DeepSeek's V4-Pro Survives Past Its Own Retirement Date

DeepSeek released DeepSeek-V4.1-Flash on Sept. 10, 2026, and told users the smaller model had "comprehensively surpassed" its own flagship, V4-Pro. The company gave itself four days to move every Pro request onto Flash pricing. Nine days past that deadline, DeepSeek's own API pricing page still lists V4-Pro as a separate, more expensive model, with no note explaining why the migration did not happen on schedule.

DeepSeek's announcement said that from 04:00 UTC on Sept. 14, every call to deepseek-v4-pro would be served by V4.1-Flash at Flash's lower rate, ahead of a future V4.1-Pro release. Independent analysis site Digital Applied checked the pricing page again on Sept. 17 and found Pro still billing at its old, higher rate. As of this writing, deepseek-v4-pro remains listed as an active model with its own price.

What changed inside the model

V4.1-Flash is a mixture-of-experts design, a model that routes each request to a subset of its parameters instead of all of them, built around what DeepSeek calls a Causal Encoder-Decoder architecture. The model card lists 552 billion total parameters, with only 8 billion active while it reads a prompt and 16 billion active while it generates a response. DeepSeek said the split cuts the memory needed for cached context to a quarter of the prior generation's footprint and shrinks disk storage for that cache to an eighth. The model carries an MIT license, ships with a 1-million-token context window, and is downloadable in 76 quantized versions on Hugging Face.

The migration DeepSeek hasn't finished

The unfulfilled deadline is the clearest gap in an otherwise confident launch. DeepSeek's own numbers make the case for retiring Pro: at off-peak rates, Flash costs $0.15 per million input tokens and $0.60 per million output tokens, against $0.66 and $1.98 for Pro, a gap DeepSeek's announcement frames as proof Flash no longer needs a bigger sibling.

Model	Input, off-peak ($/M tokens)	Output, off-peak ($/M tokens)
deepseek-flash	0.15	0.60
deepseek-v4-pro	0.66	1.98

DeepSeek has not published a reason for the delay, and neither the announcement nor the pricing page addresses it. Digital Applied's benchmark table, reproducing DeepSeek's own comparison, found Pro still ahead of Flash on GPQA Diamond and the text-only version of Humanity's Last Exam, even as Flash gained ground on agentic and coding tasks, a split that may explain why DeepSeek is keeping the bigger model billing separately rather than routing it away entirely.

An independent check on the capability claim

DeepSeek's own benchmark table shows V4.1-Flash reaching 90.9 percent on GPQA Diamond and 79.4 percent on HumanEval, both vendor-supplied and not independently reproduced by this publication. Security research firm Aikido ran the model against its own cybersecurity benchmark and reported a different kind of gain: single-run recall of the benchmark's CVEs rose to 65.6 percent, up from 55.2 percent for the prior Flash release, with pass-at-three recall reaching 84.4 percent.

1. Previous DeepSeek Flash55.2 %
2. DeepSeek-V4.1-Flash65.6 %

Source: Digital Applied, citing Aikido Security research, accessed 2026-09-19

V4.1-Flash is the third large mixture-of-experts release out of a Chinese lab this month, after Alibaba's Qwen3.8-27B and China Telecom's Xing4.0, both of which shipped open weights within weeks of each other. Unlike those two, DeepSeek's release comes with a pricing change still in progress rather than finished, and the company has not said when, or whether, V4-Pro will actually stop taking traffic of its own.

The next checkpoint is whichever comes first: DeepSeek updating its pricing page to match the announcement, or the company saying why it hasn't.

Sources

1. Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient, DeepSeek
2. DeepSeek-V4.1-Flash model card, Hugging Face
3. Pricing, DeepSeek
4. DeepSeek V4.1 Flash: Benchmarks, Prices, Pro Continues, Digital Applied