek's New Peak Pricing Hits Cache-Heavy Workloads with a 12× Increase

At 7 PM on the evening of August 13th — just now! I saw DeepSeek's official WeChat account push a pricing update, which also semi-launched V4-Pro. The new prices take effect at midnight on August 17, 2026. The price hike itself isn't surprising — once a model reaches the Pro tier, it has to start aligning with the market. But when the numbers were laid out, it was a bit harsher than expected.

How Much Did It Go Up

Let's look directly at the numbers. The old price is the tier from the official table before the hike, and the new price is split into two tiers: 'Idle / Peak', with the unit uniformly in CNY / million tokens.

Model	Billing Item	Old Price	New-Idle	New-Peak	Increase (Idle)	Increase (Peak)
V4-Flash	Cache Hit	0.02	0.05	0.10	2.5×	5×
V4-Flash	Cache Miss	1	1.5	3.0	1.5×	3×
V4-Flash	Output	2	4.5	9.0	2.25×	4.5×
V4-Pro	Cache Hit	0.025	0.15	0.30	6×	12×
V4-Pro	Cache Miss	3	4.5	9.0	1.5×	3×
V4-Pro	Output	6	13.5	27.0	2.25×	4.5×

Peak hours are Beijing time 9:00–12:00 and 14:00–18:00, which is basically domestic working hours — so the vast majority of usage scenarios will actually get the peak price.

A Few Glaring Points

1. Pro's cache hit price directly increased 12 times. During peak hours, Pro's cache hit went from 0.025 → 0.30, the steepest hike among the six billing items. What this means: the prefix cache optimizations and long-context reuse techniques you pride yourself on, this single tier is 12 times more expensive than before. Scenarios like RAG/Agents that cache a few K of system prompts and make high-frequency calls will be hit the hardest.

2. Input cache miss only increased 1.5×, but output generally starts at 2.25×. Both Flash and Pro are very restrained on the 'input cache miss' tier (both 1.5×), but the output side is pushed to 2.25× or more — clearly pushing script-type and long-reasoning output scenarios towards being more expensive. Agent orchestration, automated evaluation, long document summarization, and batch generation scripts all fall into this tier.

3. Flash is no longer 'dirt cheap'. Flash's old cache hit price was 2 cents per million tokens; the new peak-hour price is 10 cents. Flash's core selling point of being 'cheap and high-volume' is basically gone. I guess this is also to make way for Pro — with Pro semi-launched, its pricing can't be too close to Flash's.

4. Another point easy to miss: in the old table, Flash's concurrency limit was 2500, and Pro's was 500. The new table doesn't mark a concurrency cap. Whether it can still run at this level depends on the actual performance on the 17th. If concurrency is squeezed hard, the combination of 'increased unit price + squeezed concurrency' is the real heavy blow.

How Will the Actual Bill Change

Take a common call structure as an example: 1 million input tokens (cache miss) + 1 million output tokens.

Model	Old Price	New-Idle	New-Peak
V4-Flash	3 CNY	6 CNY	12 CNY
V4-Pro	9 CNY	18 CNY	36 CNY

A single Pro call during peak hours directly results in a 4× bill. If an office worker uses the API normally during the day, almost all calls fall into the peak period.

How I See It

• It's not a rip-off, but a normalization of the pricing strategy. Flash served as a loss leader for nearly a year, so it's not surprising that it's now being raised to align with the mainstream pricing of domestic models. Several companies at home and abroad are making this move.
• Pro's increase is a bit fierce. The 12× cache hit tier is clearly creating a stratification: 'high-value scenarios use Pro, high-volume but low-unit-price scenarios use free quota + queuing.' If you previously jumped on board because of Pro's price advantage, you need to recalculate your ROI now.
• Peak time-of-use pricing is a trend. DeepSeek has now explicitly stated that 'it's more expensive during working hours,' and other companies will likely follow suit. Those running batch tasks should learn to shift their scripts to late night.
• Actionable advice: In the two or three days before August 17th, either finish your batch tasks, or re-examine your cache design/call timing; do a cost review for projects using Pro, split the cacheable segments in your prompts, and don't just throw one prefix at the whole thing anymore.

When the axe falls at 0:00 on August 17th, the real story will be told by the bills that come out then.