DeepSeek's V4.1 Flash Beats Its Own Flagship — and V4 Pro Is Being Retired

The 552B-parameter open-weight model, out under MIT, adds native vision and a new encoder-decoder layout that shrinks the KV cache about 75%. DeepSeek will route all V4 Pro requests to V4.1 Flash from September 14 and bill them at Flash rates. Token prices are low, but Artificial Analysis finds higher token consumption makes it costlier per task than plain V4 Flash.

Key Takeaways

The 552B-parameter open-weight model, out under MIT, adds native vision and a new encoder-decoder layout that shrinks the KV cache about 75%.
DeepSeek will route all V4 Pro requests to V4.1 Flash from September 14 and bill them at Flash rates.
Token prices are low, but Artificial Analysis finds higher token consumption makes it costlier per task than plain V4 Flash.

AI synthesis4 sources6

DeepSeek released V4.1 Flash on September 10, an open-weight model it is positioning upstream of everything it shipped before. The company says the new model comprehensively surpasses V4 Pro — its current flagship — in performance, cost, speed and total time, and that it therefore plans an orderly retirement of V4 Pro. From 12:00 Beijing time on September 14, requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at Flash rates until a V4.1 Pro arrives. The legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also served by V4.1 Flash.

The specifications are a jump: a 552-billion-parameter core, roughly double V4 Flash, with 1 million tokens of context and up to 384K output tokens. Unlike V4 Pro and V4 Flash, V4.1 Flash handles images natively. Weights are published under the MIT licence, and the model is available through DeepSeek's API.

The engineering story is architectural. A new causal encoder-decoder design splits the workload: an encoder stage reads the input context while a decoder handles generation, with DeepSeek reporting 8 billion active parameters per token for input processing and 16 billion for generation, against 13 billion for V4 Flash. The efficiency mostly lands on input-heavy work — agentic coding and research runs that pour source code, files, web pages and tool output into the context window. DeepSeek also shrank its global KV cache about 75 percent versus V4 Flash, claiming roughly a quarter of the HBM and an eighth of the persistent SSD storage.

Pricing follows: $0.30 per million uncached input tokens and $1.20 per million output tokens at peak, halved off-peak to $0.15 and $0.60. Artificial Analysis puts peak-hour cost at about $0.27 per Intelligence Index task — roughly 60 percent less than V4 Pro and about a seventh of GLM 5.3 or Kimi K3. The catch is verbosity. V4.1 Flash emits noticeably more output tokens per task than V4 Pro, so despite higher per-token prices, V4 Flash still performs the same task for about $0.22. The model also improves on multi-step agent tasks and long-context reasoning, while hallucinating more often when it lacks knowledge.

The move compresses DeepSeek's product line to a single serving tier and puts the onus on developers: anyone pinned to V4 Pro has until September 14 to adapt, and cost planners should measure tokens, not just per-token rates.

DeepSeek's V4.1 Flash Beats Its Own Flagship — and V4 Pro Is Being Retired

Sources

Anthropic's $80.7B loss is an accounting echo. The $125.2B lease is the number that binds

Sources

Related Articles

Anthropic's $80.7B loss is an accounting echo. The $125.2B lease is the number that binds

Meta opens a $95 billion equity sale for AI — with Nvidia, AMD and Broadcom as "cornerstone" buyers

Britain's new biosecurity lab has £100 million of public money — and £200 million it cannot pay salaries with

Microsoft's diagnostic AI beat generalists 83% to 36%. Its own data says the orchestration barely mattered

Anthropic's $80.7B loss is an accounting echo. The $125.2B lease is the number that binds