DeepSeek, the Chinese artificial intelligence lab based in Hangzhou, has released its V4-Flash model, which it claims is the most cost-effective AI model in the world. The model’s Application Programming Interface (API), a set of rules that lets software talk to the model, is priced at $0.14 per million input tokens and $0.28 per million output tokens. At these rates, processing ten million tokens of data costs just $1.40, a price point that has shaken the global AI industry.
What Is the DeepSeek V4-Flash Model?
The DeepSeek V4-Flash is a large language model (LLM), an AI system trained on massive amounts of text to understand and generate human-like language. It is the lighter and cheaper sibling of DeepSeek V4-Pro, and together the two form the DeepSeek V4 family, the fourth generation of models from the lab.
A token is the basic unit of text that a model processes. It can be a word, a number, or even a punctuation mark. When a user sends a request to the model, the text is split into tokens, and the model bills for the tokens it reads (input) and the tokens it writes back (output).
V4-Flash carries 284 billion total parameters, the internal knobs that store what the model has learned, but it activates only 13 billion parameters per token through a technique called Mixture of Experts (MoE). Instead of waking up the entire network for every question, the model routes each query only to a small team of specialist sub-networks, or “experts”. This design is the core reason it can stay fast and cheap while handling complex tasks.
The model supports a 1 million token context window, meaning it can hold roughly 800,000 words, or the equivalent of several full-length novels, in memory at once. It also offers both a thinking mode, which reasons slowly through hard problems, and a non-thinking mode, which answers quickly. The legacy model names deepseek-chat and deepseek-reasoner have now been merged into this single V4-Flash model, with each name mapping to one of these two modes.
Pricing That Reshaped the Market
V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens. When a request matches text the system has already seen, called a cache hit, the input price falls to just $0.0028 per million tokens, a discount of roughly 98 percent over a fresh input, or cache miss.
To understand why this is disruptive, these rates must be compared with rival frontier models. GPT-5.5, from the American company OpenAI, costs $5 per million input tokens and $30 per million output tokens. Claude Opus 4.7, from Anthropic, costs $5 input and $25 output. That makes V4-Flash roughly 36 times cheaper on input and more than 100 times cheaper on output than GPT-5.5.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.28 |
| DeepSeek V4-Pro | $1.74 | $3.48 |
| Gemini 3.1 Flash | $0.50 | $3.00 |
| GLM-5.1 | $1.05 | $3.50 |
| Claude Opus 4.7 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
The practical effect is dramatic for companies that run AI at scale. A business processing 100 million tokens a month, a heavy but realistic workload, would pay about $21 a month on V4-Flash, compared with roughly $1,500 on Claude Opus or $1,750 on GPT-5.5. DeepSeek argues that this combination of strong performance and near-negligible cost is what makes V4-Flash the most cost-effective model available today.
How the Model Keeps Costs Low
V4-Flash achieves its low price through a mix of architectural and engineering choices, rather than by cutting corners on capability.
Mixture of Experts Efficiency
The Mixture of Experts (MoE) design is the biggest lever. With 284 billion total parameters but only 13 billion active on any single query, the model does a fraction of the computational work of a full-size model on each request. Lower compute per request directly translates into lower cost per token for the provider, savings that are passed on to the developer.
Sparse Attention
The model uses a hybrid attention architecture, combining two techniques called Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Rather than scanning all one million tokens of context with equal effort, the model focuses its attention only on the most relevant passages. DeepSeek states that at a one million token context, V4-Flash requires only about 10 percent of the compute and 7 percent of the cache that its predecessor, DeepSeek-V3, needed.
Cheaper Memory
The V4 family is trained using mixed precision, a method where parts of the model are stored at lower numerical accuracy. Expert parameters use FP4 precision, while most other parameters use FP8, which trims memory usage and hardware cost without a major hit to performance. The models are also released as open source under the MIT license, meaning anyone can download the code, run it on their own machines, and modify it, which further pushes down the effective cost of using AI.
DeepSeek’s Rise: From R1 Shock to V4 Family
DeepSeek was founded in July 2023 by Liang Wenfeng, a Chinese entrepreneur who earlier built the quantitative trading firm High-Flyer Capital Management. The lab was set up in Hangzhou, the capital of China’s Zhejiang province, as a research-focused venture backed largely by High-Flyer’s profits.
DeepSeek first grabbed the world’s attention in January 2025 with the release of DeepSeek-R1, a reasoning model that matched or beat several leading Western models while being built at a fraction of the cost. R1 was trained in about two months on less powerful, export-restricted chips, for a reported cost of under $6 million. When the model emerged, global technology stocks fell sharply, and Nvidia lost close to $600 billion in market value in a single day. Analysts at the time called it artificial intelligence’s “Sputnik moment”, a reference to the Soviet satellite that startled the United States into the space race.
The lab followed with the DeepSeek-V3 series, and on 24 April 2026 it released the full V4 family, introducing V4-Pro and V4-Flash together. The official V4-Flash API entered public beta on 31 July 2026, and DeepSeek has said the final version of V4-Pro will follow. The V4-Flash update retains the same architecture as the earlier preview but adds re-post-training that DeepSeek says sharply improves its ability to act as an AI agent, a program that plans and executes tasks on its own.
What the Launch Means for the Global AI Race
The V4-Flash launch lands at the centre of a heated contest between American and Chinese AI labs, with huge sums riding on who can deliver the best model at the lowest price.
For developers, cheap frontier-grade AI lowers the barrier to building products. Startups that previously could not afford premium models can now run coding assistants, customer-support bots, and document analysis at scale. Analysts note that this could accelerate a “race to the bottom” on pricing across the industry, forcing OpenAI, Anthropic, and Google to justify far more expensive rates.
The open-source release also carries strategic weight. Because V4-Flash can be downloaded and run locally, companies and governments worried about data privacy can deploy it on their own servers instead of sending data to foreign cloud providers. This makes the model attractive for data sovereignty, the idea that a country or organisation should control its own data and computing infrastructure.
For China, the release is part of a broader push for AI self-reliance. American export controls restrict Chinese companies from buying Nvidia’s most advanced chips. DeepSeek has stated that V4 runs efficiently on domestic Chinese hardware, including chips from Huawei, and its success signals that advanced AI can be built without access to cutting-edge American processors. The model’s low cost and strong agent performance also challenge the assumption that only massive, expensive models built in Silicon Valley can lead the field.
The Way Forward
DeepSeek has positioned V4-Flash as a workhorse for high-volume, latency-sensitive workloads, while reserving V4-Pro for the most demanding reasoning and coding tasks. As the two models continue to mature, and as rivals respond with their own price cuts, the cost of AI inference, the expense of running a trained model, is expected to keep falling.
The deeper question now is whether falling prices lead to wider adoption of AI across industries, or whether they squeeze the margins of the companies that dominate today. Either way, the V4-Flash launch marks a moment when the economics of artificial intelligence, not just its capability, became a central arena of global competition.
Key Takeaways
- DeepSeek V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, making it the most cost-effective AI model claimed by its maker.
- DeepSeek was founded in July 2023 in Hangzhou, China, by Liang Wenfeng, who also founded the quantitative trading firm High-Flyer Capital Management.
- V4-Flash has 284 billion total parameters but activates only 13 billion per token using the Mixture of Experts (MoE) architecture.
- The model supports a 1 million token context window and offers both thinking and non-thinking modes.
- V4-Flash is 36 times cheaper on input and over 100 times cheaper on output than OpenAI’s GPT-5.5.
- The models are released as open source under the MIT license, allowing free download and modification.
- DeepSeek’s earlier R1 model, released in January 2025, triggered an AI “Sputnik moment” that erased nearly $600 billion in market value from Nvidia in a single day.
