Quick Answer
Starting 2026-08-16 16:00 UTC, DeepSeek moves the V4 family from flat to peak/off-peak pricing: peak hours cost more, off-peak is half of peak. The biggest hit to heavy users is cache-hit input rising ~12x at peak — the very thing that made long agent sessions cheap.
If you don't want to watch the clock, TeamoRouter keeps flat pricing — deepseek-v4-flash $0.14/$0.28, deepseek-v4-pro $1.74/$3.48, no peak windows — plus 200 free requests a day.
What the Peak/Off-Peak Mechanism Is
DeepSeek splits the day into peak and off-peak:
- Peak windows: 01:00–04:00 UTC and 06:00–10:00 UTC (Beijing-time daytime working hours).
- Off-peak: the rest, priced at roughly half of peak.
In other words, the hours you're most likely to use the model cost double; saving money means moving heavy loads to off-peak. That's fine for manual calls, but hostile to "all-day" agent workloads.
How Much It Rose: V4 Pro per 1M Tokens
Published figures (per 1M tokens, deepseek-v4-pro):
| Item | Old flat | New off-peak | New peak |
|---|---|---|---|
| Input, cache hit | $0.003625 | $0.022 | $0.044 (~12.1×) |
| Input, cache miss | $0.435 | $0.66 | $1.32 |
| Output | $0.87 | $1.98 | $3.96 |
deepseek-v4-flash output likewise moves from $0.28 to $0.66 / $1.32 (peak/off-peak).
Two changes stand out:
- Cache-hit input up ~12x — the most relied-on saving in long agent sessions. A stable system prompt across a long session was nearly free on cache hits; at peak, that cost jumps an order of magnitude.
- Output doubles at peak — agents produce big diffs, tests, and reports, exactly the output-heavy cases now doubled at peak.
What a Month Looks Like
Example: 30M input (80% cache-hit) + 3M output per month on V4 Pro:
| Scenario | Monthly cost |
|---|---|
| Old flat | ≈ $5.31 |
| New off-peak | ≈ $10.43 |
| New peak | ≈ $20.86 |
Same usage, nearly 4x the old price at peak. For an individual that's a few to a dozen dollars; for a team running many agents all day, it's hundreds to thousands.
The Flat Alternative: TeamoRouter
TeamoRouter stays flat, no peak windows — no clock-watching:
| Model | Input | Output | Free tier |
|---|---|---|---|
deepseek-v4-flash |
$0.14 | $0.28 | 200/day (-free) |
deepseek-v4-pro |
$1.74 | $3.48 | 200/day (-free) |
One OpenAI-compatible endpoint:
Base URL : https://api.teamorouter.com/v1
API Key : sk-teamo-your-key
Model ID: deepseek-v4-pro-free # free tier; paid uses deepseek-v4-pro
Add deepseek-v4-pro-free / deepseek-v4-flash-free at 200 requests each per day, and light users' price-change anxiety is largely absorbed by the free quota.
FAQ
Does peak/off-peak affect occasional users?
Little. It mainly hurts heavy all-day agents and bulk output; the absolute extra cost of occasional calls is small.
Is the free tier affected by peak/off-peak?
No. The free tier counts requests (200/day), not peak/off-peak tokens — the price change doesn't touch it.
How should I choose between paid options?
Want predictable cost and no clock-watching: TeamoRouter flat. Prefer official list pricing and can accept peak/off-peak: go official. The model IDs differ — don't mix them up.
Will prices change again?
Pricing moves. Check the live official pricing page and TeamoRouter pricing.
Can I dodge the 12x cache-hit rise?
Move heavy loads off-peak, or use a flat endpoint. For long agent sessions, an endpoint where cache hits stay cheap is especially valuable.
See the flat plan at TeamoRouter and blunt the impact of the change.