DeepSeek points V4 Pro at Flash on September 14

Here is a deprecation that does not look like one. Starting September 14 at 04:00 UTC, every request you send to deepseek-v4-pro gets answered by V4.1-Flash instead, billed at Flash rates, until DeepSeek ships a V4.1-Pro. Your code does not change. Your model does. If you are calling DeepSeek's flagship string from production, Monday is the day a smaller model starts answering your customers.

What actually happened

DeepSeek released V4.1-Flash on September 10 — we covered the 24-hour beta that shipped with an expiry date in its model name last week. The shipped version is a 552B-parameter mixture-of-experts model that activates roughly 8B parameters to read input and 16B to generate output, with a 1M-token context window, native image understanding, and MIT-licensed weights on Hugging Face.

Per DeepSeek's own release note, the architecture is a new causal encoder–decoder that cuts KV cache requirements to about a quarter of the HBM and an eighth of the SSD footprint of the prior generation. That is the reason the price moved. VentureBeat's breakdown puts off-peak rates at $0.15 per million uncached input tokens, $0.60 per million output, and $0.003 per million on a cache hit. Peak rates are double.

DeepSeek's own benchmark numbers put V4.1-Flash at 74.2 on DeepSWE v1.1 against 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol. Those are self-reported. Treat them as a hypothesis to test, not a result.

The operational line is the routing. deepseek-v4-pro and the older V4-Flash identifiers all resolve to V4.1-Flash from September 14.

Why model routing belongs in your change log

We have written before about DeepSeek retiring model aliases outright. This is the quieter version of the same risk. An alias that errors out tells you immediately. An alias that silently resolves to a different model tells you nothing — you find out from a support ticket three weeks later.

Your bill will drop. That is not the finding. Flash pricing on Pro-tier traffic is a real saving. Whether the outputs still clear your bar is a separate question, and only your evals answer it. If you do not have a golden set of 30 real customer inputs with known-good outputs, build one before Monday. It is an afternoon of work and it is the only instrument you have.

Pin the model, don't pin the tier. Reading the model name from config instead of code is the cheap half. The valuable half is logging which model string served each response, so that when quality shifts you can correlate it with a date instead of guessing.

Watch the peak window. Doubling at peak means your bill depends on when your automations run. Batch work that can wait should wait.

Sources: DeepSeek release note, September 10, VentureBeat.