DeepSeek Readies V4.1 Flash: Cheaper and More Powerful Than the Old Pro Model

DeepSeek has announced the imminent release of its new language model, V4.1 Flash. The official launch is slated for roughly September 10, 2026, Beijing time. The news came not as a splashy press release but as a banner in the user dashboard on the platform.deepseek.com portal — those discussing it on Hacker News admitted they didn't immediately track down the original source.

According to the developers, after internal and external testing V4.1 Flash “comprehensively surpassed” the previous flagship V4 Pro on every key indicator: performance, cost, speed and task completion time. In other words, the model that ranks lower by name outpaces the higher one — a trend that commenters compare to the recent case in which GLM 5.3 Flash turned out to be both better and cheaper than GLM 5.2.

A Contentious Routing Decision

One item in the announcement drew particular attention. After V4.1 Flash launches and until V4.1 Pro is released, all requests addressed to the Pro model will be automatically rerouted to V4.1 Flash — and billed at the Flash price. DeepSeek frames this as looking out for users, but many saw the move differently.

A reasonable objection surfaced in the discussion: if a workflow has been carefully tuned and tested for a specific model, a sudden swap — even to a “better” one — can break its familiar behavior. Regression testing, participants noted, is something you want to run on your own schedule, not one imposed on you. There was also a reminder that open-weight models can be spun up with third-party providers like Together.ai or OpenRouter if the previous version is still needed.

Pricing

Starting at 12:00 Beijing time on September 10, a new pricing plan takes effect for the Flash lineup. During off-peak hours, the promised rates are $0.003 for an input-token cache hit, $0.15 for a miss and $0.6 for output tokens (per million). During peak hours the rates double. By the comparison offered in the comments, each of these figures comes in below the current Flash rates.

On the whole, the community welcomed the announcement with enthusiasm: many have long used the “flash” models as a cheap, fast workhorse for agentic tasks, praising their excellent price-to-quality ratio. Among the gripes about the current version are a tendency to occasionally reply in Chinese to English prompts, as well as limits on output length. Whether V4.1 Flash and the future V4.1 Pro will remain open-weight models the company has yet to clarify.

Source: Hacker News