DeepSeek cuts Flash prices by up to 60% — and the cache-hit line is the real story

Less than a month after raising API prices, DeepSeek is cutting them again. The company announced that from September 10, 2026, 12:00 Beijing time, the Flash series moves to new pricing — with a headline number of up to 60% off.

Here's the table, at off-peak rates (peak is double):

Both deepseek-v4-flash and deepseek-v4-flash-vision-exp move to the same rates — the vision model still bills at Flash pricing with no visual surcharge, consistent with how it launched in August (our post on that release).

The 60% headline is really a cache subsidy

Read the three lines separately and the intent is hard to miss:

• Cached input: down 60%. This is what you pay for tokens the model has already seen — a stable system prompt, a codebase, a document you keep re-sending. It just got 2.5x cheaper.
• Uncached input: down 33%. New material still costs real money, just less of it.
• Output: down 11%. The tokens the model writes barely moved.

That shape is a targeted discount on high-reuse agent workloads: long system prompts, repeated context, retrieval headers, tool schemas, multi-turn sessions that keep re-reading the same prefix. If most of your input is cached, your bill moves a lot. If most of your bill is generated output, this cut is closer to a rounding error.

The peak/off-peak multiplier still rules

Nothing changes about when you run. Peak hours remain weekdays 09:00–12:00 and 14:00–18:00 Beijing time, billed at double the off-peak rate, and weekends are entirely off-peak (the change we covered in weekends are now all off-peak).

Stack the two levers and the cheapest cell in the table is striking: a cached input token on a Sunday now costs ¥0.02 per million — while the same token at 10am on a Tuesday costs ¥0.04, and the output token that follows it costs ¥8 instead of ¥4.

A month of pricing whiplash

Line the moves up and the pattern is hard to miss:

• Mid-August: peak/off-peak pricing arrives, and reports describe steep increases on some line items (reported at up to 1,100%) — commentary at the time asked whether DeepSeek was giving up on being the price disruptor.
• August 23: weekends get folded into the cheap tier.
• September 9–10: the Flash series gets cut by up to 60%.

The fair reading: the increase overshot, and DeepSeek is walking it back — at exactly the moment a new model is waiting in the wings. Multiple outlets report V4.1 Flash is expected around September 10 and is already in internal testing, reportedly beating V4 Pro across the board. Cutting prices the same week you ship a model is a familiar play: land the new model on a cheaper meter and make "upgrade" an easy yes for everyone already running Flash. The same week, DeepSeek also started another large hiring round.

What this means if you build agents

1. Cache aggressively — it's now the highest-leverage lever you have. Stable prefixes, deterministic system prompts, tool schemas that don't churn: every token you keep in the cache costs ¥0.02 instead of ¥1. If you've been sloppy about prompt stability, the price list just made that expensive.
2. Re-run your cost model, because the ratios changed. Cached input is now 200x cheaper than output (¥0.02 vs ¥4 per million). Context-heavy agents got a real budget cut; generation-heavy agents got a rounding error. Both should re-check, only one should celebrate.
3. Schedule around the meter. Batch jobs, backfills, and evals belong in off-peak windows and on weekends — same code, half the price.
4. Don't over-index on one vendor's price list. A hike and a 60% cut inside a month is the strongest argument yet for keeping a routing layer that lets you move workloads between providers by task, price, and availability.
5. Watch for V4.1 Flash. If it ships at the new Flash rates with better benchmarks, the upgrade decision may be trivially obvious. Test it when it lands — and re-benchmark before assuming the price cut made your choice for you.

The bottom line

DeepSeek's "up to 60%" is real, but it isn't a 60%-cheaper model — it's a 60%-cheaper cache. The bill that falls the most belongs to agents that keep re-reading the same context; the bill that falls least belongs to the ones that mostly write. Know which one you are, then re-run the numbers: for the right workload, this is the cheapest Flash has ever been.