DeepSeek V4 Pro Peak & Off-Peak Pricing: Hours, Rates & Strategy

On August 16, 2026, DeepSeek replaced flat pricing with a two-tier hourly schedule across the whole V4 lineup: **peak hours cost 2x the off-peak rate**, and off-peak covers the majority of the day^[1][2].

The official announcement frames the change as flexibility: 'Off-peak rates are 50% lower than peak, enabling more flexible workload scheduling.'^[1] The new rates took effect at 16:00 UTC on August 16 — three days after the V4 Pro GA on August 13.

The important nuance: off-peak is half of the *new* peak price, not half of the old flat price. Every V4 Pro line is still more expensive than the pre-August-16 flat rate at both tiers^[2][7].

• Peak rate = 2x off-peak rate
• Off-peak covers the majority of the day (see exact hours below)
• Applies to deepseek-v4-pro and deepseek-v4-flash
• Took effect 16:00 UTC, Aug 16, 2026

Per the official pricing page, peak hours are **01:00-04:00 and 06:00-10:00 UTC**; everything else is off-peak^[2][7].

Peak hours total 7 of 24 (roughly 29% of the day); off-peak covers 17 hours. The two peak windows are both anchored to Asia business hours, which makes sense for DeepSeek's primary demand base.

If your traffic is US-centric, most of the US workday (10:00-01:00 UTC) is off-peak — a genuinely useful fact for cost models.

The per-1M-token rates for deepseek-v4-pro at each tier^[2].

For comparison, V4 Flash output is $0.66 off-peak / $1.32 peak, and Flash input on a cache miss is $0.22 / $0.44. Pro is roughly 3x Flash at every line^[2].

A concrete example: 10M output tokens of V4 Pro costs $19.80 off-peak vs $39.60 peak. On a heavy agent workload that runs around the clock, the peak windows add up quickly — the scheduling lever is worth roughly 2x on every token moved.

The playbook for exploiting the cheap hours without hurting your product^[2][7].

The win is real but bounded: off-peak is 50% of peak, not 10%. The bigger lever remains the cache-hit rate on input and `reasoning_effort` on output — see the reasoning effort guide for the output-side control.

For always-on interactive traffic you cannot defer, the hourly tier is a background cost factor; for anything batch-able, it is a 2x discount you should take.

• Batch jobs: nightly reports, eval runs, dataset enrichment → run between 04:00-06:00 or 10:00-16:00 UTC
• Retries & re-generation: queue model retries to the next off-peak window
• Pre-generation: generate content ahead of peak demand and cache it
• Cache strategy: on peak hours, lean harder on cache hits ($0.044 vs $1.32 input)
• Regional routing: if you have users in Asia, their peak is your peak — schedule accordingly

Practical notes for implementing hour-aware routing^[2][3].

• Use UTC — the schedule is defined in UTC, and your local tz math will drift with DST
• Billing is tokens × rate; reasoning tokens bill at output rates, so peak + max effort is the most expensive combination
• The pricing page footnote: prices can be adjusted — pin a rate check into your cost monitor
• OpenRouter mirrors DeepSeek rates but may lag the hourly tier; verify before relying on router prices for peak scheduling[10]

# Pseudo: pick the cheap tier for batch work
from datetime import datetime, timezone

def is_peak(now_utc):
    h = now_utc.hour
    return (1 <= h < 4) or (6 <= h < 10)

def schedule_batch(job):
    while is_peak(datetime.now(timezone.utc)):
        sleep(60)
    run(job)