Chinese AI lab DeepSeek officially launched V4.1 Flash on Wednesday, an interim model it says beats its own flagship V4 Pro on performance, cost and speed — while cutting API prices by as much as 60% from its predecessor. The release follows a brief public beta that opened Tuesday and capped access at 20 concurrent requests per account, an unusually constrained rollout for a lab that has repeatedly used aggressive pricing to pressure rivals since its V3 and R1 releases last year.
V4.1 Flash is a 552-billion-parameter mixture-of-experts model built on a new "Causal Encoder-Decoder" architecture, with roughly 8 billion active parameters per token on input and 16 billion on output, and a context window of up to 1 million tokens. DeepSeek trained it on 45 trillion multimodal tokens, marking the first time the V4 series has built image and speech understanding directly into the model rather than bolting on a separate vision module, as it did with an experimental release in August. The company published the weights under an MIT license on Hugging Face, keeping them freely available to developers.
Undercutting on Price
Under the new pricing, effective September 10, cached input tokens cost $0.003 per million during off-peak hours, down 60% from V4-Flash rates; uncached input fell 33% to $0.15 per million; and output pricing dropped 11% to $0.60 per million. Peak-hour rates run double the off-peak figures. Starting September 14, DeepSeek said API calls still directed at V4 Pro will be automatically routed to the new Flash infrastructure and billed at Flash pricing, effectively retiring the older model's dedicated capacity.
DeepSeek said internal and external testing showed V4.1 Flash outperforming V4 Pro "on performance, cost, speed and total runtime," and positioned it as competitive with Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 on agent-style tasks, according to details reported by benchmark tracker CellCog, though it trailed both on the Terminal-Bench 3.0 coding-agent test. The company also touted efficiency gains in how the model manages memory during long agent sessions, saying its cache footprint is roughly a quarter the size of the previous Flash model's.
The release lands amid intensifying price competition among frontier AI labs, as Chinese developers continue to ship frequent, low-cost model updates aimed at undercutting the per-token pricing of their better-funded U.S. rivals. For enterprise customers running high-volume agent workloads, where costs can scale quickly with usage, this week's price cuts give DeepSeek another opening to compete on economics even as it works to close the performance gap with the largest U.S. models.
