What changed
DeepSeek released V4.1-Flash on 10 September 2026 under the API name deepseek-flash, retired V4-Flash and V4-Flash-Vision-Exp, and temporarily pointed their names at the new model 1. The same post says every deepseek-v4-pro request routes to V4.1-Flash from 04:00 UTC on Sept 14, 2026 1. It still says so. The pricing page says the opposite: its footnote now keeps V4 Pro running after September 14, with billing unchanged 4. Two Wayback captures on 11 September, at 07:47 and 16:55 UTC, bracket the edit 56, and TokenCost published a dated correction on 12 September 7.
| Model name you send | Served by, per the pricing page 4 |
|---|---|
deepseek-flash | DeepSeek-V4.1-Flash |
deepseek-v4-flash, deepseek-v4-flash-vision-exp | DeepSeek-V4.1-Flash, legacy names |
deepseek-v4-pro | DeepSeek-V4-Pro-0813 |
The swap that did happen
I think the Pro deadline drew the attention while the Flash alias is the change that bites. When V4-Flash left preview in July, this radar noted that only its post-training had changed. This time the architecture changes too: V4.1-Flash is a causal encoder-decoder that activates 8B parameters per token in prefill and 16B in decode 2, with 552B in total against 284B for V4-Flash 3. A request that still sends deepseek-v4-flash is accepted and served by that model 4. An eval score, a latency budget or a cost baseline keyed on the name now describes a model you no longer call.
A reprieve with no end date
Before the edit, the same footnote said V4.1 Flash had surpassed V4 Pro and that DeepSeek planned to retire V4 Pro in an orderly manner 5. The new text sets no end date and only promises further notice of any change 6. I read it as a reprieve on notice, and I would not plan a quarter around it.
Impact on your team
If you pinned deepseek-v4-pro, change nothing today. Spend the reprieve running your own tasks on deepseek-flash and deepseek-v4-pro side by side while the two names still reach distinct models 4, so the comparison is ready the day DeepSeek gives notice. If your stack still sends deepseek-v4-flash, rename it to deepseek-flash now and re-baseline: the legacy names are a temporary compatibility route 1, and every result recorded under them since the release measures V4.1-Flash. Log, next to each eval run, the model version the pricing page lists, because the name alone has changed meaning twice since July.