DeepSeek has released V4-Flash version 0731, keeping the aggressive pricing the line is known for — and attaching an unusual warning to it.

The numbers

V4-Flash costs $0.14 per million input tokens on a cache miss and $0.28 per million output tokens. A cache hit drops the input cost to $0.0028 per million — fifty times cheaper, which makes prompt structure a first-order cost decision rather than an optimisation detail.

The larger V4-Pro sits at $0.435 input and $0.87 output per million, with cache hits at $0.003625. Both models offer a 1M token context window and support thinking and non-thinking modes.

The warning

The pricing documentation states plainly that DeepSeek plans to raise overall API pricing in the near future, with a significant increase expected. Vendors rarely pre-announce their own price rises. Publishing it means current rates should be read as promotional rather than structural.

Why it matters

At $0.14 per million inputs, DeepSeek is the cheapest credible option in the frontier-adjacent tier, undercutting OpenAI's freshly reduced Luna at $0.20. That gap is worth something — but only for as long as it lasts, and the vendor has said it will not last.

The practical lesson is about lock-in rather than price. Building a high-volume pipeline around whoever is cheapest this quarter is fine if switching costs are near zero. It is expensive if the integration is deep. The cache-hit discount cuts both ways here: designing prompts around one provider's caching behaviour is exactly the kind of work that does not transfer.

How to read the cache pricing

The fifty-fold gap between cache hit and cache miss is the most actionable number here. It means the shape of a prompt — what stays constant across calls and what varies — drives cost more than model choice does. Systems that put a large stable preamble first and vary only the tail get most of their input tokens at $0.0028 per million.

That is real money at volume, and it is also a subtle form of lock-in: prompts engineered around one vendor's caching behaviour do not transfer cleanly to another.

What to plan for

Treat current DeepSeek pricing as temporary, because the vendor has said it is. If a workload is being sized on these rates, model what happens at two or three times the price before committing. A pipeline that only works at $0.14 per million is a pipeline with a countdown on it.