DeepSeek shipped V4.1-Flash today, publishing MIT-licensed weights to Hugging Face and setting a September 14 cutover after which every deepseek-v4-pro API call silently reroutes to the Flash endpoint at Flash prices. The move compresses the frontier price gap with Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol into something closer to a rounding error, and it hands operators four days to decide whether their production stacks tolerate a swap.

The architecture is where the cost story actually lives. V4.1-Flash is a 552B-parameter mixture-of-experts backbone paired with a 196B "Engram" conditional-memory module, bringing total weights to roughly 748B. Only 8B parameters activate per token during prefill, 16B during decode, drawn from 384 routed experts per layer plus one shared expert, with six routed experts firing per token across a 40-layer stack. The context window runs to 1M tokens; maximum output length is 384K.

The KV cache is where DeepSeek's engineering shows up in a bill. FP4 quantization plus Compressed Sparse Attention 2 and a Hierarchical Sparse Indexer bring the global cache to roughly 890 bytes per token, about a quarter the size of V4-Flash and, per Intelligent Living's comparison table, 437 times smaller than V1. That's what makes the off-peak $0.60-per-million output token price legible instead of aspirational. DeepSeek-ViT, a from-scratch vision encoder, handles images up to roughly 1344×1344 pixels natively, and a three-block DSpark drafter replaces the MTP module used in V3.

Reasoning effort is now a dial from 1 to 100 exposed through the API. Inference paths cover vLLM, SGLang, and Transformers on day one, and the RL training scaffold names read like a competitive census: Claude Code, Codex, OpenCode, Pi, mini-SWE, and DeepSeek's own Harness.

The beta preview ran September 8 under the temporary ID deepseek-v4.1-flash-expires-on-0910, capped at 20 concurrent requests. Off-peak pricing went live at 04:00 UTC today. The correct name going forward is deepseek-flash; deepseek-v4-flash and the retired deepseek-v4-flash-vision-exp endpoint remain as legacy aliases pointing at V4.1-Flash.

The forced-migration mechanic echoes the recent frontier-tier price cuts across the industry, and it lands against a backdrop of Reuters reporting that DeepSeek has tapped CITIC Securities to prepare for a possible Shanghai STAR Market listing. Independent numbers from Artificial Analysis and Vals.ai will settle whether Flash actually holds Pro-tier quality. Until then, the routing table decides for you.

Sources

https://api-docs.deepseek.com/updates/
https://www.marktechpost.com/2026/09/10/deepseek-ai-released-deepseek-v4-1-flash-with-1m-context-fp4-kv-cache-and-cross-layer-attention-reuse/
https://www.intelligentliving.co/deepseek-v41-flash-pricing-release/
https://time.news/deepseek-launches-v4-1-flash-a-faster-multimodal-mixture-of-experts-model/
https://superpowerdaily.com/posts/deepseek-launches-v4-1-flash-and-plans-to-replace-v4-pro-api-traffic