Pricing

Since 16:00 UTC on 2026-08-16, what a DeepSeek token costs depends on the hour you spend it in. This page carries the schedule, both rate cards, and a strip that reads your clock – the same data the CLI's estimates switch on, so the page and ds pricing can never disagree.

Right now

The schedule

One table drives everything on this page: each row is a billing period, its daily window in UTC minutes, the multiplier on that era's base card, and the instant the row takes effect. The strip above, the CLI's estimates and ds pricing all read the same rows.

The windows are defined in UTC, not Beijing – in Beijing time (UTC+8) the peak hours read 09:00–12:00 and 14:00–18:00, which is the Chinese working day. The off-peak window covers the whole European and American working day, so batch work that can move west should.

The card

USD per 1M tokens, in force since 2026-08-16 16:00 UTC. Off-peak is half of peak by construction, but on a higher base than the flat card it replaced – even the cheapest hour now costs more than any hour did before. A flash cache-miss input token went from $0.14 to $0.22 per 1M off-peak, and to $0.44 in peak hours:

The context cache stays the biggest lever: a cached input token costs about 1/30th of a miss, against 1/50th (flash) and 1/120th (pro) on the card this replaced. Prompt structure still dominates a bill – it is worth up to 30×, where the hour of the day is worth 2×.

What it replaced

The flat card published 2026-08-02, billed at every hour until the switch. It is here because the ledger keeps token counts rather than dollars, so a call made before 2026-08-16 16:00 UTC still reprices under the card it was actually billed at:

Peak output on pro is 4.6× that card; peak cached input is 12× it. The published increase is steepest exactly where the old card was cheapest.

The same answer in the terminal

ds pricing            # period right now, local + UTC + Beijing, both cards
ds pricing --json     # the same, as JSON for scripts

Computed locally from the same schedule – no network call, nothing spent. The cost estimates price each call with the card in force at the moment it was made, which is what lets one ledger span both eras. The ledger stores exact token counts rather than dollar amounts, so historical calls can always be repriced under whatever card was, or becomes, real.

Source

The numbers are DeepSeek's own, from the official Models & Pricing page, which dropped the flat card and now shows only these two periods. The CLI carries the same page offline – ds docs show quick_start/pricing – and ds docs sync refreshes it, so the terminal copy, this page and the estimates trace to one upstream.

They are also checked against the bill. A 188,542-token cache-miss call to pro, made off-peak on 2026-08-17, settled at 0.84 CNY – 4.46 CNY per 1M against the published off-peak 4.5, where the old flat card would have made it 3.0. The page and the billing system agree.