DeepSeek's Price Sheet Has Two Levers, and the Clock Is the Smaller One

Everyone reads DeepSeek's pricing page for the peak windows. Between 01:00 and 04:00 UTC, and again between 06:00 and 10:00, Monday to Friday, Chinese public holidays excluded, every rate doubles. That is real, it is seven hours out of twenty-four, and it is worth planning around. But it sits two rows below a number that changes the arithmetic far more: a cache hit costs $0.003 per million input tokens on DeepSeek V4.1 Flash, against $0.15 for a cache miss. A prefix that lands in cache is fifty times cheaper than the same prefix sent cold — and unlike the clock, that is a lever entirely inside your control.

Put the two levers side by side and the ordering is obvious. The peak window is a 2× penalty you avoid by rescheduling. A cache hit is a 50× discount you earn by structuring requests. Almost every article written about DeepSeek's price sheet, including earlier versions of this one, spends its words on the 2×.

The card, exactly as DeepSeek publishes it

Read from DeepSeek's Models & Pricing page on 30 October 2026. Off-peak first, peak in brackets, per million tokens:

• Input, cache hit — $0.003 for Flash ($0.006 peak); $0.022 for V4 Pro ($0.044 peak).

• Input, cache miss — $0.15 for Flash ($0.30 peak); $0.66 for V4 Pro ($1.32 peak).

• Output — $0.60 for Flash ($1.20 peak); $1.98 for V4 Pro ($3.96 peak).

• Peak hours — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Off-peak rates are exactly half of peak.

• Concurrency — 2,500 for Flash, 500 for V4 Pro. A fifth, as the price ratio suggests.

• Shape — both tiers serve a 1M-token context with a 384K maximum output; deepseek-flash is served by DeepSeek-V4.1-Flash-1017 and supports vision, deepseek-v4-pro is served by DeepSeek-V4-Pro-0813 and does not.

The cache-hit column is the one to stare at. Three tenths of a cent per million input tokens means a workload that reuses a long system prompt, a large document, or an agent's accumulated history is paying almost nothing for the part it re-sends — provided the request is shaped so the prefix is recognised as the same prefix.

What the two levers are worth on one workload

Take a realistic agent-shaped job: 20 million input tokens and 2 million output tokens per day, where 80% of the input is a reused prefix.

Lever one, caching. Sixteen million cached tokens at $0.003 is $0.048. Four million uncached at $0.15 is $0.60. Output at $0.60 per million is $1.20. Total: about $1.85 a day.

Lever two, scheduling. The same job with no caching, all of it off-peak: 20 million at $0.15 is $3.00, plus $1.20 of output — $4.20. Run it entirely inside a peak window and it is $8.40.

So caching saves $2.35 here and the best possible scheduling saves $4.20 against the worst. Both matter; neither is the whole story, and the combination is where the real number lives: cached and scheduled off-peak, the same job is about $1.85 instead of $8.40 — a 4.5× swing on identical tokens.

What that means practically is that the ordering of decisions should be: shape the requests for prefix reuse, then move what can move out of the peak windows, then compare vendors. Most teams do it in the reverse order, which is why the clock gets all the attention — it is the only lever that appears on the page as a bold heading.

Why the cache line is easy to miss and easy to lose

Three things can quietly cost you the 50× discount./h2> p>A prefix that changes at the front. Caching works on the leading tokens. Put a timestamp, a session id, or a shuffled tool list at the top of the system prompt and every request is a cache miss — while looking identical to the person who wrote it. This is the single most common way a cached workload turns into a full-price one.

A provider path that does not share cache. Cache is a property of the serving infrastructure, not of the model name. A router with automatic failover can move a request to a different provider path mid-workflow, and the cache benefit does not travel with it. That is not a reason to avoid failover — a failed request costs more than a cache miss — but it is a reason to know which share of your traffic is landing on the primary path.

An identifier whose checkpoint changed. deepseek-flash now resolves to DeepSeek-V4.1-Flash-1017, and the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but billed at the Flash price and served by the same model. A cost model written against an older checkpoint is describing a card that has already moved once this month.

The clock, where it does bite

The peak window is still a real constraint for anything that cannot move. Both windows sit inside a European and Asian working day and outside a US one: 09:00–12:00 and 14:00–18:00 in Beijing, 02:00–05:00 and 07:00–11:00 in London, and for a US East Coast team the second window starts at 02:00 local. An interactive product serving European users is in peak rates for most of its traffic whether or not anyone planned it.

Two second-order effects are worth knowing before rewriting a scheduler:

• Concurrency, not just price. V4 Pro allows 500 concurrent requests against Flash's 2,500. Moving batch work into the off-peak window to save money does not help if everyone else moved there too.

• The holiday calendar. The windows exclude Chinese public holidays, which means the calendar your scheduler ships with is not the one you need. It is a small recurring tax, and it surfaces on the one day a year it matters.

The rule that survives a price change

Structure for cache hits first, schedule second, and keep the per-token comparison for last — because the first two levers move your bill by 50× and 2× respectively, and no vendor switch reliably moves it by either. Then log the UTC hour and the cache-hit share on every request for a week. Those two numbers, not the rate card, are what your next month costs.

DeepSeek V4.1 Flash and DeepSeek V4 Pro are both on OrcaRouter at DeepSeek's published rates, passed through with no markup on the same key as a 224-model catalogue. The cache-hit line and the peak doubling apply exactly as DeepSeek defines them, which cuts both ways: the discount is live the same day it is published, and so is the surcharge.