15 SEP 2026 — DeepSeek told its API customers on 10 September that from noon Beijing time on 14 September, every request sent to its V4 Pro model would be answered by the new V4.1 Flash instead, and billed at Flash rates. By the deadline it had changed its mind. Its change log now says the company will keep serving V4 Pro after 14 September, with billing unchanged.
The reversal itself is minor. The plan behind it deserves more attention than the price of either model, because of why developers objected.
What DeepSeek announced
V4.1 Flash arrived by notification to API users on 10 September. DeepSeek said the model had surpassed V4 Pro in performance, cost, speed and total completion time in internal and external testing, and priced it low: at off-peak rates, RMB 0.02 per million input tokens on a cache hit, RMB 1 on a cache miss and RMB 4 per million output tokens, with peak-hour rates double those figures.
On that basis the company proposed a clean switch. Requests addressed to deepseek-v4-pro would be routed to V4.1 Flash from 14 September, and customers would pay Flash prices for them. The model names deepseek-v4-flash and deepseek-v4-flash-vision-exp were also pointed at the new model.
DeepSeek's reasoning is easy to follow. A cheaper model that it believes is better serves more users on the same hardware, and one model is simpler to run than two.
Capacity matters too, if less visibly. Every model a provider keeps online occupies accelerators whether or not it is busy, and a large model reserved for a shrinking share of traffic is expensive idle hardware. Consolidating traffic on the newer, cheaper model frees that capacity for the model the company wants customers on. That is a legitimate operational goal, but it is the provider's goal, and the customer carries the risk of meeting it.
How a better model still breaks things
V4.1 Flash may well be better on most tasks. Developers objected anyway, because a model name is a contract.
A developer who builds on deepseek-v4-pro has tuned prompts against it, measured its outputs on their own evaluation sets and set cost budgets from its behaviour. Swap the model behind that name and every one of those measurements goes stale at once, silently, with no code change on the customer's side to trigger a review. Output formats drift, refusal behaviour shifts, a prompt that relied on a particular habit of the old model stops working. A system can get better on a benchmark and worse on the one task a customer actually runs.
This is the reason serious API providers version their models with dated identifiers and give notice before retiring them. The alias that always points at the latest model is a convenience; the pinned identifier is what production systems depend on. Redirecting a pinned name is the move that breaks that arrangement.
The usual practice among large model providers is to publish a deprecation schedule, announce retirements months ahead and leave the old identifier working until the date, so that customers can run their own comparisons before anything changes underneath them. Four days between the notice and the switch is not a migration window. It is barely enough time to read the notice, let alone rerun an evaluation suite across a production workload.
What stayed and what changed
The reversal is narrower than it may look. V4 Pro remains available at its existing prices. The Flash model names were still pointed at V4.1 Flash, so anyone calling those names already has the new model, whether they tested it or not.
Reporting on the reversal has attributed it to user demand. DeepSeek's change log states the decision without elaborating, and it gives no retirement date for V4 Pro. The question has been postponed, not settled.
There is also a billing asymmetry in the original plan that is easy to miss. Redirected requests would have been charged at Flash rates, which are lower, so most customers would have seen their bills fall. A falling bill is welcome, but it also signals that the work has changed. In many organisations the finance team, noticing an unexplained drop, would be the first to discover a silent model swap.
A future V4.1 Pro has been discussed in coverage; the change log we read gives it no date.
A pattern in DeepSeek's pricing
This is not the first time DeepSeek has moved quickly on the terms of its API. It has retired legacy endpoints before, and its move to peak and off-peak pricing — doubling rates during busy hours while Western providers were cutting theirs — showed a company willing to use pricing actively to shape when and how customers use it.
Aggressive pricing is a large part of why developers adopted DeepSeek in the first place. Terms that move quickly in the customer's favour can also move quickly in ways the customer did not choose. For anyone costing a DeepSeek deployment, the peak multiplier and the RMB denomination both matter; our own calculator for Chinese model token costs works through both.
What developers should take from it
Pin the model you validated, and treat any provider notice about aliases or redirects as a change request rather than an announcement. Where a provider offers only moving aliases, keep an evaluation set small enough to rerun whenever the provider says anything at all.
And plan for V4 Pro to go eventually. The reversal bought time; it did not come with a date. A migration tested on your own workload before a deadline is much cheaper than one forced on you by a redirect.
View author profile → · Editorial policy
