DeepSeek just slipped a new model out for testing with almost no fanfare, and if early chatter is right, it might quietly outclass the company’s own current flagship. Here’s what’s actually confirmed versus what’s still rumor.
What’s Actually Confirmed
On September 8, at around 3pm Beijing time, DeepSeek posted a notice in its official user community group announcing an “intermediate version” of V4.1 Flash open for testing. The test model, listed under the API id deepseek-v4.1-flash-expires-on-0910, is billed at the same rate as the existing deepseek-v4-flash and capped at 20 concurrent requests per account. As the name suggests, the test access expires on September 10.
The relayed notice promised a new model structure, native multimodal support, more capability, more speed, and lower cost. What it didn’t include is just as notable: no benchmarks, no confirmed pricing, and no official launch date from DeepSeek itself.
The Rumor Going Around
Chinese financial outlet Cailianshe reported that DeepSeek plans to officially release V4.1 Flash around September 10 Beijing time, claiming the model comprehensively surpasses V4 Pro, DeepSeek’s current top-tier model, across performance, cost, speed, and total processing time. Worth flagging: that’s a secondhand report, not a statement from DeepSeek itself, so treat the “beats our own flagship” claim as unconfirmed for now.
Where V4 Currently Stands
For context, V4 Flash launched back in April 2026 with a hybrid attention architecture and a 1-million-token context window with up to 384K tokens of output. V4 Pro left preview and reached general availability on August 12, 2026, with the same context and output specs but noticeably higher-end pricing, alongside a very cheap cache-hit input rate for repeated queries.
Why This Matters
DeepSeek has built its reputation on shipping fast, cheap, open-weight models that keep pressure on Western labs over pricing, including recent flagship launches like GPT-6 Astra and Claude Fable 5.1, which shipped at identical price points just weeks ago. If V4.1 Flash genuinely does outperform the company’s own flagship at flash-tier pricing, that’s one more move in the ongoing price war between US and Chinese AI labs, and good news for anyone building on a budget. Meanwhile, OpenAI has been pushing its own agents toward doing research-level work, a sign of just how fast this competitive landscape keeps shifting.
We’ll likely know for sure around September 10, whenever DeepSeek decides to make things official. Worth bookmarking and checking back.
