The Pulse
DeepSeek V4.1-Flash Cuts Agent Costs With Causal Encoder-Decoder
DeepSeek released V4.1-Flash, a 552B MoE model with native vision, 1M context, and KV cache reduced to one-quarter prior size. The model routes V4-Pro traffic from Sept 14 and undercuts frontier pricing on cached input.
AI.info Team ·
“DeepSeek-V4.1-Flash is now live on the DeepSeek API with native multimodal support.”
DeepSeek, V4.1-Flash release note, 10 September 2026
DeepSeek Ships V4.1-Flash on September 10
DeepSeek published its V4.1-Flash model announcement on the company API documentation site on September 10. The release introduces the smallest model in a new Causal Encoder-Decoder architecture family and adds native visual understanding to the line.
The 552-billion-parameter mixture-of-experts model activates only 8 billion parameters during input processing and 16 billion during generation. It supports a 1-million-token context window, a maximum output of 384,000 tokens, and ships with open weights under the MIT licence on Hugging Face.
Architecture Targets Agent Workloads
The new design splits a 40-layer Transformer into a 20-layer causal encoder and 20-layer decoder. KV cache from the encoder projects directly to the decoder, cutting active computation for input-heavy agent tasks that repeatedly process tool results and long histories.
Additional techniques—Compressed Sparse Attention 2, hierarchical sparse indexing, FP4 KV caching, and SWA Bounded Replay—shrink the global KV cache to roughly 890 bytes per token. Persistent storage on SSD falls to about one-eighth of the prior V4-Flash generation.
Pricing and Migration Details
Off-peak cached-input pricing starts at $0.003 per million tokens, with output at $0.60 per million. The model replaces both V4-Flash and the experimental V4-Flash-Vision-Exp on the API; legacy names now route to V4.1-Flash.
Beginning 04:00 UTC on September 14, the release note said, all requests to the V4-Pro endpoint would route to V4.1-Flash at the lower rates until a V4.1-Pro version arrived. That has since been walked back: DeepSeek’s pricing page now says that “in response to user demand” it will keep serving V4-Pro after September 14 with billing unchanged. Official partners WorkBuddy and OpenCode have already integrated the new model.
The gap between the two prices is the point of the release. Off-peak, a cache hit on V4.1-Flash costs $0.003 per million input tokens against $0.022 on V4-Pro, a cache miss $0.15 against $0.66, and output $0.60 against $1.98 — roughly a seventh, a quarter and a third. Peak rates are double the off-peak ones throughout. The concurrency limit rises with it, from 500 requests on V4-Pro to 2,500 on Flash, which matters more than the headline price for anyone running agents in parallel.
Benchmark Performance
The benchmark table is DeepSeek’s own, published on the model card rather than measured independently. It puts V4.1-Flash at 74.2 on DeepSWE v1.1, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0; at 90.6 on Terminal-Bench 2.1; and at 88.1 on CyberGym. On harder reasoning the same table shows the model behind the closed frontier: 36.8 on Humanity’s Last Exam against 56.3 for Opus 5, and 90.9 on GPQA Diamond against 94.1 for GPT-5.6 Sol.
Independent measurement so far is thinner. Artificial Analysis places the model at 40 on its Intelligence Index, well above the median of 18 for open-weight models of comparable size. DeepSeek’s own release note says only that “tests by multiple parties” put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, without naming the parties or publishing the figures.