DeepSeek Launches V4.1 Flash and Deactivates V4 Pro by Routing Requests to the New Model

DeepSeek officially launched the V4.1 Flash model on September 10 and introduced a new pricing table for the Flash series, according to a notification sent to API users and reported by TechNode. The point that matters most for those with code in production is not the announcement itself, but the operational consequence: V4 Pro will no longer be served directly.

Starting at 12:00 pm (Beijing time) on September 14, requests sent to V4 Pro will be routed to V4.1 Flash and billed at Flash series rates. In practice, if you point your application to the V4 Pro endpoint, it will keep responding, but behind the scenes it's Flash answering, and the bill changes.

What DeepSeek Claims About Flash

According to the company, in internal and external tests the V4.1 Flash outperformed V4 Pro in performance, cost, speed, and total completion time (the total completion time, which adds up latency plus the time to generate the full response). This is the claim that justifies the automatic routing: from DeepSeek's point of view, there's no reason to keep a more expensive, slower model when the successor is better on all four fronts.

An editorial note is worth making: these numbers come from the manufacturer itself, not from an independent benchmark. Until third-party measurements appear, the "better at everything" thesis is DeepSeek's own version, and each team needs to validate it against its own use case before assuming the switch is seamless.

The New Pricing Table

Pricing for the Flash series is split between peak and off-peak hours. The figures below, provided by DeepSeek, apply to the off-peak period; during peak hours, prices double.

| Component | Off-peak price (per million tokens) | |---|---| | Input with cache hit | RMB 0.02 | | Input with cache miss | RMB 1 | | Output | RMB 4 |

During peak hours, this becomes RMB 0.04 / RMB 2 / RMB 8, respectively. The distinction between cache hit and cache miss is the detail that makes a real difference in the bill: repeated prompts or ones with a stable prefix (fixed system prompt, reused few-shot examples) fall into the RMB 0.02 per million tokens tier, fifty times cheaper than uncached input. Those who structure their calls to maximize cache hits pay a fraction of what those who resend fresh context on every request pay.

What Changes for the Brazilian Developer

From the perspective of those building software here, three practical points:

1. Forced migration, not optional. Unlike a long-timeline deprecation, here the switch is automatic and immediate starting September 14. If your application depends on specific V4 Pro behavior (output formatting, tone, response structure), the best move is to run your test suite against V4.1 Flash before that date, because after it you'll already be on the new model whether you like it or not.

2. The bill in reais depends on exchange rate and time zone. Prices are in RMB (yuan). When converting to reais, factor in exchange rate variation in your cost-per-thousand-requests calculation, and keep in mind that Beijing time defines when it's peak or off-peak, which doesn't line up with Brazilian business hours. Running batch loads scheduled for China's off-peak hours can cut costs in half, but that requires the execution window to tolerate delay.

3. The price-performance baseline shifts. With each iteration like this, the reference value teams use to decide between proprietary models and lower-cost alternatives changes. If Flash delivers what DeepSeek promises, it pushes down the acceptable cost for tasks like classification, extraction, and high-volume summarization, tasks where the quality gap between models is smaller and price per token dominates the decision.

The Ecosystem Around It

The launch didn't come in isolation. On the same day, DeepSeek released Harness 0.1.5 with V4.1 Flash support, adding file uploads and sidebar previews, a sign that the company is working on the tooling around the model, not just the model itself. V4.1 Flash had already gone through a time-limited multimodal beta before this formal release, which explains why the V4 Pro transition is happening so fast: the model isn't a blind debut.

What Remains Open

The source doesn't detail whether V4 Pro will be shut down entirely or remain only as a redirected label, nor whether there are differences in context window or feature support between the two. There is also, as of now, no independent public benchmark confirming the claimed superiority. For those operating in production, the sensible playbook is clear: test V4.1 Flash against real cases before September 14, measure cost under the new pricing table accounting for your application's cache-hit pattern, and only then decide whether to keep the DeepSeek endpoint or reassess the model mix.