DeepSeek will route every V4 Pro request to V4.1 Flash, and bill you the Flash price
The new model is cheaper on every line and, DeepSeek says, better on every metric, but developers are arguing with the swap rather than the benchmark.
A model ID is a promise. DeepSeek is about to break one politely, and hand you a discount for the trouble.
The company says V4.1 Flash launches around 10 September 2026 Beijing time, that it "has comprehensively surpassed V4 Pro across all key metrics, including performance, cost, speed, and task completion time", and that until V4.1 Pro ships, "all requests to the Pro model will be routed to V4.1 Flash and billed at Flash's price."
鉴于 DS V4.1 Flash 模型在性能、费用、速度、总用时等各项指标上都全面超越了 V4 Pro。再以更高的价格、更慢的速度和更多的算力消耗给 DS 用户提供原有的性能较差的 V4 Pro 模型会不太合适。V4.1 Flash 正式上线之后，V4.1 Pro 上线之前，V4 Pro 模型的请求都会被路由到 V4.1 Flash，并按 Flash 的价格计费。
The new off-peak prices are $0.15 per million input tokens on a cache miss and $0.60 per million output, with peak hours charged at double that (the working day in Beijing costs more than the night). Every line is a cut. Anyone currently paying V4 Pro's off-peak rates, $0.66 and $1.98, is about to pay roughly three-quarters less for input and seventy percent less for output, whether they asked to or not.
The objection is not about capability
The Hacker News thread, past 390 points by the middle of the week, argues about none of the benchmarks. Its top comment argues about the swap. "If a user has validated a workflow on V4 Pro, they might not want to suddenly start testing it in production on V4.1 Flash. Instead, keep V4 Pro around but deprecated for a defined period of time, then remove it."
The counter-argument in the thread is that model output is non-deterministic anyway, so nobody should depend on specific weights. It drew the sharpest replies of the day. One commenter running a chatbot for an education department pointed out that "every other model upgrade has a measurably different adherence to the existing system prompts about not talking to the kids about sex and drugs and mental health issues."
Another put it formally. A user "cares about the distribution of outputs", and "changing the underlying weights necessarily invalidates whatever characterization may have been done."
Can you mitigate it? Partly. Constrained decoding removes one class of regression, the malformed output, and none of the behavioural ones. V4 Pro is open-weights, so other providers can keep serving it. But that is a migration, not a continuation of the contract you had on Tuesday.
The price environment explains the behaviour without excusing it
DeepSeek is the floor of this market and keeps lowering it. In April it cut cache-hit input prices across its entire API to a tenth of what they were. One widely shared comparison this week set o1 Pro's old rates against a current Flash-class model and called it a thousandfold collapse in eighteen months, which is roughly the water everyone here is swimming in. We looked at the other end of the same pricing strategy in DeepSeek's V4.1-Flash costs $0.15 per million tokens for four-fifths of the week.
On capability, Artificial Analysis puts DeepSeek V4 Pro at 36 on its index, behind GLM-5.3 and Kimi K3 at 44 among open-weights models. V4.1 Flash has no published score yet, which means the claim to have surpassed Pro is, for now, DeepSeek's own.
Right about the model, wrong about the contract
Routing a model ID to different weights, however good those weights are, is a breaking change shipped as a favour. The honest version of this announcement is a thirty-day deprecation window for the Pro endpoint with the Flash price applied on request, and it would have cost DeepSeek nothing except the headline.
We would expect the swap to go through on the tenth as written, with at most a handful of visible regressions, because the new model is probably as good as claimed.
Which is the actual problem. Every quiet success teaches the market that model IDs are suggestions, and every team that built an evaluation suite on the assumption that the same string returns the same distribution finds out otherwise at two in the morning.
Somebody will sell the fix. We would bet a major provider markets pinned model versions as an enterprise line item before the end of 2026. And if DeepSeek publishes a deprecation window before the tenth instead, a forum thread will have done what the market has not, and we will say so.
