梁圣回来了，缓存价格直降60%Saint Liang Is Back: Cache Prices Drop a Straight 60%

Let's start with the bottom line.

News Flash: Flash Series Repriced, Effective Noon on September 10

Let's start with the bottom line.

Starting at 12:00 on September 10, 2026, Beijing time, DeepSeek is adjusting pricing for the flash series. The new off-peak prices are (unit: yuan / million tokens):

• Input cache hit: ¥0.02
• Input cache miss: ¥1
• Output: ¥4

Peak-hour prices are 2× the off-peak prices, i.e. ¥0.04 / ¥2 / ¥8.

The biggest winners are multi-turn conversation, long-context, and agent workflow scenarios — in their bills, the cache-hit portion usually dominates, and that tier just dropped a full 60%.

The original official notice reads:

Starting at 12:00 on September 10, 2026, Beijing time, we will adjust pricing for the flash series: during off-peak hours, the unit price for input cache hits is ¥0.02, for input cache misses ¥1, and for output ¥4; peak-hour prices are 2× the off-peak prices. Please plan your usage accordingly.

Old vs. New Pricing

Based on the current published rates on the official pricing page (deepseek-v4-flash, verified in the early hours of 2026-09-09) and this announcement, the three tiers change as follows:

One framing point has to be spelled out here: the "straight 60% cut" in the headline refers to the cache-hit tier (¥0.05 down to ¥0.02) — not every tier drops 60%. Cache misses fall by a third, and output by about a tenth. The three reductions differ widely, so budgeting off "40% of the old price" will get the math wrong.

Another thing that didn't change is worth noting: peak is still 2× off-peak, and the multiple holds exactly as before. Under the current rules, peak hours are Monday–Friday 9:00–12:00 and 14:00–18:00, and off-peak prices are half of peak — the new notice says the same thing, which means the peak/off-peak mechanism itself wasn't touched.

How Much You Actually Save Depends on Your Cache Hit Rate

The reduction isn't a fixed number; it depends on the share of your requests that hit the cache. Take an example (the figures below are estimates based on the table above, purely to illustrate the structure, and do not represent any actual bill):

Suppose an agent workflow processes 1 million tokens of input per run, 90% of which hit the cache and 10% miss, and produces 50,000 output tokens.

• At current off-peak prices: 0.9 × 0.05 + 0.1 × 1.5 + 0.05 × 4.5 ≈ ¥0.42
• At new off-peak prices: 0.9 × 0.02 + 0.1 × 1.0 + 0.05 × 4.0 ≈ ¥0.32

That's an overall drop of about 24%. With a higher hit rate — say 95% — the overall reduction gets closer to 30%; conversely, if a task keeps feeding in brand-new long documents with a very low hit rate, what you mainly get is the 33% cut on the cache-miss tier.

The conclusion is straightforward: this new price sheet is friendliest to workflows that re-read the same batch of context. Multi-turn agent iteration, repeated Q&A over long documents, retrieval from a fixed knowledge base, and batch runs over the same prompt set all fall into that category. One-shot tasks whose context changes every time benefit comparatively little.

An Old Piece of Advice, Now Worth More: Move What You Can to Off-Peak

Peak is 2× off-peak, and that multiple still holds under the new price sheet. Which means moving a batch job that used to run at ten on a Tuesday morning to the evening is effectively another 50% off — and that discount stacks with this price cut.

Using the example above: the same task at peak hours costs about ¥0.84 at current prices and about ¥0.64 at the new ones; moved to off-peak, the new price is about ¥0.32. Stacking the price cut with off-peak scheduling is the only way to get the full benefit of this repricing.

In the DSH ecosystem specifically, two things we covered in our last issue line up neatly here:

First, the Little Whale balance widget, which already supports peak/off-peak price conversion — with a token configured it can pin token usage down to the hour. Once the new prices take effect, glancing at real-time consumption against the new price sheet is far more intuitive than digging through the bill afterward. Second, dsh-web's "Liang God mode" agent preset, which is built around hands-on tuning for the flash series (community benchmark average 98.5); one of the design goals behind its two-stage anchoring is to eliminate unnecessary tool overhead in the first round. We won't expand on either here — see last issue's detailed coverage.

About "Saint Liang Is Back"

The moment the repricing news broke, people in the community were already shouting "Saint Liang is back." To be clear, this is the community's emotional reaction to the price cut and a running joke; the full official notice covers only pricing and time-slot arrangements and mentions no personnel matters whatsoever. This article does not speculate about the reason for the repricing, nor does it draw any causal link between it and any individual's decisions.

Effective Date and Sourcing Notes

• The new prices take effect at 12:00 on September 10, 2026, Beijing time; before that, billing continues at current prices.
• The price information in this article comes from two places: the effective time, the three new tier prices, and the peak/off-peak multiple come from the official notice; the current prices and the peak/off-peak time rules come from the official pricing page (deepseek-v4-flash, verified in the early hours of 2026-09-09). The "Reduction" and "New peak price" columns in the table are derived from those two sources.
• Final billing is subject to the official notice and the platform's actual bills. If your usage is heavy, we suggest running one or two small tasks after the change takes effect and checking them against your bill before scaling up.

One last reminder: price is a long-term variable, not a one-off red envelope. If you really want to save money, cache hit rate and off-peak scheduling will outlast any single price cut.