DeepSeek V4.1 Flash cuts agent costs and forces a V4 Pro migration
The multimodal model brings aggressive API pricing, but DeepSeek is also retiring old Flash versions and redirecting V4 Pro traffic on September 14.
What shipped
DeepSeek released V4.1 Flash on September 10 as a native multimodal, 552-billion-parameter MoE model with a one-million-token context window. The company says its architecture activates 8 billion parameters during input processing and 16 billion during generation, while its global KV cache is roughly one quarter the size of the previous V4 Flash model.
The API is live as deepseek-flash. Off-peak rates are $0.15 per million uncached input tokens, $0.003 for cached input and $0.60 for output; peak rates are double.
The forced migration
The previous V4 Flash and experimental Vision models are retired, with their old identifiers temporarily routing to V4.1 Flash. More importantly, at 04:00 UTC on September 14, DeepSeek will begin redirecting every deepseek-v4-pro request to the new Flash model and charging the Flash rate.
For production systems, that is more than a cheaper invoice. The same request using the same identifier can acquire different model behavior, making existing evaluations, quality thresholds and incident records potentially misleading.
Our move
Virtual Arc would put V4.1 Flash into an immediate controlled evaluation for input-heavy agents, but we would not promote it on vendor benchmarks alone. Alongside success rate, we would measure total cost per accepted task, tool latency, structured-output stability and the cost of retries.
Teams using V4 Pro should not wait: regression testing needs to finish before the automatic redirect. Production adoption should begin behind a gateway, on a small traffic slice, with a clear route back to another provider.
DeepSeek V4.1 Flash is the rare model launch that changes both agent economics and the operational plan for existing users. Its cache pricing, native vision and million-token context make it worth testing on repository-scale coding, document-heavy research and long tool trajectories. Virtual Arc would not make it the default yet. DeepSeek has already turned the old Flash identifiers into mutable aliases, and on September 14 it will redirect V4 Pro requests to V4.1 Flash. A model name is therefore no longer a sufficient record of what actually ran. We would immediately freeze a representative production task set, preserve request dates, reasoning settings and returned model metadata, then compare cost per accepted result, wall-clock latency and retry rates. If the economics survive that test, we would canary a small share of traffic behind our own gateway with a fast rollback path. Cheap tokens are a reason to experiment; forced routing is a reason to treat this as a dependency migration, not a free upgrade.