DeepSeek V4.1-Flash: cheap to train or cheap to use?

When an AI model is called "cheap", the claim can refer to three different bills: what it cost to train the model, what each response it generates costs, and what it ultimately costs to deliver a complete task. DeepSeek launched V4.1-Flash in September 2026, saying its new architecture lowers cost and price. In the two pages consulted — the release announcement and pricing page — we found usage prices but no monetary figure for training. Separating the three bills keeps you from comparing things that cannot be compared.

That is the question behind XMACNA's diary #8, a Portuguese-language video titled "DeepSeek 4.1: barato para treinar ou para usar?" ("cheap to train or cheap to use?"), published on YouTube on October 21, 2026. The video raises the question; this article lays out the reasoning with DeepSeek's own public figures, so you can apply the same method to any model.

Three bills that often get mixed up

Training cost. This is what the developer spends to create the model: compute, data, staff and the tuning rounds after pre-training. The vendor pays it. For anyone using the model through an API, it shows up only indirectly, built into whatever price the company decides to charge.

Inference cost. This is the cost of putting the model to work: every time it reads text and writes a response, it consumes compute. With APIs, that bill reaches the customer as a price per token. A token is a piece of text, which can be a word, a number or a punctuation mark.

Delivered cost per task. This is the bill that matters to whoever runs the operation: what it costs to resolve a customer request, summarize a contract or classify an order from start to finish. It includes tokens, but also retries, the tools the model calls, human review and the surrounding infrastructure.

The three bills are related, but one does not stand in for another. A model can be expensive to train and cheap to use, or cheap per token and expensive per task if it makes many mistakes and requires rework.

What DeepSeek published about V4.1-Flash

On its release page, dated September 10, 2026, DeepSeek describes V4.1-Flash as a model with a MoE (mixture of experts) architecture and 552 billion parameters in total. According to the company, only part of them is activated for each token: 8 billion when reading the input and 16 billion when generating the output. In the API, the model is called `deepseek-flash`.

The idea behind MoE, in simple terms, is to split the model into specialized blocks and activate only some of them at each step. That tends to reduce the compute needed to generate each response. DeepSeek says the architecture makes the model more efficient and that it is passing the savings on through lower API prices. These are the company's own claims.

One caveat matters here. Active parameters per token describe the work done on each response. They do not tell you what it cost to train the model, nor do they allow a direct comparison between training and usage.

What the price list shows

DeepSeek's models and pricing page, checked on October 21, 2026, lists DeepSeek-V4.1-Flash-1017 at these prices, in US dollars per million tokens:

Three details change the math:

• Input and output are priced differently. The text the model writes costs more than the text it reads.
• Caching makes repeated input cheaper. When part of the input has already been processed and is reused, it is billed at the cache-hit rate, which is much lower.
• Timing matters. Peak hours run from 01h to 04h and from 06h to 10h UTC, Monday to Friday, excluding Chinese public holidays. Outside those hours, the rate is half.

The same page states that model usage is charged as the number of tokens multiplied by the price, and that the company may adjust its prices. The Search and Fetch APIs, for searches and lookups, have their own price list and are billed per successful request. There is a temporary exception: when the DeepSeek API is used through DeepSeek Harness, Claude Code or Codex, the server-side tool calls to Search and Fetch are currently free, and the page says per-request billing for them starts on November 1, 2026, at 00h Beijing time (16h UTC on October 31). The exception covers only those calls, not every direct use of Search and Fetch. Check the current page before making any decision.

An illustrative calculation

Picture a volume of 1 million input tokens without cache and 1 million output tokens. At the prices above, the bill comes to US$ 0,15 + US$ 0,60 = US$ 0,75 off-peak, or US$ 0,30 + US$ 1,20 = US$ 1,50 at peak.

This arithmetic shows how the price list works; it is not an estimate of the typical cost of real usage. It leaves out infrastructure, implementation, tools, taxes and the option of running the model on your own servers, which swaps the per-token price for the cost of hardware and operations.

What about training cost?

In the two DeepSeek documents used in this article, the release page and the price list, we did not find the monetary training cost of V4.1-Flash. The release page mentions new pre-training methods and larger-scale reinforcement learning, but gives no figures. So those sources offer no sufficient public basis to say whether the model was cheap to train, or to compare that number with the price of using the API.

For a company that buys the model through an API, that gap matters less than it seems. Training cost belongs to the vendor. What lands on your invoice is inference, and what lands on your results is the cost per task.

Why price per token is not cost per task

The same price per token can produce very different bills, because each task uses tokens differently. Consider two hypothetical tasks on V4.1-Flash, at the off-peak rate, with a thousand runs of each:

Task B uses about four times fewer tokens than task A and still costs more than three quarters as much, because output is the most expensive line on the price list. Task A, meanwhile, becomes about five times cheaper when the repeated input comes from the cache. The volumes are made up to show the mechanism; only the prices come from DeepSeek.

Beyond the price list, other factors move the bill:

• Context size. The more text the model has to read on each request, the more expensive each request becomes. AWS makes this point in a February 2026 technical article about its own inference service; on DeepSeek's price list, it shows up as more input tokens billed.
• Urgency. Not every task needs an immediate answer. When it launched a batch-processing API in April 2025, Cloudflare gave the example of summaries of many documents that can wait. With DeepSeek, the equivalent is moving flexible work outside peak hours, when the rate drops by half. That does not mean batch processing is always cheaper with every provider.
• Who runs the infrastructure. Using a managed API and hosting the model yourself are different bills: with one, you pay per token; with the other, for servers, energy and staff. Google Cloud describes these two options, for its own products, in a February 2025 post. In a March 2025 guide, the company says its managed service can save hours of management and avoid paying for idle GPUs, while virtual machines with direct access to the hardware give full control; the guide itself warns that not every tip fits every use case.
• Switching models. Moving to a cheaper model has a cost too. In an MIT Sloan article from January 2026, one of the authors of a study on open models notes that companies build processes around the model they already use and that switching is not trivial.

Delivered cost per task: the bill that matters

The price per token is only the starting point. A practical way to get to cost per task:

1. Pick a real, repetitive task, such as answering a customer question or extracting data from a document.
2. Measure the tokens in a sample: how much goes in, how much comes out and how much can be served from the cache.
3. Count retries and failures. A wrong answer that must be redone adds cost; record the tokens, tools and review needed for each retry.
4. Add everything around it: tools called, human review time, integration with your systems.
5. Divide by the number of tasks completed to standard, not by the number of calls.

This is where models that are cheap per token can win or lose. A model that costs less per million tokens but gets fewer answers right can end up more expensive per resolved task. A fair comparison uses identical tasks and criteria defined in advance, as we show in Blind AI testing: choose models with evidence. The same reasoning about operating cost appears in GPT-5.6 and the operational cost of AI.

V4.1-Flash is also part of a wider contest between open and closed models, the subject of The new race of open models has shifted power in AI. The release page points to a model page on Hugging Face and says DeepSeek will work with the open-source community on inference support.

Three questions before choosing a model on price

• Which bill is being advertised? An API price is a usage cost, not a training cost.
• What does your usage look like? Lots of repeated input favors the cache; long responses weigh on the output line; tasks that can wait benefit from off-peak hours.
• What does a completed task cost? Only that number allows an honest comparison between models.

For an overview of what goes into a company's AI budget, see the guide to how much AI costs a company.

Watch the diary

Diary #8, "DeepSeek 4.1: barato para treinar ou para usar?" (YouTube, ID `lToEm9nslIU`), is on the XMACNA YouTube channel, in Portuguese. The video poses the question; this article details the method and the sources.

Frequently asked questions

Does DeepSeek's API price show what it cost to train the model?

No. The price list shows the cost of usage, billed per token. In the DeepSeek documents used here, the release page and the price list, we did not find the monetary training cost of V4.1-Flash, so those sources offer no public basis for comparing the two.

What does it mean to have 552 billion parameters but only 8 or 16 billion active?

According to DeepSeek, V4.1-Flash uses a MoE architecture: the model has 552 billion parameters in total but activates 8 billion per token on input and 16 billion on output. That describes the compute behind each response, not the training cost.

What do 1 million uncached input tokens plus 1 million output tokens cost with V4.1-Flash?

Based on the price list checked on October 21, 2026, 1 million input tokens without cache plus 1 million output tokens add up to US$ 0,75 off-peak or US$ 1,50 at peak. It is an illustrative calculation: it does not include infrastructure, implementation, tools or taxes, and prices may change.

How can I compare the cost of two AI models for my company?

Run both on the same real task, with the same quality criteria, and divide the total spend by the number of tasks completed correctly. The price per token alone does not show how much the task reads and writes, how much it draws from the cache, or the rework, tools and human review involved.

Next step

If your company wants to understand what it would cost to put AI to work on a real process, XMACNA's free diagnosis suggests a starting point based on your answers to the form.