你的 V4 Pro 请求，明天起会被自动换成另一个模型，而且更便宜Tomorrow Your V4 Pro Requests Get Auto-Swapped to a Cheaper Model

Yesterday we reported the Flash series price change. At the time we only told half of it — that wasn't an isolated price cut, but a companion move to a new model launch.

Yesterday We Only Told Half the Story

Yesterday we reported the Flash series price change. At the time we only told half of it — that wasn't an isolated price cut, but a companion move to a new model launch.

All three things happened on September 10:

1. DeepSeek released the V4.1 Flash model (officially worded as "around September 10, Beijing time");
2. Starting at 12:00, the Flash series moved to new pricing;
3. V4 Pro requests are automatically routed to V4.1 Flash and billed at Flash rates.

The third is the most unusual: without changing a single line of code, the server hands your V4 Pro request to a newer model and then charges you the cheaper price. An upgrade that also cuts the price — that's genuinely rare in this industry.

What the Official Notice Says

Here is the original text of the DeepSeek open platform notice:

DeepSeek plans to officially release the V4.1 Flash model around September 10, 2026, Beijing time. After extensive internal and external testing, V4.1 Flash has comprehensively surpassed V4 Pro across every metric, including performance, cost, speed, and total elapsed time. In the spirit of an attitude of responsibility toward users, after V4.1 Flash officially goes live and before V4.1 Pro launches, we will route all requests to V4 Pro to V4.1 Flash and bill them at V4.1 Flash rates. If you find any problems in comparative testing between V4 Pro and V4.1 Flash, please report them to us promptly. Thank you for your support!

One caveat: "comprehensively surpassed V4 Pro across every metric" is the official wording, which we relay verbatim without seconding it. To date, the company has published no benchmark score table for V4.1 Flash at all — and that absence is itself information.

The Boundaries of Auto-Routing: Don't Misread Them

Let's draw the lines where confusion is most likely:

• This is an interim policy, valid only "after V4.1 Flash goes live and before V4.1 Pro launches." What happens after V4.1 Pro ships, the company has not said.
• Users need no code changes at all; routing happens server-side. In other words, the model name you call is still V4 Pro, what actually runs is V4.1 Flash, and the bill is calculated at Flash rates.
• The stated reason is "an attitude of responsibility toward users" — the wording itself is notably restrained, with no marketing language whatsoever.
• This policy appears on no public documentation page; it comes from a logged-in open platform notice and an official community group, relayed by Juejin, Shanghai Securities News, ITHome, STAR Market Daily, and others.

In other words, this is an "act first, document later" move: the savings are already live, the docs haven't caught up.

The Math Is the Most Persuasive Part

For the same request, how much does the price differ when it moves from V4 Pro to V4.1 Flash?

Peak-hour prices are each tier's off-peak price ×2, and the multiple is unchanged (peak hours are Monday–Friday 9:00–12:00 and 14:00–18:00). V4 Pro's current prices come from the official pricing page; V4.1 Flash's new prices come from the official repricing notice; the reductions are derived from the two.

Combine the three tiers into one concrete scenario (the figures below are estimates based on the table above and do not represent any actual bill):

An agent task with 1 million input tokens, 80% of which hit the cache, and 50,000 output tokens.

• On V4 Pro, current off-peak price: 0.8 × 0.15 + 0.2 × 4.5 + 0.05 × 13.5 ≈ ¥1.70
• On V4.1 Flash, new off-peak price: 0.8 × 0.02 + 0.2 × 1.0 + 0.05 × 4.0 ≈ ¥0.42

For the same task, the bill drops by about three quarters. This is why "a stronger model" and "less money" can both be true at once: the savings don't come from squeezed margins, but from the price gap between the two model generations — V4 Pro's output price has always been ¥13.5, and Flash has always been in the single-digit yuan range. Routing simply settles Pro's work at Flash's rates.

While we're at it, here's yesterday's Flash repricing table again for comparison (unit: yuan / million tokens):

Two things to note: first, the -60% in the repricing table refers only to the cache-hit tier; second, that table describes what changes for Flash users themselves, whereas the V4 Pro comparison above is the real blast radius of the routing policy.

Community Tests: Some Pleasant Surprises, Some Cold Water

First the positives (all from community users testing in their own scenarios, not official figures):

• 3D "S-shaped path parking" task: V4.1 finished in about 25.4 seconds with zero collisions; V4 took 11 minutes 27 seconds and rendered the main view upside down.
• Small shop task (Node.js + shopping cart + coupon codes + idempotent checkout): V4.1 took 1 minute 41 seconds, V4 took 5 minutes 54 seconds.
• Multimodal hallucination: a blogger posted a photo of a suit, and the model answered "pinstripe suit" — zooming in confirmed it was correct.
• Speed: the community broadly reports it's noticeably faster than V4-Flash, but the circulating numbers are all over the place (claims range from 200 to 600 tok/s, with 284 / 355 / 420 / 507 each having their own source), and they all come from different people in different scenarios, so they can't corroborate each other and don't represent official performance.

Now the negatives — this part matters just as much:

• One developer ran 14 sets of task tests under the headline "it's actually more expensive than the previous generation": for the same small game, V4.1 took 34.5 minutes end to end, versus 30.5 minutes for V4. The reason is plain — fast token output doesn't mean fast delivery; the generation time saved gets eaten by tool calls and verification (18.7 minutes waiting on tools vs 6.6 minutes).
• Yicai reported that developers said "¥10 in 5 minutes" and "dozens of yuan gone in an instant," casting doubt on the "lower cost" conclusion and speculating that faster generation means burning more tokens per unit of time.
• Overthinking is still a problem: someone burned 4M tokens without finishing a single Frogger game; the Seven Wonders benchmark ran 2 hours and cost $2.60.
• Self-identity confusion: asked "who are you," it answered "Claude."
• Beta limitations: video isn't supported yet — only images were tested; some third-party integrations haven't caught up.

These two sides don't contradict each other. A model reacting faster on a single turn and a task costing less overall are two different things — if your bottleneck is waiting for tokens, the new model is a net win; if your bottleneck is tool-chain round trips and repeated verification, then fast token output may actually burn more per unit of time.

What DSH Users Should Watch

On the ecosystem side there are two confirmed data points so far: NetEase Youdao's LobsterAI announced on September 8 that it was the first to integrate the V4.1 Flash beta, the fastest move in this cycle; we have not found any DSH official announcement of dedicated support for v4.1-flash.

But one thing is worth DSH users' early attention: under the V4-native protocol, DSH uses deepseek-v4-flash by default as its high-throughput channel and automatically escalates to deepseek-v4-pro when it hits a hard problem. Once V4.1 Flash ships and Pro requests are routed server-side, that default channel and escalation path will most likely change — exactly how depends on subsequent model identifiers and protocol adjustments.

The practical advice: after September 10, run a round of your own real tasks, compare time and bills before and after the switch, and only then decide whether to touch your configuration. Someone else's benchmark scores and your workflow are usually two different things.

Notes on Sourcing and Reliability

• As of the night of September 9, there was no V4.1 Flash entry anywhere in DeepSeek's official public documentation: the api-docs pricing page still listed only three models — v4-flash, v4-pro, and v4-flash-vision-exp; the newest changelog entry was dated August 21; and the homepage headline was still V4-Pro.
• Therefore all release and routing information in this article comes from DeepSeek open platform notices and the official community group, relayed by third-party media, not from official public documentation pages. Please treat the official final announcement and the platform's actual behavior as authoritative.
• The release timing is the official phrasing "around September 10," not a fixed moment; there is only one fixed moment: the repricing takes effect at 12:00 on September 10.
• V4.1 Flash's context length, output cap, concurrency limit, and whether it supports a thinking mode are all unstated by the company. This article makes no guesses and does not carry over the previous generation's specs.
• All community test data names the tester and the scenario; it reflects observations in individual environments, does not represent official performance, and does not constitute any model-selection conclusion.

One last point: the most memorable thing about this round isn't how much the price dropped, but the fact that "the server can swap your model for you." Model iteration is now moving so fast that users changing their own code can't keep up.