The model told you when it would die before it told you what it was. At about 3 p.m. Beijing time on September 8, 2026, DeepSeek posted a notice in its official user community group opening an API model id called `deepseek-v4.1-flash-expires-on-0910` — an intermediate checkpoint of an unreleased model, billed at existing V4-Flash rates, capped at 20 concurrent requests per account and, per the name, scheduled to stop answering on September 10. No blog post, no X announcement, no changelog entry.
A day later a notice on DeepSeek's open platform, the signed-in developer console, turned a curiosity into a story: the company plans to officially release V4.1 Flash around September 10 Beijing time, and says that in internal and external testing the model comprehensively surpasses V4 Pro on performance, cost, speed and total time.
That claim is the one worth staring at. V4 Pro is DeepSeek's flagship: 1.6 trillion total parameters, 49 billion active, per the April 24 V4 Preview announcement. V4-Flash is 284 billion total and 13 billion active, with the line's shared 1M-token context, built on token-wise compression plus DeepSeek Sparse Attention. DeepSeek is asserting that a Flash-class successor, a fraction the size on paper, now beats its own frontier tier — with no benchmark table, no parameter count for 4.1 and no model card to support it.
The company also said what it will do about it. Once V4.1 Flash goes live, and until a V4.1 Pro exists, every request to the `deepseek-v4-pro` endpoint routes to V4.1 Flash and bills at V4.1 Flash prices — the first time DeepSeek has publicly named a V4.1 Pro. It also amounts to a silent swap of production traffic: customers who provisioned the flagship get the small model, whether they asked or not.
The price is the actual weapon
New Flash-series rates take effect September 10 at noon Beijing time. Per the relayed notice, off-peak cached input drops from 0.05 to 0.02 yuan per million tokens, a 60 percent cut; uncached input falls from 1.50 to 1.00 yuan, down 33 percent; output moves from 4.50 to 4.00 yuan, down 11 percent. Peak is double off-peak. In dollars, the current V4-Flash card reads roughly $0.22 per million input and $0.66 per million output off-peak, cache hits at $0.007.
The steepest cut lands where agentic workloads live: cached input, the tokens a long-running agent re-reads every turn. For anyone paying V4 Pro rates — 0.15 yuan cached, 4.5 uncached, 13.5 output off-peak — the routing change is arithmetic more than architecture. CellCog founder Nitish Garg calculates that if Pro requests are served by V4.1 Flash at the new rates, uncached input falls 78 percent and output 70 percent. His caveat is the right one: “That is the claim to test on September 10, and it is DeepSeek's claim, from its own internal and external runs, not ours.”
Everything else is community telemetry. Machine Heart relayed self-reported throughput near 400 tokens per second and, from X user @NFT_Chen, speedups of 3.9x to 6x versus V4-Flash-Vision-Exp — self-tests against an expiring checkpoint whose 20-request cap, roughly 1 percent of production concurrency, rules out any real load test.
The other unverified claim is structural: a new model structure with native multimodal support, text, image and speech in one unified model. That would depart from V4-Flash-Vision-Exp, open-weighted under MIT on August 31, whose Hugging Face card describes “incorporating visual modules and undergoing continued training to unlock visual understanding capabilities.” Built-in versus bolted-on is a real distinction, but as of September 9 the deepseek-ai organization carried no 4.1 repository at all.
Analysis: cadence, price pressure and a bad week to be quiet
DeepSeek's 2026 has been metronomic: V4 Preview April 24, V4-Flash-0731 July 31, V4-Pro GA August 13, Vision-Exp August 21, its weights August 31 — the last three landing 13 and 8 days apart. The V4.1 test id arrived 18 days later. If a structural change ships now, it is the fastest architectural update DeepSeek has made, and the first previewed through an expiring endpoint rather than simply shipped.
The cadence is itself the competitive instrument. Western labs price against quarterly releases; DeepSeek is repricing its catalog twice in a month, having raised V4 prices sharply in August only to cut Flash rates now. Pricing floors do not survive a competitor that treats its rate card as a release artifact. Its own tables have shown small models punching up before: the V4-Flash-0731 card reported 82.7 on Terminal Bench 2.1, 54.4 on DeepSWE and 70.3 on Toolathlon-Verified, beating the V4-Pro preview at a fraction of the active parameters.
It arrives at an awkward moment. On September 8, the day the endpoint opened, the NSA, CISA and FBI published joint advisory AA26-251A naming DeepSeek alongside Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as running industrial-scale distillation campaigns against US frontier models since at least late 2024. It asserts DeepSeek's widely cited $5.6 million training cost is misleading because it excludes data acquired through malicious distillation, and tells providers to watch for “24/7 sustained usage without human variation/idle periods.” Nothing in it addresses V4.1 Flash. But a lab claiming a dramatic efficiency jump with no technical report, in the same week it is accused of manufacturing efficiency by copying, has picked the hardest possible venue in which to be unfalsifiable. Reviewing the beta, explainx.ai's Yash Thakker wrote that “this reads like an internal test DeepSeek opened to a wider pool of testers than intended to keep private, not a product launch.”
What to watch
Four things decide whether this is a release or a rumor with a rate card. An entry on the API news page, which as of September 9 still ended at Vision-Exp. A permanent model id without an expiry suffix. Weights and a card on Hugging Face, which would give the first real parameter count and either confirm or retire the native-multimodal line. And the routing itself: whether `deepseek-v4-pro` starts returning V4.1 Flash output from launch day, and how long before V4.1 Pro turns out to be. Until a benchmark table exists, comprehensively surpasses is DeepSeek's word about DeepSeek's model, tested by DeepSeek.
“That is the claim to test on September 10, and it is DeepSeek's claim, from its own internal and external runs, not ours.”— Nitish Garg, Founder and CEO, CellCog