DeepSeek V4 opens the million-token era at a discount

On 24 April, DeepSeek released the preview of its V4 series and open-sourced the weights under the MIT licence, and the sentence from the announcement that deserves the most attention was also the most matter-of-fact: a million-token context is now the standard configuration for every official DeepSeek service. The line was the end of a quiet competition. A few weeks earlier the million-token context window was the frontier’s newest selling point; by the end of April it was the free default of a Chinese lab whose API prices are a small fraction of its rivals’.

The architecture is the serious part. V4 keeps DeepSeek’s mixture-of-experts design and introduces a compressed-attention scheme, with sparse attention applied along the token dimension to shrink the key-value cache, so that long contexts cost far less in memory and compute than a straightforward attention mechanism would demand. The Pro version, with 1.6 trillion total parameters of which 49 billion are active per token, and the Flash version, smaller and faster, both ship with the million-token window. The technical claim, that a model can hold a million tokens without the usual quadratic blow-up in cost, is the reason the pricing is possible, and the two belong together.

The benchmark story is less important than the market story, but it is still a strong one. DeepSeek reports V4-Pro at the top of open models on agentic coding, ahead of every other open model on maths, STEM and competitive programming, and close behind the best closed models on world knowledge. The company also notes that it now uses the model internally for agentic coding, and that its engineers rate it above one prominent closed model on day-to-day use. None of these claims will settle anything permanently, but they establish the pattern that matters: an open model from a Chinese lab is now part of the same conversation as the closed frontier, and it sets the price floor.

The reaction of the domestic hardware ecosystem was immediate, which is itself a story. Baidu’s cloud platform had the model adapted on day one, Cambricon published open-source adaptation code, Moore Threads and the BAAI community deployed the Flash version on their flagship GPUs, and Huawei’s Ascend line announced full support across its product range. The signal here is that the Chinese AI stack has reached the point where a major model release triggers coordinated adaptation by domestic chip makers within days, an integration layer that did not exist two years ago. The constraint is still hardware: DeepSeek says V4-Pro throughput is limited by high-end compute supply, and prices for the Pro tier are expected to fall once Ascend 950 volume production lands in the second half of the year.

What the release says about the industry is worth spelling out. The frontier labs have been competing on context windows and capability, with prices set to pay for enormous training runs. DeepSeek’s move cuts the floor out from under that logic: if a million tokens is the default and open-sourced, the long-context race stops being a differentiator and becomes table stakes, and the pressure moves to serving cost, where the Chinese lab’s compressed attention and cheap compute give it an edge. The honest reaction to this is not cheerleading for any single company but an observation about how the market works: open weights from a low-cost producer have a way of turning the entire industry’s pricing into a race to the bottom, which is excellent for customers and uncomfortable for the labs whose business models depend on scarcity. The models themselves are close in capability; the economics are not.

DeepSeek has promised that million-token context remains standard for all official services going forward, and the old model names are being retired in July. The preview is a preview, with the usual caveats about performance under load, but the direction is clear. The company that made the price war famous in January 2025 has now done the same thing to context windows, and the rest of the industry, including the American labs that set the frontier’s prices, is going to have to answer the question it posed: if a million tokens is free, what exactly are you charging for?