DeepSeek says V4.1 Flash will replace V4 Pro API traffic at lower prices

I found the source in a banner on DeepSeek's platform usage page. The original post made me hunt for it for a couple of minutes.

If Flash keeps surpassing the preceding Pro model at this pace, I am very interested. Anthropic and OpenAI may tighten their anti-distillation defenses, though.

I hope this is open weight and keeps the current V4 Flash Vision architecture and size. That model is probably the best LLM I can run on a 128 GB machine.

I found that the new model already works when called through the API, even though the API does not list it.

I use V4 Flash for most personal projects and reach for V4 Pro or Kimi 3 only on harder work or project-level reviews. A better, cheaper Flash model would fit that pattern well.

I do not want DeepSeek to silently swap a production model after I have validated a workflow on V4 Pro. Keep V4 Pro deprecated for a stated period, because a replacement can change JSON formatting and force regression testing on DeepSeek's schedule.

I have been impressed by V4 Flash's cost and performance. It is my default agent in Omarchy, and I have used it to build Rails apps and tests.

I mostly use Chinese Flash models now because they cost pennies per answer. I want a fast, cheap, consistent model more than one that invents mathematics.

I used to avoid Chinese models over surveillance concerns, but I no longer think any major AI provider keeps my data private. I am willing to use them while hoping the models eventually run locally.

I work on agents, and Chinese Flash models are often better than Gemini Flash on tool-following tasks at 20 to 30% of Gemini's task price. DeepSeek still lacks non-text inputs, and failed tool calls can send an agent into retries unless prompts handle errors carefully.

I want a real medium reasoning setting. DeepSeek's low is close to reasoning off, while high and max take too long and push costs up.

I have seen Flash answer English prompts in Chinese or German, sometimes after retries. Adding 'in English' helps, but I should not need to force the language repeatedly.

I have also seen V4 Flash 7/31 behave badly in coding harnesses, including infinite loops, hallucinations, premature termination, and invalid tool calls.

The claim that output pricing rose nearly fourfold is wrong. Current Flash output is 0.66, so the announced 0.60 rate is a reduction.