Electricity has been priced this way for decades. Run the dishwasher at night, pay less, because the grid is quieter and the capacity is sitting there unused.
Nobody thought to do it with AI until recently, and DeepSeek has now done it in the most literal way available. Two short windows of the day cost double. Every other hour costs half. Same model, same quality, same everything, and the only variable is what time it is in London.
It is a small idea with a surprisingly large consequence for a European business, because the expensive hours land almost entirely outside a European working day. Most of your usage would already be in the cheap window without you doing anything clever at all.
What actually shipped
DeepSeek released V4.1-Flash on 10 September 2026. It is a 552 billion parameter mixture-of-experts model, activating roughly 8 billion parameters for input and 16 billion for output, with a context window of one million tokens.
Those architecture numbers matter less than they look. The activated parameter count is what governs cost and speed, and activating 8 to 16 billion out of 552 billion is what allows a large model to be served at these prices. It is the same architectural approach behind most current frontier models, applied aggressively.
The weights are published on Hugging Face under an MIT licence, which is about as permissive as open licensing gets. That places it alongside the other genuinely open releases we have covered this year, including Kimi K3 and GLM-5.2, and it matters for reasons that have little to do with whether you ever download anything.
How the clock pricing works
The mechanics are simple enough to memorise, which is unusual for AI pricing.
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. During those windows you pay double. Every other hour of every day, including all weekend, is off-peak, and off-peak is the rate that matters: $0.15 per million input tokens on a cache miss, $0.60 per million output tokens, and $0.003 per million input tokens on a cache hit.
To put the off-peak rate in context, GPT-6 Astra costs $10 per million input and $50 per million output. Claude Opus 5 sits at $5 and $25. DeepSeek off-peak is $0.15 and $0.60. That is not a competitive discount, it is a different order of magnitude, and the obvious caveat applies: this is not a frontier model and it is not trying to be. It is a fast, capable, extremely cheap model for work that does not need frontier reasoning.
The peak windows are also genuinely short. Seven hours a day, five days a week, is 35 hours out of a 168 hour week, so roughly 79% of all hours are off-peak before anyone schedules anything. This is not a pricing scheme that demands clever engineering to benefit from.
Why this favours European businesses
Work out where those peak windows land in your own time zone and the picture gets better.
In Central European Summer Time, which is UTC+2, the peak windows are 03:00 to 06:00 and 08:00 to 12:00. So the expensive hours cover the small hours of the morning, when almost nothing is running, and then the stretch from eight in the morning until noon. That second block is real working time and it is the only part of a European day that costs double.
Everything from midday onward is off-peak. Your entire afternoon, your evening, overnight batch work, and the whole weekend all sit at the lower rate. If your business is customer-facing with traffic spread across a normal day, the majority of your volume is already landing in the cheap window with no action required from you.
And the one expensive block is the easiest to plan around. Batch work is the obvious candidate: document processing, bulk classification, report generation, overnight enrichment jobs. Anything that does not need to happen at a specific moment can simply be scheduled after midday, which is a scheduling change rather than an engineering project. Moving a nightly job from 09:00 UTC to 13:00 UTC halves its cost and nobody notices.
The cache rate is the real number
Buried in the pricing table is a figure that deserves more attention than the time-of-day mechanic, and it tells you where this industry is genuinely heading.
A cache hit costs $0.003 per million input tokens. A cache miss costs $0.15. That is a fifty-fold difference for sending the same text, and the only distinction is whether the model has seen that exact content recently enough to have it stored.
We went through what caching actually is in the Claude Fable 5.1 article, and the short version is that most business AI requests contain a large block of identical context every time: your instructions, your policies, your product information. Only the specific question changes. Caching stores the repeated part so you are not billed full price to resend it.
At a fifty-fold difference, anything stable across your requests is effectively free to include. A system prompt, a knowledge snippet, a tool schema, your company voice guidelines: all of it costs almost nothing once cached. That changes how you should design prompts, because the instinct to keep context minimal for cost reasons is largely obsolete for the repeated portion. Keep the stable part rich and the variable part tight.
What MIT licensing actually buys you
The MIT licence on the weights is the sort of detail that gets mentioned in passing and matters more than the mention suggests, even though almost no small business should self-host anything.
The direct benefit is that you could run this model on your own hardware, and for the overwhelming majority of readers that is a theoretical benefit rather than a practical one. The infrastructure required to serve a 552 billion parameter model is not something a nine-person company should be maintaining, and the arithmetic almost never works against $0.15 per million tokens.
The real benefit is structural. Published weights mean multiple providers can host the same model, which creates genuine price competition and, more importantly, an exit. A workflow built on a hosted proprietary model has exactly one supplier who can change the terms whenever they choose. A workflow built on an open-weight model has several, and that difference only becomes visible on the day a vendor sends an awkward email.
It also means the model cannot be retired out from under you. Proprietary models get deprecated on the provider's schedule, as anyone who built on DALL-E discovered this summer. Open weights that exist on Hugging Face today will still exist in three years regardless of what the company that made them decides to do.
The honest caveats
Three things worth being straight about before anyone moves a production workload.
This is not a frontier model and it should not be used as one. For work requiring genuine reasoning depth, nuanced judgement, or handling of situations the prompt did not anticipate, the expensive models remain expensive for a reason. The right mental model is a fast, cheap, capable workhorse for high-volume structured work, which is most of what a small business actually needs, rather than a replacement for your best model on your hardest task.
Data governance deserves a proper look rather than a shrug. DeepSeek is a Chinese company, and where your data is processed and under what terms is a question with real weight for an EU business under GDPR. That is not an argument against using it, and it is an argument for reading the terms before customer information touches it. Using it for internal document processing is a different decision from using it on your customer database, and the second one needs an answer you can defend.
And time-of-day pricing adds a variable to your forecasting that most businesses have never modelled. Your bill now depends partly on when your customers happen to contact you, which is not something you control. That is a minor complication at small scale and worth knowing about before it surprises you, in the same way the Notion credit model surprised teams who had only ever budgeted for per-seat software.