The list price is now a range: peak hours, promo windows and host floors

For most of 2026 you could look up a model's price, write it in a spreadsheet, and trust it for a quarter. ModelPriceWatch, which tracks list prices across ten flagship models, noted in August that across 40 readings no basket constituent had ever changed its own list price. Prices moved when models were replaced, not because a vendor edited a rate card.

August broke that. Two constituents repriced in one month, and both reprices came with conditions attached. One depends on the clock. One depends on the calendar. Meanwhile the cheapest GPT-4-class price in the market is now held by a third-party host rather than the lab that trained the model. Put those three together and "list price" stopped being a single number. It became a range, and the width of the range is set by things your routing layer can control.

The clock: DeepSeek's peak and off-peak schedule

On 16 August DeepSeek moved its V4 family from flat pricing to a peak and off-peak schedule. The official pricing footnote is short: off-peak rates are half of the peak rates, and peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Every other hour is off-peak, including all 48 hours of the weekend.

The headline numbers were reported as a price increase, and at peak they are. InfoWorld's write-up put the V4 Pro rise at 51% to 203% on input and 127% to 355% on output, with cache-hit input up as much as 1,100%. V4 Pro now costs $1.32 input and $3.96 output per million at peak, $0.66 and $1.98 off-peak. V4 Flash is $0.44 and $1.32 at peak, $0.22 and $0.66 off-peak.

Two details matter more than the percentages.

First, the schedule is set in UTC around DeepSeek's home market. Peak lands in Beijing's working day, 09:00 to noon and 14:00 to 18:00 local. For a team in California or New York, the entire 9 to 5 falls outside peak. A Central European team has about 38% of its working day inside peak. An Indian team has 56%. Where your traffic originates now changes what the same request costs.

Second, this is the first frontier-tier API to sell a discount for time rather than for latency. Batch tiers give you half price if you accept a 24-hour turnaround. DeepSeek gives you half price for the same synchronous request, same latency, at 11:00 UTC instead of 09:00 UTC. That is a lever nothing else in the market has, and most teams are not pulling it because nobody built the scheduler.

The calendar: promotional windows with printed end dates

Five days after DeepSeek's schedule went live, OpenAI moved in the other direction. On 21 August it cut GPT-5.6 Sol from $5 and $30 to $4 per million input and $20 per million output, a 20% input cut and a 33% output cut. Reuters framed it as a response to competition from Anthropic and Chinese models.

The condition is in the fine print. OpenAI calls the new rate promotional and says it is "available at least through November 21, 2026." There is no published date on which it reverts, and no published price it reverts to. The Terra and Luna cuts from 30 July are permanent. The Sol cut is a floor date and an open question.

Google did the same with a harder edge. Gemini 3.8 Flash launched on 2 September at $0.75 and $3.75 per million tokens, the same introductory price as 3.7 Flash, with a footnote stating $1.50 and $7.50 from 1 January 2027. The model does not change. The bill doubles at midnight.

Smaller windows are closing faster. Inception's Mercury 2.5 Preview shipped on 31 August at an 80% discount that ends 8 September at 07:00 UTC. Z.ai's GLM-5.3-Flash promotional rate of $0.075 and $0.25 runs to 9 September, then doubles. And the one that went the other way: Anthropic had scheduled Sonnet 5 to rise from $2 and $10 to $3 and $15 on 1 September, then on 10 August made the introductory price permanent. Teams that had budgeted a 50% increase got to un-budget it.

Every one of those is a date your finance team does not have on a calendar.

The host: when a reseller sets the floor

The third shift is quieter. ModelPriceWatch tracks the cheapest model clearing a fixed GPT-4-class bar (GPQA Diamond of 70 or better). Through August that floor stayed at $0.113 blended per million, held by DeepSeek V4 Flash.

But not DeepSeek's V4 Flash. On 16 August DeepSeek raised its own first-party peak rate on V4 Flash from a blended $0.175 to $0.66. The floor did not move because DeepInfra kept serving the same weights at $0.09 input and $0.18 output, 5.9x below the maker's peak rate. As the report puts it, that is "a different kind of floor from the four before it: set not by the lab that trained the model but by a third party that hosts it, and therefore a price one vendor can withdraw."

We wrote about the quality side of this in same weights, twelve providers. The price side is now equally real. For open weight models the lab's rate card is one row in a table, not the table. A cost model that only knows "DeepSeek V4 Flash costs X" is wrong by up to 6x depending on which host the request lands on, and wrong in the other direction on the day that host changes its mind.

What this does to a cost model

A year ago the effective price of a request was a function of one variable: the model. Then cache read pricing split it into two, and we covered how far apart cache discounts sit in the price nobody compares. As of this month it is a function of at least five:

• Model. Still the biggest lever.
• Host. Same weights, up to 6x spread, and the cheapest host is the one most likely to change.
• Hour. For DeepSeek, 2x between 09:00 UTC and 11:00 UTC on a weekday.
• Date. Sol until 21 November, Gemini 3.8 Flash until 31 December, GLM-5.3-Flash until 9 September, Mercury 2.5 until 8 September.
• Cache state. Which the Fable 5.1 launch just turned into a competitive axis.

Nobody is going to maintain that by hand. It is a routing problem, and the routing layer is the only place in the stack that sees all five variables at request time.

What to do this week

Put price expiry dates next to model deprecation dates. You already track when a model ID gets retired. Track when its price changes on the same calendar: 8 September, 9 September, 21 November, 31 December. Set an alert on spend per model a week before each so a scheduled doubling shows up as a planned event.

Separate deferrable work from interactive work. Evals, synthetic data, backfills, nightly summarisation and re-indexing do not need to run at 09:00 UTC. Tag them with request metadata so you can see what share of spend is deferrable, then schedule that share into off-peak windows. For DeepSeek that is a clean 50% on the moved portion.

Keep two hosts in every open weight route. If DeepInfra is setting your floor on V4 Flash, put a second host behind it in a fallback policy and use load balancing to keep both warm. The day the floor host reprices, traffic shifts on its own instead of at the end of the month when the invoice arrives.

Compare blended cost per task, not rate cards. With five variables in play, the rate card is the least informative document in the discussion. Pull actual cost per completed task from cost tracking and make the routing decision on that. We made the longer version of this argument in cost per task, not cost per token.

Assume the next reprice has conditions. August's two reprices both did. The pattern is set: labs now price with a schedule or a promo window because it lets them move volume without committing to a permanent number. Build for a price that has a shape, not a value.

The takeaway

Frontier pricing used to be sticky and simple. In one month it became fluid and conditional: two reprices, one indexed to the hour and one to a date, plus a price floor held by a reseller. None of that is bad for buyers. Off-peak DeepSeek is cheaper than it was. Sol is cheaper than it was. The floor held. But collecting those savings now requires infrastructure that knows what time it is, what day it is, and who is serving the weights. If your routing layer cannot answer all three, the list price you are paying is the top of the range.