DeepSeek is replacing its flagship V4-Pro model with a cheaper open model that it claims can outperform its own top-tier system on several coding and agent-focused workloads, while cutting the cost of serving AI requests.
The new model, DeepSeek V4.1-Flash, has 552 billion parameters and is being released under the MIT licence, meaning developers can download, modify and deploy it themselves. More significantly for existing customers, DeepSeek says all V4-Pro requests will begin routing to V4.1-Flash from September 14 at 04:00 UTC, with users billed at the lower Flash pricing until a future V4.1-Pro arrives.
That makes this more than another model launch. DeepSeek is effectively retiring its current flagship and betting that efficiency matters as much as raw scale, particularly for AI agents that repeatedly read large amounts of context, call tools and generate relatively small amounts of output.
V4.1-Flash uses what DeepSeek describes as a causal encoder-decoder architecture, activating 8 billion parameters per token when processing input and 16 billion during output generation. The company says this structure helps reduce the cost of agent workloads, where models may spend far more time ingesting context than producing responses.
The model supports a context window of up to one million tokens and can process images natively. DeepSeek also says V4.1-Flash was pre-trained on 45 trillion tokens.
One of the more interesting technical claims involves memory use. DeepSeek says the model requires 890 bytes of KV cache per token, significantly reducing the high-bandwidth memory needed during long-running sessions. That could matter for operators trying to serve agentic workloads at scale, where memory consumption can become as important as compute.
DeepSeek’s benchmark numbers are aggressive. The company reports a score of 74.2 on DeepSWE v1.1, narrowly ahead of Anthropic’s Claude Opus 5 at 74.0 and OpenAI’s GPT-5.6 Sol at 73.0. It also claims an 88.1 score on CyberGym.
Those results should be treated cautiously. They are DeepSeek’s own figures and have not been independently verified. The same benchmark table also shows areas where V4.1-Flash falls behind leading closed models. On Humanity’s Last Exam, for example, DeepSeek reports 36.8 versus 56.3 for Claude Opus 5, while its ProgramBench result also trails substantially.
Pricing may prove just as important as benchmark leadership. DeepSeek says off-peak output pricing is $0.60 per million tokens, while cached input costs $0.003 per million tokens, with peak weekday rates doubling those figures. The cut comes shortly after DeepSeek raised prices around its previous model generation, underscoring how aggressively AI providers are adjusting economics as inference efficiency improves.
For developers, the more consequential question may not be whether V4.1-Flash wins every benchmark, but whether it delivers strong enough coding and agent performance at a cost that is difficult to ignore. If independent testing confirms even part of DeepSeek’s claims, rivals may have to respond on pricing as much as capability.
