DeepSeek's cache discount is a bigger cost lever than its peak-hour surcharge

Two levers on one price page

As OrcaRouter reported, DeepSeek's pricing has two levers, and most coverage looks at the smaller one. From Monday to Friday, in two windows (UTC 01:00 to 04:00 and 06:00 to 10:00), every rate doubles. Chinese public holidays are excluded. That adds up to 7 hours a day.

Two lines below those headline rates sits the larger number. On DeepSeek V4.1 Flash, input text that hits the cache costs $0.003 per million tokens. The same text on a cache miss costs $0.15, which is fifty times more.

What a cache hit actually is

A token is a small chunk of text, and providers bill by the million. A cache lets the provider recognise the start of a prompt it has already seen and charge much less for repeating it. This suits agents and chatbots that resend long instructions, big documents or a growing history on every call.

The catch is that the repeated part must be recognised as identical. The peak-hour surcharge is a 2x penalty you can dodge by rescheduling. The cache is a 50x discount you earn by how you build the request.

The worked example

OrcaRouter models a daily agent-style job where 80% of the input is a reused prefix. With caching, the daily cost comes to about $1.85. Without caching but run entirely off-peak, it is $4.20. Run entirely at peak, it is $8.40.

So caching saves $2.35 in this case, and the best scheduling saves $4.20 against the worst. Combined, the same tokens cost about 4.5x less. The suggested order of work is to shape requests for reuse first, move what you can out of peak windows second, and compare vendors last.

How the discount slips away

Caching matches on the leading tokens. If you put a timestamp, a session ID or a reshuffled tool list at the top of the system prompt, every request looks new and misses. The text looks the same to the person who wrote it, but not to the cache.

Two other traps are less obvious. Caches belong to the provider's serving infrastructure, so a router that fails over to a different provider path mid-task may lose the discount. And the name deepseek-flash now points to DeepSeek-V4.1-Flash-1017, so a cost model built on an older checkpoint describes a card that has already changed.

Where the clock still bites

The peak windows fall inside working days in Europe and Asia and mostly outside the US working day. A product serving European users will land much of its traffic at the doubled rate whether or not anyone planned for it.

• Concurrency limits: 2,500 for Flash and 500 for V4 Pro, so if everyone shifts batch jobs off-peak, you may compete for capacity.
• The windows follow China's holiday calendar, which your scheduler probably does not use.

The outlet also sells access to both models, so read its advice with that in mind. The figures come from DeepSeek's own pricing page as OrcaRouter read it.