What happened on 13 August
DeepSeek repriced its API together with the general availability of V4-Pro. The announcement says "Off-peak rates are 50% lower than peak, enabling more flexible workload scheduling", and the new prices take effect at 16:00 UTC, Aug 16, 2026. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; every other hour is off-peak.
The new price list (USD / 1 million tokens):
| Model | Tier | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|---|
| DeepSeek-V4-Flash | off-peak | $0.007 | $0.22 | $0.66 |
| DeepSeek-V4-Flash | peak | $0.014 | $0.44 | $1.32 |
| DeepSeek-V4-Pro | off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek-V4-Pro | peak | $0.044 | $1.32 | $3.96 |
On the same day OpenAI announced a new tier of its own: Ultrafast mode, by its own description "up to 14x faster than Standard processing", for now "available in limited preview to select customers". No price was published. The two announcements are the two ends of the same phenomenon: the price of a token is no longer one number but a table, and time and patience are line items in it.
DeepSeek was not the first. Alibaba Model Studio has been giving a night-time discount on the Qwen3.7 series since 23 June 2026: according to the price list, between 22:00 and 08:00 Beijing time the Plus model runs at a 60% and the Max model at an 80% discount, weekends and public holidays included.
hour (UTC) 0 6 12 18 23
DeepSeek .###..####..............
Qwen ..............nnnnnnnnnn
# = peak (double) n = night (-60%)
CEST = UTC + 2 hours (in summer)
This is not a price rise, it is a repricing
On DeepSeek's April price list, the one without tiers, a million output tokens of V4-Pro cost $3.48 and of V4-Flash $0.28. With the new prices:
- V4-Pro is 43% cheaper off-peak ($3.48 → $1.98) and 14% more expensive at peak ($3.96);
- V4-Flash gets more expensive either way: 2.4× off-peak, 4.7× at peak ($0.28 → $0.66 / $1.32).
So the big model got cheaper, the small one more expensive, and both are now tied to the clock. If you have been running everything on the small model because that was the penny option, this move is not a discount, it is a bill.
At Western providers the clock does not matter, patience does
There the tier is set not by the time of day but by how long you are willing to wait. The commitments come from their own documentation, as of August 2026:
| Provider | Fast tier | Cheap but synchronous | Batch |
|---|---|---|---|
| OpenAI | Fast mode: 2× price, up to 2.5× faster on GPT-5.6 Sol (it replaced Priority Processing on 30 July) | Flex: at batch price, a 10–15 minute timeout recommended | −50%, within 24 hours, "and often more quickly" |
| Anthropic | — | — | −50%, 24-hour expiry, "most finish in less than 1 hour" |
| Google Gemini | Priority: +75–100%, served at standard price when capacity is short | Flex: −50%, 1–15 minute target, no automatic fallback when it drops out | −50%, 24-hour target |
| AWS Bedrock | Priority: +75%, up to 25% better output speed | Flex: −50%, longer latency, served after standard | −50% |
| Microsoft Foundry | Provisioned (reserved capacity, in PTUs) | — | Global Batch: −50%, 24-hour target, no real-time SLA |
| SpaceXAI (Grok) | Priority: 2× price, on every token type | — | −20%, and only on a few models |
| Mistral | — | — | −50%, no deadline published |
| Alibaba | — | — | −50%, the waiting window is adjustable between 1 and 14 days |
Two things come out of this. One: batch is not a matter of negotiation, it is a standard — seven providers out of eight sit at exactly 50%, and the outlier is xAI with its 20%. Two: the fast tier costs more than the speed it buys — double the price gets you 2–2.5× the speed, which only pays off where a human is sitting in front of the screen.
What this is in money
Take a job of realistic size: 2 million input and 1 million output tokens a month — roughly what it takes to generate the descriptions for a 500-product webshop, plus summarising the incoming customer messages. The same work, in five tiers:
| Where and how | Monthly cost |
|---|---|
| OpenAI GPT-5.6 Terra, Fast mode | $32 |
| OpenAI GPT-5.6 Terra, standard | $16 |
| OpenAI GPT-5.6 Terra, batch | $8 |
| DeepSeek V4-Pro, at peak | $6.60 |
| DeepSeek V4-Pro, off-peak | $3.30 |
Between the two extremes there is almost a tenfold difference — for the same task. And the decision is not about which model is better: it is about when and with how much patience you ask.
On top of this comes the prompt cache, which is a multiplier of its own. At Anthropic a token read from the cache costs a tenth of the input price, writing costs 1.25× (five-minute) or 2× (one-hour) — and the documentation states that "The pricing discounts from prompt caching and Message Batches can stack." If the same system prompt goes out ten thousand times, that is a bigger item than the time of day.
Four traps
1. A batch can expire, and you pay anyway. After 24 hours OpenAI puts the batch into expired state: the unfinished requests are lost, but you pay for the responses already produced. At Anthropic the expiry is 24 hours as well. If you submit a big batch at midnight and count on the full result in the morning, you need a retry branch.
2. Flex can drop out, and it does not fall back on its own. According to Google's Flex documentation, standard traffic can crowd flex out at peak, and then you get a 503 or a 429 — and there is no automatic fallback to standard. You have to handle that yourself, otherwise the "cheaper tier" sometimes simply does not answer.
3. A tier boundary reprices the whole request. Both xAI and Alibaba spell it out: if the prompt crosses the length threshold, every token is billed at the more expensive rate, not just the part above the threshold. One token more of input can double the price of the request.
4. Off-peak is not your night. DeepSeek's peak hours in Central European summer time are 03:00–06:00 and 08:00–12:00 — meaning the start of the European working day falls into the expensive band, while the afternoon and the evening are cheap. Alibaba's night window falls roughly between 16:00 and 02:00 in the same zone. If you plan on a "I'll schedule it for the night" basis, it is easy to run straight into the most expensive hours.
What to do now
- List what runs on an immediate response today that could wait. Report generation, product descriptions, document summarisation, tagging, embedding recomputation, translation rounds, evaluation — all of these survive 24 hours. Whatever a human is sitting in front of stays on the standard or the fast tier.
- Look at where the biggest repetition in your prompts is. The cacheable part (system prompt, product list, style guide) can be 90% cheaper — this is often a bigger item than switching tiers, and it needs no scheduling at all.
- Measure the actual cost per request, not the total on the invoice. Most providers return the tokens used and in several places the price of the request too. Without that you will not know whether the saving actually happened.
- If you use time-based pricing, put the tier boundary in your calendar. Knowing that "it is cheaper at night" is not enough: you need to know what hour that is in UTC, and what hour that is where you are.
If model choice is itself a question, AI pricing decoded and fast, cheap or smart deal with that separately.
Common misconceptions
- "Batch is half price everywhere." At seven providers out of eight, yes. At xAI it is 20%, and only on a few models — nothing at all on image and video generation.
- "Flex is the same as batch, only faster." No. Flex is a synchronous call, so your code waits for it — but it errors out when capacity is short, and with Google it does not fall back to the more expensive tier by itself.
- "A cheaper tier means a weaker model." No. Same model, same output; the difference is the order of service and the latency.
- "The night-time discount is at night for me too." The time windows are published in UTC or in Beijing time. Convert them before you schedule.
- "The cache discount and the batch discount rule each other out." Anthropic's documentation says explicitly that they stack — although under batch a cache hit is only best-effort.
What to take away
In the summer of 2026 the price of a token stopped being a single number. Today they price along at least four dimensions: which model, how long the request is, how long you wait for it, and lately when you send it. That is bad news for anyone who would like to type one number into their calculation — and good news for anyone willing to separate their jobs out. The practical minimum: know what can wait on your side, and put it in batch. That is a single switch, it takes half off the total, and it needs no change of model or of provider.