DeepSeek plans V4.1 Flash for September 10th and will route Pro traffic to it

DeepSeek will release V4.1 Flash around September 10th in Beijing and temporarily route all V4 Pro API requests to the cheaper model until V4.1 Pro arrives, cutting token prices for Pro workloads by more than 70%. The company claims V4.1 Flash outperforms V4 Pro in internal and external tests, but independent evaluations are pending. New pricing takes effect at 4:00 UTC on September 10th, with off-peak rates for cached input at $0.003, uncached input at $0.15, and output at $0.60 per million tokens.

DeepSeek plans V4.1 Flash for September 10th and will route Pro traffic to it The temporary backend swap will give Pro API users the new model at Flash rates while DeepSeek prepares V4.1 Pro. By Ryan Merket /author/ryan-merket · Published Why it matters DeepSeek is using a cheaper model to replace its premium API tier, forcing developers onto V4.1 Flash while cutting some Pro workloads' token prices by more than 70%. DeepSeek https://www.deepseek.com/en/transparency/ plans to release V4.1 Flash around September 10th in Beijing and temporarily route every V4 Pro API request to the cheaper model until V4.1 Pro arrives. The switch amounts to an early retirement for DeepSeek's current premium endpoint. DeepSeek says V4.1 Flash beat V4 Pro across performance, inference cost, generation speed and total task completion time in internal and external testing. Those comparisons remain company claims until independent evaluations establish how the model performs across coding, reasoning, agent and multimodal workloads. An early look from Open Design https://open-design.ai/llm-arena-for-design/ offers one narrow comparison focused on design tasks. Its arena scored DeepSeek V4.1 Flash at 81.2 out of 100, behind GPT-6 Astra at 82.7 and ahead of GPT-5.6 Sol /models/openai/gpt-5.6-sol at 77.6. The displayed results put V4.1 Flash at 28.4 out of 30 for requirement fulfillment, 52.8 out of 70 for design quality and a 57.7% delivery rate, with a 5.3-minute completion time and a reported cost of $0.023 per run. The comparison is an early benchmark rather than a broad test of the coding, reasoning and agent workloads that will determine whether Flash can replace Pro in production. DeepSeek will preserve the deepseek-v4-pro endpoint during the transition, but requests sent to it will be served by V4.1 Flash and billed at Flash prices. Developers will therefore receive a different underlying model without changing the model name in their applications. That arrangement cuts bills for current Pro users, though it also turns a routine model alias into a forced migration. Applications tuned around V4 Pro's output, tool use or reasoning behavior will need to be tested against V4.1 Flash even if their API configuration remains unchanged. Founder Liang Wenfeng https://apnews.com/article/deepseek-founder-liang-wenfeng-china-ai-0673d5c39d90108189cc31b88d85b9f8 built DeepSeek out of the machine-learning work of quantitative hedge fund High-Flyer. Since founding DeepSeek in 2023, Liang has pushed a model strategy centered on open releases, aggressive inference pricing and smaller architectures that compete with far larger systems. V4.1 Flash extends that bet by putting the lower-cost model in front of customers who had been paying for DeepSeek's flagship tier. Flash gets cheaper again DeepSeek's new prices take effect at 4:00 UTC on September 10th, equivalent to noon in Beijing, midnight Eastern time and 9 p.m. Pacific time on September 9th. The new rates per 1 million tokens are: | Billing item | Off-peak | Peak | |---|---|---| | Input, cache hit | $0.003 | $0.006 | | Input, cache miss | $0.15 | $0.30 | | Output | $0.60 | $1.20 | Peak hours run from 1:00 a.m. to 4:00 a.m. UTC and 6:00 a.m. to 10:00 a.m. UTC, Monday through Friday. All other periods use the off-peak rate. DeepSeek's pricing documentation https://api-docs.deepseek.com/quick start/pricing/ on September 9th still listed V4 Flash at $0.007 for cached input, $0.22 for uncached input and $0.66 for output during off-peak hours. The V4.1 schedule reduces those lines by about 57%, 32% and 9%, respectively. Peak rates receive the same proportional reductions. The difference is larger for workloads currently pointed at V4 Pro. Its listed off-peak rates are $0.022 for cached input, $0.66 for uncached input and $1.98 for output. Routing those requests to V4.1 Flash lowers the corresponding prices by roughly 86%, 77% and 70% during the period before V4.1 Pro launches. For agent developers, the output reduction is only part of the calculation. Long-running coding and research agents repeatedly submit large context windows and can generate millions of cache-hit tokens. Cutting cached input from $0.007 to $0.003 per million gives DeepSeek another lever to attract persistent agent workloads, where repeated prompts and tool histories can dominate consumption. A two-day test before the switch DeepSeek opened an interim V4.1 Flash build for limited testing on September 8th. The temporary model identifier, deepseek-v4.1-flash-expires-on-0910 , worked through the existing API address, carried V4 Flash pricing and was capped at 20 concurrent requests per account, according to TechNode's report on the beta https://technode.com/2026/09/09/deepseek-v4-1-flash-multimodal-limited-beta/ . The test model uses a new architecture with native multimodal support, according to a notice reported by IT Home https://www.ithome.com/0/999/795.htm . DeepSeek also asked testers whether the interim model could replace the production V4 Pro service, a question the routing plan has now answered at the product level. The release follows a fast sequence of V4 updates. DeepSeek introduced the V4 preview on April 24th https://deepseek.com/en/news/v4-preview/ with a 1.6 trillion-parameter Pro model and a 284 billion-parameter Flash model, both supporting a 1 million-token context window. DeepSeek moved an updated Flash API into public beta on July 31st and released the experimental V4 Flash Vision model https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp on August 21st. V4.1 Flash compresses that product line again. DeepSeek is betting that a faster and cheaper model can absorb the premium tier's work before the next Pro model is ready. For customers, the immediate result is a lower bill and a mandatory change in model behavior arriving under an existing endpoint.