If your app still sends deepseek-v4-flash, the name no longer tells you which model answers. DeepSeek has retired V4-Flash and routes its compatibility aliases to DeepSeek-V4.1-Flash. The current API name is deepseek-flash.
The September 10 release adds native visual understanding. DeepSeek describes a 552B-parameter mixture-of-experts model with an asymmetric encoder–decoder: 8B active parameters on the input side and 16B on the output side. Those are the provider's architecture figures, not a claim that the full checkpoint fits on a small machine. Release announcement.
For an application, the more immediate details are vision, tool calls, JSON output, and both thinking and non-thinking modes. The API lists a million-token context and up to 384K output tokens. Current API documentation.
Pricing depends on when the request runs. Peak uncached input/output rates are $0.30 / $1.20 per million tokens; off-peak rates are half. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays. The site uses peak rates for its estimates, so a budget does not quietly assume overnight execution.
One detail has already changed since the announcement. The original post said V4 Pro requests would be rerouted after September 14. The live API docs now say V4 Pro service will continue, with billing unchanged. Use that current statement when planning a migration rather than assuming the original retirement plan still applies.
The practical next step is a replay of representative requests. A compatible endpoint can still produce different tool choices, reasoning lengths, or patches. Compare those against the old behavior before directing a production workload to the new model. The model tools include V4.1 Flash; the older V4 examples keep their version labels.