What peak pricing for AI APIs is — the same prompt, twice the cost
Peak pricing for AI APIs means the per-token rate depends on when a request is sent — a congested peak window costs more than a quiet off-peak one, typically double — and it exists because GPU compute cannot be stockpiled for later while demand concentrates into a few hours of each day
The three lines
- Definition — the same model and the same prompt cost different amounts depending on the hour. Off-peak is typically exactly half of peak
- Why — GPU compute is not storable. Idle hours vanish; busy hours queue. It is the structure electricity and telecom tariffs have used for decades
- In practice — keep requests a human is waiting on at peak rates, and move only batch work off-peak. Windows are usually set in UTC, so their local impact depends on your time zone
Key questions
- What is peak pricing for an AI API?
- A pricing structure where the per-token rate depends on the hour a request is sent. Send the same prompt to the same model during a congested window and it costs more; send it during a quiet window and it costs less. In deployed cases, off-peak has been set at exactly half of peak. It is the same structure as time-of-use electricity tariffs, overnight telecom discounts or off-season airfares, applied to AI inference.
- Why charge differently by time of day?
- Because GPU compute cannot be held as inventory. An idle GPU-hour at 3am simply disappears; there is no way to save it for the afternoon crush. Meanwhile API demand tracks human working hours and concentrates into a few hours a day. A provider therefore has two options: buy enough GPUs to cover the peak and leave them idle most of the time, or use price to spread demand out. Peak pricing is the second option.
- How should a team respond?
- Start by splitting requests in two. Requests where a person is waiting at a screen — chat responses, search, code completion, live translation — should stay at peak rates, because latency is quality. Requests where nobody is waiting — overnight document indexing, bulk classification and summarisation, evaluation and regression runs, data cleaning — move off-peak. That split alone usually shifts a large share of token volume. One trap: peak windows are typically published in UTC, so convert them to your own local hours before assuming you are safe.
An AI API price sheet had two columns for a long time. Input costs this, output costs that.
In 2026 a third column started appearing. What hour did you send it?
Same model, same prompt, same answer. Twice the price.
1. The definition
| Axis | Old pricing | Peak pricing |
|---|---|---|
| Model | Varies | Varies |
| Input / output | Varies | Varies |
| Cache hit or miss | Varies | Varies |
| Time of request | Irrelevant | Varies |
One axis was added, nothing more. In deployed cases off-peak is set at exactly half of peak.
The idea is old. Time-of-use electricity tariffs, overnight telecom discounts, peak and off-season airfares, early-morning transit fares — pricing a congested resource by the clock is a well-worn tool. It has now arrived in AI inference.
2. Why it appears: inventory you cannot keep
The core is one sentence. GPU compute cannot be stockpiled.
| Resource | Can surplus be stored? |
|---|---|
| Chips, clothing, physical goods | Yes — put it in a warehouse |
| Electricity | Barely — storage at scale is expensive |
| GPU compute-hours | No — they simply vanish |
An idle GPU-hour at 3am is gone. There is no mechanism to bank it for the afternoon.
And demand is not flat. API calls follow human working hours. That leaves a provider with two options.
| Option | Consequence |
|---|---|
| Buy enough GPUs to cover peak | Hardware sits idle most of the day — costs rise |
| Use price to spread demand | More requests served on the same hardware |
Peak pricing is the second. When DeepSeek introduced the structure in August 2026, its stated reason was to "allocate resources more reasonably" and to have developers "schedule their tasks based on actual usage."
3. Increase, or redistribution? How to tell
This is where confusion sets in. Introducing peak pricing can carry a price increase and a demand-shifting incentive at once, and in practice they arrive mixed.
There is one way to separate them. Compare the off-peak rate to the old flat rate.
| Case | Off-peak vs old rate | Reading |
|---|---|---|
| A | Off-peak = old rate | Pure redistribution (peak surcharge only) |
| B | Off-peak < old rate | Effective cut for anyone who can shift |
| C | Off-peak > old rate | An increase, with redistribution layered on |
Put DeepSeek's August 2026 change into that table. V4-Flash output was $0.28 per million tokens flat; the new off-peak rate is $0.66.
Even moving everything off-peak costs 2.4 times the old rate. That is case C — an increase first, with a shifting incentive on top. The specific figures are in "DeepSeek API prices rise August 16."
This is why "off-peak is half price" is not, by itself, information. You have to know where the baseline moved.
4. Practice: what moves and what stays
Split requests in two. The test is whether a person is waiting.
| Type | Examples | Decision |
|---|---|---|
| Someone is waiting | Chat responses, search, code completion, live translation | Stay at peak — latency is quality |
| Nobody is waiting | Overnight indexing, bulk classification and summarisation, evaluation and regression runs, data cleaning, log analysis | Move off-peak |
In many organisations the second category dominates token spend. The invisible batch pipeline usually burns far more tokens than the user-facing chat. So this split alone tends to free up a large share of volume.
The arithmetic, for a team emitting 10 million output tokens a month with 60% of it batch work:
| Scenario | Volume | Rate | Cost |
|---|---|---|---|
| All at peak | 10M | peak | peak × 10M |
| Move 60% off-peak | 6M | 50% of peak | 30% saved |
| 4M | peak | — | |
| Total | — | — | ~70% of the original |
About 30% saved, with no code change — only a change in when jobs run. (This is arithmetic on published rates, not a measured case study.)
5. Three traps
① The windows are in UTC. Peak hours are typically published in universal time, so their local effect depends entirely on where your team sits. Windows of 01:00-04:00 and 06:00-10:00 UTC land at 10:00-13:00 and 15:00-19:00 in Seoul — most of a working day — while falling largely outside business hours on the US east coast. The same price change costs two teams very different amounts. Convert before you assume.
② Retries cross the boundary. A large job started near the end of an off-peak window that fails and retries can have its retries billed at peak. Schedule batch runs with headroom before the boundary.
③ Cache-hit rates move independently. Cache-hit input pricing can be adjusted sharply and separately from the peak/off-peak split. The more your architecture repeats the same prompt — conversational products, looping agents — the more that line matters. In practice the cheapest line on a price sheet has taken the largest increase.
6. What to watch
Whether peak pricing becomes an industry standard or stays a one- or two-provider measure is not settled. But the underlying conditions do not change quickly: GPU supply does not expand on short notice, and demand clusters into working hours.
Providers moved in opposite directions in the same period. One opened its newest model and unlimited text chats to free users; another halved a model's price. That divergence continues the competitive picture described in "A new AI model every two weeks" and "Five in eight weeks: China's model offensive."
The base structure of token pricing — why output costs more than input, what a cache hit is — is in "What AI token prices are."
7. What is not established
No complete list of adopters. The concrete case documented here is DeepSeek.
Time-zone basis varies. Whether windows are a single UTC standard for all customers differs by provider and was not confirmed.
Stacking with batch discounts is unconfirmed. Providers running separate batch-API discounts may or may not combine them with off-peak rates.
Sources
- InfoWorld — DeepSeek raises some V4 prices by more than 10x as AI demand strains capacity
- Engadget — DeepSeek's AI models are about to cost four times more
- Quartz — DeepSeek raising API prices by up to 1,100% starting Aug. 16
- Computerworld — DeepSeek raises some V4 prices by more than 10x as AI demand strains capacity
- South China Morning Post — DeepSeek signals 'significant' price hike amid surge in demand for low-cost AI models