DeepSeek debuts sparse attention, cuts API prices by half

DeepSeek launched DeepSeek-V3.2-Exp on Monday, an experimental model designed to make processing long texts cheaper through a new sparse-attention architecture. The Chinese company has also cut some API prices by more than 50%, putting renewed pressure on pricing across the industry.

The model is available on Hugging Face with its weights under an MIT license, alongside a technical paper published on GitHub. That will allow researchers and outside providers to test whether the announced savings hold up beyond DeepSeek’s own evaluations.

Fewer calculations for working with long texts

Large language models use a mechanism called attention to determine which parts of a text are relevant when generating each response. In its conventional form, the system compares tokens—the units into which text is divided—with one another. The longer the document or conversation, the more memory and computing power that operation requires.

DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention, a variant that avoids processing the entire available context with the same intensity. First, a component called the lightning indexer identifies excerpts that are likely to be useful. A second system then makes a more precise selection of the tokens that will enter the main attention calculation.

The idea of sparse attention is not new in academic research. Various labs have spent years trying to replace exhaustive token-to-token comparisons with selective methods. DeepSeek’s contribution is to integrate it into one of its large models, retain a 128,000-token context window, and publish the weights to make independent testing easier.

The benefits should be most apparent in tasks such as analyzing code repositories, summarizing lengthy contracts, querying document databases, or maintaining very long conversations. For short requests, where the context contains relatively few tokens, the difference may be much smaller.

Cached input drops below three cents

The new pricing table puts cached input at $0.028 per million tokens. That is the figure behind the claim that the price falls below three cents, but it comes with an important qualification: it applies only when DeepSeek has already stored the context and can reuse it.

Input that is not cached costs $0.28 per million tokens, while generated output is priced at $0.42 per million tokens. DeepSeek describes these changes as a reduction of more than 50% from its previous rates.

Caching is especially useful for applications that repeatedly send the same instructions or documents. An enterprise assistant can reuse its internal handbook across hundreds of queries; an application that receives a different text with every request will not see the same savings.

It is also important to distinguish the commercial price from the technical improvement. DeepSeek may subsidize its API or accept lower margins to attract users, while sparse attention is intended to reduce the actual cost of running the model. Publishing the weights makes it possible to study that second part without relying on the rates set by the company.

An experiment aimed at preserving V3.1’s quality

DeepSeek describes V3.2-Exp as an experimental version and compares it with DeepSeek-V3.1-Terminus, its immediately preceding reference model. The company’s internal evaluations point to similar performance across various tests, despite using fewer resources to handle long contexts.

That balance is the critical issue. Selecting fewer tokens reduces computation, but it also creates the risk of discarding a sentence, figure, or instruction needed to answer correctly. A model could be cheaper while losing accuracy when the relevant information is buried in a lengthy document.

Third-party tests will need to measure both cost and quality: speed, memory usage, retrieval of data located in different parts of the context, and behavior as the input grows longer. The company’s preliminary results are a signal, not definitive validation.

Another price offensive from China

DeepSeek already disrupted the market at the beginning of 2025 with R1, a reasoning model that competed with U.S. systems at a significantly lower price. V3.2-Exp is less spectacular, but it targets one of the costs that weighs most heavily on companies’ books: inference—the cost of running an already-trained model each time a user submits a request.

For developers, the new rates make it cheaper to experiment with agents, internal search tools, and document analysis. For competitors, the challenge is not just to match the published price: they will have to show that their models offer enough quality, reliability, or ease of integration to justify higher rates.

DeepSeek does not present V3.2-Exp as a definitive replacement for its earlier family of models. Its immediate purpose is to test sparse attention at scale. If independent evaluations confirm that it preserves quality in long contexts, the technique could become a practical way to reduce an inference bill that grows as applications feed models more information.