DeepSeek will route every V4 Pro request to V4.1 Flash from September 10, and bill you the Flash price
The new model is cheaper on every line and, DeepSeek says, better on every metric. Developers on Hacker News are not arguing with the benchmark. They are arguing with the swap.
DeepSeek said this week that V4.1 Flash will launch around September 10 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 Flash prices, off-peak from noon Beijing time on the 10th: $0.003 per million input tokens on cache hits, $0.15 on cache misses, $0.60 per million output. Peak hours are double. Every line is a cut against today's V4 Flash sheet. For anyone on V4 Pro, whose off-peak rates are $0.66 on cache misses and $1.98 on output, it is a 77 percent cut on input and 70 percent on output, whether they asked for it or not.
The objection
The Hacker News thread, 390 points and 201 comments by Wednesday afternoon, is not about capability. The top comment: "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, that LLM output is non-deterministic anyway so nobody should depend on a specific model, drew the sharpest replies. One commenter running a chatbot for an education department: "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 more formally: the 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. Structured outputs with constrained decoding remove one class of regression (malformed JSON) and none of the behavioural ones. V4 Pro is open-weights, so Together, OpenRouter and others can keep serving it. But that is a migration, not a continuation of the API contract you had on Tuesday.
Context
DeepSeek is the price floor of the market and keeps lowering it. In April it cut cache-hit input prices across its whole API to a tenth. The framing from @Yuchenj_UW this week, that o1 Pro cost $150/$600 per million tokens 18 months ago and GLM-5.3 Flash costs $0.15/$0.50 today, "a ~1000x collapse," is the environment DeepSeek competes in. Artificial Analysis's index puts DeepSeek V4 Pro (0813) at 36, behind GLM-5.3 and Kimi K3 at 44 among open-weights models. V4.1 Flash's score isn't published yet.
Our read
DeepSeek is right about the model and wrong about the contract, and it will probably keep being wrong, because nobody with pricing power has ever paid for being right about it. A model ID is an API promise. Routing it to different weights, however good, is a breaking change shipped as a favour. The honest version of this announcement is a 30-day deprecation window for deepseek-v4-pro with the Flash price applied on request, and it would have cost DeepSeek nothing except the headline.
We'd expect the swap to go through on the 10th as written, and at most a handful of visible regressions, because V4.1 Flash is probably as good as claimed. That is the problem. Every quiet success teaches the market that model IDs are suggestions, and every team that built an eval suite on the assumption that the same string returns the same distribution learns otherwise at 2am. Someone will sell the fix: we'd bet a major provider markets "pinned model versions" as an enterprise line item before the end of 2026. If DeepSeek reverses and publishes a deprecation window before the 10th, a Hacker News thread will have done what the market has not, and we'll note it.