DeepSeek Releases V4.1 Flash, Cuts API Prices and Sets End Date for V4 Pro
DeepSeek officially released V4.1 Flash, a cheaper and faster multimodal successor to V4 Pro with a 1-million-token context window, and said it will reroute all V4 Pro API traffic to the new model from September 14.
DeepSeek officially released DeepSeek-V4.1-Flash on September 10, a lower-cost, faster successor to V4 Pro built to handle both text and images natively rather than through a bolted-on vision module.
What changed under the hood
V4.1 Flash is a sparse mixture-of-experts model with a 1,048,576-token context window and support for up to 384,000 tokens of output. According to MarkTechPost's technical write-up, the model brings native multimodal input into its base architecture and uses several efficiency techniques — including an FP4 key-value cache and cross-layer attention reuse — to cut the memory cost of holding long contexts in memory compared with V4 Flash. DeepSeek said internal and external testing showed V4.1 Flash beating V4 Pro on performance, cost, speed and total task time, despite Flash traditionally being positioned as the cheaper, lighter-weight tier.
Pricing moves and V4 Pro's retirement
Alongside the release, DeepSeek cut Flash's API pricing, with reporting from Intelligent Living and OpenRouter's own listings pegging the new uncached input rate at a fraction of what V4 Pro charged, plus a steep discount on cache-hit input tokens. DeepSeek also set a firm transition date: starting at 12:00 Beijing time on September 14, all API requests still targeting deepseek-v4-pro will be automatically routed to V4.1 Flash and billed at Flash's lower price, effectively retiring V4 Pro until a V4.1 Pro model arrives.
Why it matters
The release keeps DeepSeek in the middle of the pricing war among frontier and near-frontier labs: rather than shipping a flagship "Pro" upgrade, it pushed its efficiency gains into the cheaper Flash tier and is forcing existing Pro customers onto it. That mirrors a broader pattern this quarter of labs competing less on headline benchmark scores — which are converging across Claude, GPT-6 Astra, Gemini and DeepSeek's own models — and more on cost per token, context-window efficiency and how gracefully a model handles very long inputs. For developers building on open-weight or budget-tier models, it also lowers the cost of long-context workloads like large-codebase analysis or document review that previously required a pricier "Pro"-class model.
Sources
AI-assisted reporting, overseen by the AgentsAI team. Spotted an error? Let us know.
More ai news
Google Unveils Gemini 4 Argon, Claims Benchmark Lead but Limits Access to Cyber Defenders
Google's new top-tier Gemini 4 model claims leads over OpenAI's GPT-6 Astra and Anthropic's Opus on most disclosed benchmarks, but it is initially available only to vetted defenders through the Fairwind Program.
EU Set to Propose Barring Under-15s From AI Chatbots and Social Media in 'Kids Act'
The European Commission is preparing to unveil an EU Kids Act that would bar unsupervised access to AI chatbots, social media, video platforms and online games for under-15s, with tiered rules and mandatory age verification for 13-14 year-olds.
Amodei's 'Pace the Frontier' Plan Draws Same-Day Backing From OpenAI, DeepMind and xAI
Anthropic CEO Dario Amodei published an essay arguing frontier AI labs should deliberately slow capability gains, and within hours Sam Altman, Demis Hassabis and Elon Musk publicly endorsed the idea, with Microsoft's Satya Nadella following a day later.
Positron Raises $875M to Build an HBM-Free AI Inference Chip
Chip startup Positron closed an $875 million Series C at a $5 billion post-money valuation to fund its Asimov inference accelerator, which pairs its compute architecture with up to 2,304GB of commodity LPDDR5X memory instead of scarce high-bandwidth memory.