DeepSeek v4.1 Flash launches cheaper and outperforms v4 Pro, company says

DeepSeek announced the launch of v4.1 Flash, scheduled for September 10, 2026 (Beijing time). According to the company itself, the new model "comprehensively surpassed V4 Pro across all key metrics, including performance, cost, speed, and task completion time," that is, it surpassed the previous Pro across every indicator that matters to those putting a model into production: performance, cost, speed, and task completion time.

The information didn't come from a blog post or a paper. As user oefrha noted in the Hacker News thread, the source is understated: "Source is apparently a banner announcement on https://platform.deepseek.com/usage. Had me searching for a couple minutes...". A banner on the platform's usage dashboard, not a formal announcement.

The detail that stands out: Pro becomes Flash on the bill

The most unusual point of the announcement is what DeepSeek will do between the Flash launch and the arrival of the future v4.1 Pro. According to the announcement:

In keeping with our commitment to user responsibility, following the official launch of V4.1 Flash and prior to the release of V4.1 Pro, all requests to the Pro model will be routed to V4.1 Flash and billed at Flash's price.

>

-- DeepSeek, via platform banner

In practice: whoever calls the v4 Pro endpoint will now be served by v4.1 Flash and pay the Flash price, which is cheaper. For those who already have code pointing to Pro, this means a better model and a smaller bill without changing a single configuration line, even though there's no option to keep using the old Pro during this window.

This move didn't go unnoticed. In the thread, nicce reacted to the passage with irony about American competitors: "Wow. Imagine OpenAI/Google/Anthropic doing this! Nope." It's the comparison developers make when closing a budget: the big providers rarely lower the price of a call automatically when they launch something better.

Flash pricing starting September 10

DeepSeek also announced the adjustment to the Flash series pricing table, effective from 12:00 pm Beijing time on September 10, 2026. The values below are for off-peak hours; during peak hours, the company says prices double.

| Component | Off-peak price (per 1M tokens) | Peak price | |---|---|---| | Input, cache hit | $0.003 | $0.006 | | Input, cache miss | $0.15 | $0.30 | | Output | $0.6 | $1.2 |

The cache structure matters for the wallet: repeated calls that hit the input cache cost $0.003 per million tokens, fifty times less than a cache miss. For applications with long, stable system prompts (agents, RAG with fixed context, standardized few-shot), the difference between reusing the cache and not reusing it changes the order of magnitude of the monthly bill.

The time-based variable pricing model is another point that requires attention from whoever operates the system. Since peak hours double the cost, workloads that can be scheduled (embedding batches, asynchronous processing, overnight jobs) get a direct incentive to run outside Beijing's peak hours, which, for Brazil, falls into specific windows of our day because of the time zone.

What changes for those building software in Brazil

The practical takeaway is inference cost. Model choice in Brazil is rarely about the absolute top of quality: it's about the relationship between cost per token, latency, and "good enough" for the task. A Flash that surpasses the previous Pro at a lower price changes exactly that math, and puts pressure on the pricing bar that all providers practice.

The cadence is what the market is watching. swiftcoder summed up the community's expectation: "If they can keep up this cadence of Flash leap-frogging the previous Pro, we're in for a good time". If each Flash generation leapfrogs the previous Pro, the developer who standardizes on Flash tends to reap cost and capability gains each cycle without migrating code.

But there are caveats that remain open, and they're worth noting without optimism:

• Communication came via banner, not documentation. Without a formal paper or changelog, the claims of surpassing "across all metrics" are a company statement, not a verifiable independent benchmark.
• API access may be confusing at first. User nickweb pointed out, citing a post on X, that "the new model can be used if summoned via the API but the API won't list it", meaning the model would work through the API but wouldn't appear in the listing, which complicates automation and discovery of available models.
• Open weights are not confirmed. In the thread, tarruda hopes that "it will be open weights and have the same architecture and size as the current v4 flash vision, which is probably the best LLM that can be run on 128G devices". For those thinking about running it locally (privacy, fixed cost, no API dependency), the availability of open weights is what decides whether the model makes it into the self-hosted stack, and that hasn't been announced yet.

For Brazilian teams already using DeepSeek via API, the takeaway reading is direct: run your own comparative test between v4 Pro and v4.1 Flash on the application's real task, measuring output quality, latency, and cost per call during the hours the operation actually uses. The company itself asks for this feedback in the announcement, saying to report any issues found in the comparative test. As always, a vendor's benchmark number is a starting point, not a verdict.