> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# DeepSeek introduces V4.1 Flash and new Flash pricing
- URL: https://globalfeed.ai/en/deepseek-introduces-v4-1-flash-and-new-flash-pricing/
- Published: 2026-09-09T18:15:10.000Z
- Updated: 2026-09-09T18:15:10.000Z
- Description: DeepSeek says V4.1 Flash is cheaper and more capable than V4 Pro, with revised Flash series pricing taking effect on 10 September 2026.
- Author: GlobalFeed Editor
- Tags: DeepSeek, pricing, API, x-deepseek_ai, dil-en, denetim-gerekli, video

DeepSeek has announced V4.1 Flash, describing it as cheaper and more capable than V4 Pro. The company is also revising pricing for the Flash series.

The new rates apply from 12:00 Beijing Time on 10 September 2026\. Off-peak unit prices are:

- Input, cache hit: $0.003
- Input, cache miss: $0.15
- Output: $0.6

Peak-hour prices are double the off-peak rates. The company advised users to plan their usage accordingly.

## The debate moved to model swaps

Most of the Hacker News discussion focused less on price than on V4.1 Flash replacing V4 Pro. One commenter argued that a user who validated a workflow on V4 Pro may not want to suddenly test it in production on V4.1 Flash, and suggested keeping V4 Pro available but deprecated for a defined period before removal. Because these are open weights models, the same commenter noted, V4 Pro can still be run through providers such as Together.ai or OpenRouter for as long as they keep it up.

Another commenter, who works for an education department serving a chatbot to students, said every model upgrade produced a measurably different adherence to existing system prompts on sensitive topics, which is why content safety reviews are required. Others said changing the underlying weights invalidates any prior evaluation of the output distribution, so teams need to run regression tests on their own schedule rather than have one forced on them.

Responding to the argument that model outputs are non-deterministic anyway, one commenter cited the DeepSeek-V4 paper, which states the team implemented end-to-end, bitwise batch-invariant and deterministic kernels with minimal performance overhead. Providers may not implement deterministic inference, the commenter added, but it is possible.

The full thread is on [Hacker News](https://news.ycombinator.com/item?id=49624603&ref=globalfeed.ai).