DeepSeek cuts prices and gives its Flash model eyes

On September 10, 2026, DeepSeek released a new version of its AI model, V4.1 Flash. It can work with both text and images, and the prices for using it have been cut at the same time.

Understands images

According to 36Kr Europe, the model processes text and images in the same core model. It can describe images, recognize text in screenshots and analyze charts. This can be done via public links, direct encoding or file upload. The older V4 Pro does not support images, according to DeepSeek's own documentation.

The model has a context window of 1 million tokens, which is how much text it can keep track of at once. The maximum response is 384,000 tokens. It also supports JSON output, tool calls and thinking mode.

TechNews Taiwan reports that the model has 552 billion parameters in a so-called Mixture-of-Experts architecture, where only part of the model is active at a time. It activates 8 billion parameters for input and 16 billion for output. According to the article, the need for cache memory has been cut to a quarter of the previous generation, and the need for SSD storage to one eighth.

Lower prices and two pricing tiers

Prices were cut starting September 10, and there are now two tiers: a higher price during peak hours and a half-price rate during off-peak hours. According to DeepSeek's pricing page, Flash costs the following per million tokens:

- Input with cache hit: USD 0.003-0.006 - Input without cache hit: USD 0.15-0.30 - Output: USD 0.60-1.20

By comparison, the pricing page lists V4-Pro at USD 0.66-1.32 for input without a cache hit and USD 1.98-3.96 for output. 36Kr Europe reports that the price of cached input has fallen 60 percent compared with V4 Flash. Peak hours are Monday-Friday, 9 a.m.-12 p.m. and 2-6 p.m. China time.

Older models being phased out

V4 Flash and V4 Flash Vision Exp have been discontinued, and V4 Pro is being phased out in stages. Starting September 14, 2026, V4 Pro requests were scheduled to be handled by the new model at V4.1 Flash prices until a V4.1 Pro is released.

For developers, the model is called deepseek-flash in the API. Older model names are still accepted, but the corresponding models have been retired. The API is compatible with both OpenAI and Anthropic formats and, according to the documentation, is supported by coding tools such as Claude Code and GitHub Copilot.

The figures are DeepSeek's own

According to press coverage, V4.1 Flash outperforms V4 Pro in benchmarks, including a score of 90.9 on GPQA Diamond and a Codeforces rating of 3471. The figures come from DeepSeek via the media and have not been independently verified.