On September 14, 2026, DeepSeek's API was scheduled to lose a model. Every request sent to the deepseek-v4-pro endpoint was set to be automatically rerouted to the newer, cheaper V4.1 Flash model starting at 04:00 UTC, with V4 Pro's premium billing swapped for Flash's lower rates. DeepSeek framed it as a simple upgrade: testing it described as coming from "multiple parties" showed V4.1 Flash beating V4 Pro on performance, cost, speed, and total runtime, so V4 Pro was being phased out.
That is not what happened. By the time the cutover deadline arrived, DeepSeek's API documentation carried a different line: V4 Pro would keep running past September 14 after all, at the same prices it always charged, "in response to user demand." It's a narrow, dated story, but a telling one. Even DeepSeek's normally blunt approach to shipping and retiring models has a limit, and developers who build production systems on a single vendor's API apparently made enough noise to reverse a decision that had already been announced in writing with a fixed deadline.
DeepSeek announced on September 9-10, 2026 that starting at 04:00 UTC on September 14, every request to its deepseek-v4-pro API would be automatically rerouted to the newer V4.1 Flash model, billed at V4.1 Flash's lower rates — effectively retiring V4 Pro. After developer pushback over the short notice and the forced, silent migration, DeepSeek reversed the decision: as of September 14, its official documentation says V4 Pro API service will continue, with billing unchanged, "in response to user demand." The reprieve is open-ended, not guaranteed permanent — DeepSeek says the arrangement holds until its next-generation V4.1 Pro model ships, and it will give further notice before any future change.
The plan DeepSeek walked back
DeepSeek released V4.1 Flash on September 10, 2026, and paired the launch with a retirement notice for its existing flagship. The company's own announcement was blunt about why: "Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We're phasing out V4-Pro." The mechanism was equally direct: "Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches." Sources: [Introducing DeepSeek-V4.1-Flash](https://deepseek.com/en/news/deepseek-v4-1-flash/)
In other words, this was not a gradual sunset with an opt-in migration window. Any application, script, or agent still calling deepseek-v4-pro after the deadline would have kept working — but silently answered by a different model, at different prices, with different behavior, without the caller changing a line of code. DeepSeek had already done exactly this to its older deepseek-v4-flash and deepseek-v4-flash-vision-exp model IDs, which the company's change log confirms are permanently retired and now "temporarily routed to V4.1 Flash" with no reprieve offered. V4 Pro looked headed for the same fate.
Why developers pushed back
The core complaint was not really about whether V4.1 Flash is a good model — independent benchmark trackers and DeepSeek's own comparison charts do show it competitive with, or ahead of, V4 Pro on several coding and agentic tests. The complaint was about control. Teams that had validated prompts, evaluation harnesses, and safety filters against V4 Pro's specific behavior were told, with only a few days' notice, that their production traffic would start hitting a different model without any change on their end. A GitHub issue opened against one downstream tool tracking the announcement captured the concern plainly: users wanted confirmation the replacement would hold up to V4 Pro's standard before their integrations were switched over involuntarily. Sources: [DeepSeek plans to discontinue the V4 Pro service — GitHub issue](https://github.com/Hmbown/Codewhale/issues/6025)
That timeline reads as short by the standards developers have come to expect from major API providers, whose formal deprecation notices typically run to months rather than days, precisely so production systems have time to re-test before a swap. DeepSeek's pattern here — announce a new model, then declare the old paid tier obsolete and reroute it within days — is also not new for the company, which is what made the abruptness of doing it to a model developers were actively paying a premium for the part that generated real pushback, rather than the routine grumbling that follows most model updates.
The reversal, in DeepSeek's own words
DeepSeek's current API change log and pricing pages both now carry the same sentence: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged." Outlets that track DeepSeek's API changelog closely reported the reversal appeared on DeepSeek's site before the scheduled cutover took effect, meaning V4 Pro reportedly never actually stopped answering requests under its own name. Sources: [DeepSeek V4 Pro API Call Service Will Not Be Discontinued — PANews](https://panews.io/articles/01a09045-be46-7632-8fd6-e486b08c3c21)
Two things are worth flagging about how DeepSeek framed the U-turn. First, it kept the original justification intact — DeepSeek has not walked back its claim that V4.1 Flash outperforms V4 Pro, only its decision to force everyone onto it. Second, the continuation is explicitly conditional: it lasts until V4.1 Pro, a full successor to the current Pro tier, actually ships, at which point this conversation will likely restart. Developers reading the reversal as a permanent reprieve rather than a postponement are reading it too optimistically.
V4 Pro vs. V4.1 Flash, by the numbers
The pricing gap is the clearest evidence for why DeepSeek wanted to retire V4 Pro in the first place, and why keeping both models live indefinitely is unusual for the company. DeepSeek bills both models on a peak/off-peak schedule, and V4.1 Flash undercuts V4 Pro at every tier.
| Spec | DeepSeek V4 Pro | DeepSeek V4.1 Flash |
|---|---|---|
| Status as of Sept 14, 2026 | Still live, billing unchanged | Live since Sept 10, 2026 |
| Context window | 1M tokens | 1M tokens |
| Vision / image input | Not supported | Supported |
| Input, cache hit (off-peak/peak, per 1M tokens) | $0.022 / $0.044 | $0.003 / $0.006 |
| Input, cache miss (off-peak/peak, per 1M tokens) | $0.66 / $1.32 | $0.15 / $0.30 |
| Output (off-peak/peak, per 1M tokens) | $1.98 / $3.96 | $0.60 / $1.20 |
At standard off-peak, cache-miss rates, V4 Pro's output price of $1.98 per million tokens runs more than 3x V4.1 Flash's $0.60, and its cache-hit input price of $0.022 is roughly 7x Flash's $0.003. Flash also does something Pro cannot: it accepts image input natively, while V4 Pro is text-only. Third-party model trackers put V4.1 Flash's architecture at roughly 552 billion total parameters in a mixture-of-experts design, with only a small fraction active per token, which is consistent with DeepSeek's own cost and speed claims for the model relative to Pro. Sources: [Models & Pricing — DeepSeek API Docs](https://api-docs.deepseek.com/quick_start/pricing/)
DeepSeek prices its API on a fixed weekly schedule: peak hours run 01:00-04:00 UTC and 06:00-10:00 UTC on weekdays, with every other hour billed as off-peak at half the peak rate, for both models. Batching non-urgent jobs outside those two weekday windows cuts costs further regardless of which model handles them.
What this means if you build on DeepSeek's API
- If you already have a validated integration calling deepseek-v4-pro, no action is required right now — it's still live, at the price you were already paying, with no forced cutover date on the books.
- For new projects, or anything cost-sensitive, V4.1 Flash is the more logical default: it's several times cheaper across the board and adds vision support V4 Pro lacks.
- Treat V4 Pro's survival as provisional, not permanent. DeepSeek has said the current arrangement holds only until V4.1 Pro ships, and hasn't committed to a long-term support window for V4 Pro after that — keep a migration plan ready rather than assuming today's pricing page is a permanent decision.
- Pin model versions explicitly wherever DeepSeek's API allows it, and monitor the change log directly rather than relying on default routing behavior — the company has now shown it will silently reroute a retired model ID to a successor without any version bump on your end.
This is a narrow episode, but it's a useful data point on how DeepSeek runs its API business compared with the more conservative deprecation practices at larger US labs: fast to ship, fast to try to retire the previous generation, and — this time, at least — fast enough to reverse course when the people actually paying the bills objected.
01Is DeepSeek V4 Pro still available through the API after September 14, 2026?
Yes. DeepSeek's official change log and pricing pages confirm V4 Pro API service continues past that date, with billing unchanged from before the reversal.
02Why did DeepSeek originally try to retire V4 Pro?
DeepSeek said internal and third-party testing showed its newer V4.1 Flash model, released September 10, 2026, beating V4 Pro on performance, cost, speed, and total runtime, making the pricier Pro tier redundant in the company's view.
03Is V4 Pro's reprieve permanent?
DeepSeek hasn't said so. Its documentation ties the continuation to the future release of V4.1 Pro and says it will give further notice before any future change, which suggests this is a postponement rather than a cancellation of the retirement plan.
04Should I migrate from V4 Pro to V4.1 Flash anyway?
For most new or cost-sensitive workloads, yes — V4.1 Flash is priced several times lower per million tokens at every tier and supports image input that V4 Pro doesn't. Teams with production systems tightly tuned to V4 Pro's specific behavior may reasonably wait, but shouldn't assume V4 Pro pricing is locked in indefinitely.
05What happened to the older V4 Flash model?
Unlike V4 Pro, the original V4 Flash and V4 Flash Vision Exp model IDs weren't spared. DeepSeek's change log confirms both are retired and permanently rerouted to V4.1 Flash, with no reversal offered for those names.
Keeping up with fast-moving, model-by-model pricing and routing changes like this one across every provider gets tedious fast. LumiChats lets you compare DeepSeek alongside ChatGPT, Claude, Gemini, and other leading chatbots side by side, and chat with several of them from one place, so a change like this is easier to spot before it quietly affects the model you're actually using.
