On September 9, DeepSeek released V4.1 Flash for general availability, a 552-billion-parameter MoE model built on a new Causal-Encoder-Decoder architecture. The same announcement said V4 Pro would be retired on September 14, with all requests automatically routed to V4.1 Flash and billed at Flash rates.
Developers had roughly four days of notice. There was no migration window, no parallel operation period, no guidance on how prompts and agent workflows tuned for V4 Pro would behave on the new model.
By September 11, DeepSeek reversed the decision. V4 Pro would stay online. The automatic migration was cancelled.
Three changes to the retirement schedule in under 72 hours.
Faster Per Token, More Expensive Per Session
The pricing math looked simple on paper. V4.1 Flash costs less per token than V4 Pro. Output dropped roughly 33%, cache-miss input fell 67%. A direct cost reduction.
Early beta testers reported something different. V4.1 Flash is faster — peak throughput around 507 tokens per second, compared to roughly 97 tokens per second for the older V4 Flash. But speed changes consumption behavior. Same wall-clock time, more tokens generated.
“Five minutes, ten yuan,” one developer wrote. Another: “Dozens of yuan gone in an instant”.
The official cost reduction is real per token. The developer's bill is not necessarily lower. For long-context agent workloads with high cache-hit rates, Flash wins clearly. For generation-heavy tasks with low cache reuse, the faster model can cost more in total, not less.
DeepSeek's efficiency narrative was built for infrastructure — HBM demand cut to one-quarter, SSD to one-eighth. Developers experience the other side of that equation: their bill.
Four Days' Notice Is Not a Migration Window
The most pointed criticism came from developers who had already integrated V4 Pro into production systems.
Prompt engineering is not portable. A prompt tuned for one model's quirks — how it handles edge cases, when it refuses, how it structures tool calls — may behave differently on another model, even if the new model scores higher on benchmarks.
One developer's summary of the problem: “You really need to get some proper software engineering training”.
The reference was to version control. In normal software practice, you do not replace a production dependency without parallel operation, a deprecation window, and rollback capability. DeepSeek planned to swap the model underneath live applications with four days of notice.
A research team working with DeepSeek-V4-Pro-0813 said their papers were unfinished. Once the endpoint was gone, experiments could not be replicated under the same conditions.
“Listening” Happens After the Damage
Chinese tech media described the V4.1 Flash release as “listening to advice” — a turnaround from V4 Pro's troubled launch.
V4 Pro had a rough month. Launched in August, developers immediately reported degraded reasoning on long tasks and premature termination on complex workflows. Within hours of wide release, reports of “dumbing down” appeared. Then came the August price hike: some billing items rose by as much as 1,100%. Founder Liang Wenfeng's reputation in developer communities shifted from “Liang the Saint” to “Liang the Problem”.
V4.1 Flash was the correction. Smaller, faster, cheaper, multimodal native. It killed the model developers had complained about and replaced it with something better.
Then DeepSeek tried to force the migration with four days' notice — and had to reverse again.
The pattern is consistent: make a change, receive backlash, adjust. “Listening” happens after the damage, not before the decision.
The Next Swap Will Look the Same
DeepSeek kept V4 Pro alive. That solves the immediate problem for developers running production systems.
It does not solve the underlying issue. There is still no published deprecation policy. No guaranteed migration window. No commitment to parallel model endpoints when replacements ship. The next time DeepSeek releases a model it believes is better, developers have no reason to expect different treatment.
DeepSeek is preparing for a STAR Market IPO in Shanghai, and has been actively hiring backend and infrastructure engineers — 150 roles, almost none touching model training. The company understands it needs to become a platform, not just a model lab.
Platforms have versioning standards. DeepSeek just demonstrated it doesn't have them yet.
P.S. V4 Pro stays online at unchanged pricing. DeepSeek has not said for how long, or what happens when V4.1 Pro ships. Developers running production workloads now have to plan around a model with no published end-of-life date — and a vendor with a three-day track record of changing its mind.
Frequently Asked Questions
Q: What did DeepSeek announce on September 9?
A: DeepSeek released V4.1 Flash for general availability and said V4 Pro would be retired on September 14. All V4 Pro requests would be automatically routed to V4.1 Flash and billed at Flash rates. Developers had roughly four days of notice.
Q: Why did DeepSeek reverse the decision?
A: Developers pushed back, citing no migration window, no parallel operation period, and no guidance on how prompts tuned for V4 Pro would behave on V4.1 Flash. By September 11, DeepSeek cancelled the automatic migration and kept V4 Pro online.
Q: Is V4.1 Flash actually cheaper?
A: Per token, yes. Output pricing dropped roughly 33%, and cache-miss input fell 67%. But V4.1 Flash is faster — peak throughput around 507 tokens per second versus roughly 97 for the older V4 Flash. Faster generation means more tokens consumed in the same wall-clock time, so some developers report higher total session costs.
Q: What did developers complain about most?
A: The lack of software engineering discipline. In normal software practice, a production dependency is not replaced without parallel operation, a deprecation window, and rollback capability. DeepSeek planned to swap the model underneath live applications with four days of notice.
Q: What about research teams using V4 Pro?
A: A research team working with DeepSeek-V4-Pro-0813 said their papers were unfinished. Once the endpoint was removed, experiments could not be replicated under the same conditions.
Q: What happened with V4 Pro's launch?
A: V4 Pro launched in August and drew immediate complaints about degraded reasoning on long tasks and premature termination on complex workflows. Then an August price hike raised some billing items by as much as 1,100%. Founder Liang Wenfeng's reputation in developer communities shifted from “Liang the Saint” to “Liang the Problem”.
Q: What is the underlying issue DeepSeek hasn't solved?
A: There is still no published deprecation policy, no guaranteed migration window, and no commitment to parallel model endpoints when replacements ship. The next time DeepSeek releases a model it believes is better, developers have no reason to expect different treatment.
Q: How does this connect to DeepSeek's IPO?
A: DeepSeek is preparing for a STAR Market IPO in Shanghai and has been hiring backend and infrastructure engineers — 150 roles, almost none touching model training. The company understands it needs to become a platform. Platforms have versioning standards. DeepSeek just demonstrated it doesn't have them yet.
Q: When will V4 Pro actually be retired?
A: DeepSeek has not said. V4 Pro stays online at unchanged pricing, with no published end-of-life date. Developers running production workloads now have to plan around a model with no stated timeline — and a vendor with a three-day track record of changing its mind.
