1. DeepSeek V4.1 Flash arrives with open weights and low rates
DeepSeek released its V4.1 Flash model on September 10, 2026, and followed it on October 17 with an update labeled V4.1-Flash-1017. The open-weight system lists at $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours, and DeepSeek's own benchmarks put it close to Anthropic's Claude Opus 5 on coding. Those benchmark claims have not been independently verified.
The release matters less for any single score than for what it implies about pricing. A model that is competitive on software engineering tasks, free to download under a permissive license and cheap to call through an API gives buyers of AI services a new point of comparison. This article walks through what the company shipped, how its prices are structured, where the model falls short, and what the market may do in response.
2. What did DeepSeek ship with V4.1 Flash?
DeepSeek-V4.1-Flash is a large language model built on a sparse mixture-of-experts design. Its main body holds 552 billion parameters, but only a fraction is switched on for any given request: roughly 8 billion parameters are active when the model reads input, and roughly 16 billion when it generates output. Other published descriptions of the release count additional components on top of that main body, so the total size of the download depends on how it is measured.
The context window is one million tokens. In practical terms, that is enough room to hold a large codebase or a stack of lengthy documents in a single prompt, which is the kind of workload where coding assistants and automated agents tend to run into limits. The model also processes images natively, so a developer can pass screenshots or diagrams alongside text without routing them through a separate vision system.
The weights are published on Hugging Face under the MIT license, one of the most permissive licenses in common use. That means companies can download the model, run it on their own hardware, modify it and build commercial products on it without paying DeepSeek anything. Customers who prefer not to manage their own infrastructure can instead use DeepSeek's hosted API under the same model name.
V4.1-Flash succeeds DeepSeek-V4-Pro, which the company has largely retired since mid-September 2026. Replacing a flagship with a model that carries the cheaper "Flash" label is unusual, and it signals that DeepSeek is betting on efficiency and price rather than on maximum capability at any cost.
3. What did the October 17 update change?
The V4.1-Flash-1017 update went out on October 17, 2026, under the same API. According to the account of the release, it improved performance metrics and shortened the time the model needs to finish tasks. One widely shared post on X claimed that the updated model runs at more than 237 tokens per second and costs 90% less than rivals. Those figures come from a single social media post and have not been confirmed, so they should be treated as a claim rather than a measured result.
Speed matters because it shapes how a model feels in use and how much it costs to operate. An agent that makes dozens of calls to complete one job accumulates both latency and expense with each step, so even modest improvements in throughput compound across a long task. Whether the update delivers a sustained gain on real workloads is something outside testers will need to establish.
4. How much does DeepSeek V4.1 Flash cost to use?
At off-peak hours, DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens. During peak hours those rates double, so the headline price applies only to customers who can shift their usage to quieter periods. A team that runs batch jobs overnight can capture the lower rate, while one that serves live users during the working day will pay closer to the higher one.
The company also discounts inputs the system has already processed. When a prompt reuses content held in the cache, the price can fall to between $0.003 and $0.006 per million tokens. That is roughly 30 times cheaper than certain frontier closed models, according to the comparison drawn at the time of the release, though the specific competing products behind that figure were not named. Caching matters most for agents and coding tools, which tend to resend the same long instructions or files with every request.
Because the weights are open, there is a second route to low costs. A company with its own graphics processors can run the model without paying per-token fees at all, accepting instead the cost of hardware, power and engineering time. For high-volume users, that trade can be attractive, and it puts a ceiling on what hosted providers can reasonably charge for similar capability.
5. How does it compare with Claude Opus 5 and GPT-5.6 Sol?
On DeepSWE v1.1, a software engineering benchmark, V4.1-Flash reportedly scored 74.2. Anthropic's Claude Opus 5 scored 74.0 and OpenAI's GPT-5.6 Sol scored 73.0. A lead of a few tenths of a point is within the range that small changes in test setup can produce, so it is better read as rough parity than as a win.
Independent tests describe a more nuanced picture. V4.1-Flash is strongest at coding and agentic work, where a model must plan steps, call tools and correct its own errors. On the hardest multi-step reasoning problems, it still trails the leading closed models, which keep an edge when a task demands long chains of careful logic.
That split is common among open-weight releases. They tend to close the gap quickly on well-defined tasks that benchmarks measure cleanly, and they lag longer on open-ended reasoning where the best proprietary systems have invested heavily. Buyers choosing between them need to match the model to the job: a coding assistant may be well served by the cheaper system, while a research tool that depends on difficult reasoning may still justify a premium.
6. What is a mixture-of-experts model, and why does it lower costs?
A mixture-of-experts model divides its parameters into specialized groups, often called experts, and uses a routing mechanism to decide which groups should handle each piece of input. Only the selected experts perform calculations for that token, while the rest stay idle. The model can therefore hold a great deal of learned knowledge in its total parameter count while spending the computing power of a much smaller system on each step.
This is why the gap between 552 billion total parameters and about 8 billion active ones is so significant. Computing cost scales mainly with the parameters actually used, so a sparse design can respond to a request far more cheaply than a dense model of the same overall size. The trade-off is memory: all of the parameters must still be stored and available, which makes self-hosting a large mixture-of-experts model a serious hardware commitment even though each individual response is inexpensive to compute.
Pricing in this industry is quoted per million tokens, which are small chunks of text roughly the size of a short word or word fragment. Input tokens are the material a user sends to the model, and output tokens are what the model writes back. Output is priced higher because generating text is more demanding than reading it, which is why DeepSeek's rate of $0.60 for output is four times its $0.15 rate for input.
7. Why does DeepSeek keep undercutting rivals?
The mixture-of-experts architecture is central to the strategy. By activating only a slice of the network for each request, DeepSeek can offer a very large model without paying the full computing cost of running every parameter on every prompt. The savings show up directly in the per-token price.
Pairing that design with an open license adds a second pressure. Competitors who sell access to closed models must now justify their rates against a system that is nearly as capable on common tasks and that customers can run themselves. That does not make the closed models obsolete, but it changes the conversation from what the technology can do to what a customer is paying extra for, such as stronger reasoning, vendor support, security assurances or integration with existing tools.
8. What does it mean for enterprises and the AI market?
The MIT license is especially relevant to organizations with strict data residency rules or reluctance to send sensitive material to an outside provider. Running the model inside their own infrastructure keeps the data in-house and removes dependence on a single vendor's API, availability and pricing decisions.
For the wider market, the release adds to the case that frontier-level coding ability is becoming a commodity. If rivals match it at a fraction of the price, established labs face a choice: lower their rates, improve their models enough to justify the premium, or compete on service and trust instead. Any of those responses benefits customers, although the outcome will take months to appear in published price lists.
There are limits to the conclusions that can be drawn. Benchmarks test particular tasks under particular conditions, and peak-hour pricing doubles the advertised rates, so real-world bills depend on when and how a customer uses the service. Moving a production system to a new model also carries switching costs, from rewriting prompts to retesting behavior.
9. What happens next for DeepSeek V4.1 Flash?
Three things are worth watching. The first is independent confirmation of the 237 tokens-per-second figure across broader workloads, since a single reported test is not enough to establish sustained performance. The second is the pace at which developers migrate existing applications to the model, which will show whether the price advantage outweighs the friction of switching.
The third is the reaction from Western labs. Anthropic, OpenAI and others could respond with lower prices, new tiers or stronger reasoning features. No such response has been announced, and the coming weeks will show whether DeepSeek's pricing pushes the rest of the market to adjust.