{"slug": "deepseeks-new-architecture-slashes-agentic-costs-by-80", "title": "DeepSeek’s New Architecture Slashes Agentic Costs by 80%", "summary": "DeepSeek released V4.1 Flash on September 10, 2026, cutting cache-hit costs to $0.003 per token during off-peak hours from the $0.022 charged for the outgoing V4-Pro, a 77-80% price reduction, and raising the concurrency limit from 500 to 2,500 requests. The model uses the industry's first Causal Encoder-Decoder architecture, in which a 552B-parameter Mixture-of-Experts backbone activates only 8B parameters during prefill and 16B during decode, compressing the KV cache to 890 bytes per token — a 75% improvement over V4-Flash and 1/437th of DeepSeek V1. DeepSeek, which carries a $71 billion valuation, said that starting September 14 all requests for the retiring V4-Pro will be automatically routed to V4.1-Flash at the lower price point.", "body_md": "DeepSeek’s release of [V4.1 Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) on September 10, 2026, forces a recalibration of the economics governing autonomous agents. By slashing cache-hit costs to $0.003 per token during off-peak hours-a steep drop from the $0.022 charged for the outgoing V4-Pro-the company is aggressively attacking the primary financial bottleneck currently stifling long-running agentic loops. This is not merely a marginal improvement; it is a calculated effort to commoditize the heavy lifting of AI automation by fundamentally altering the underlying cost structure.\n\nThe technical engine behind this shift is the industry’s first Causal Encoder-Decoder (CED) architecture. Unlike conventional designs, the CED projects the decoder’s global KV cache directly from encoder hidden states, bypassing the need to derive it from every individual decoder layer. This structural innovation facilitates a highly asymmetric activation pattern: the 552B parameter [Mixture-of-Experts](/glossary/mixture-of-experts/) backbone is engineered to utilize only 8B parameters during prefill and 16B during decode. The result is a drastic reduction in memory overhead, with the KV cache compressed to just 890 bytes per token-a 75% improvement over the previous V4-Flash and a staggering 1/437th of the original DeepSeek V1.\n\nFor enterprise developers, these technical gains translate into immediate operational capacity. According to the official [DeepSeek pricing](https://api-docs.deepseek.com/quick_start/pricing), the new model is 77-80% cheaper than its predecessor, while simultaneously expanding the concurrency limit from 500 to 2,500 requests. This move serves as a direct escalation in the ongoing [pricing wars](/three-labs-cut-frontier-prices-in-72-hours-the-ai-pricing-wars-dual-track-emerges/), where the ability to minimize HBM demand and SSD persistence is rapidly becoming the primary metric for competitive survival in the agent economy.\n\nPerformance data suggests that this aggressive cost-cutting has not compromised capability. In max reasoning effort benchmarks, V4.1-Flash achieved a score of 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 88.1 on CyberGym, consistently outperforming the V4-Pro. While the model trails in pure reasoning tasks like GPQA Diamond-scoring 90.9 against Opus at 93.4 and GPT-5.6 Sol at 94.1-it manages to surpass the V4-Pro on all agentic benchmarks while utilizing 3x fewer active parameters.\n\nThese developments occur as DeepSeek, now carrying a [$71 billion valuation](/the-71b-shadow-market-what-deepseeks-valuation-signals-for-chinese-ai/), continues to push the limits of [Chinese open-weight compression](/chinese-open-weight-frontier-compresses-five-labs-thirty-days-two-licensing-models/). The company has signaled its confidence in this new architecture by announcing that, starting September 14, all requests for the retiring V4-Pro will be automatically routed to V4.1-Flash at the lower price point, as outlined in the [DeepSeek changelog](https://api-docs.deepseek.com/news/news260910).\n\nBy pairing a 1M token context window and native multimodal capabilities with an MIT license, DeepSeek is positioning V4.1-Flash as a utility-grade engine for the next generation of agents. The broader market must now contend with a new reality: extreme efficiency and top-tier reasoning are no longer mutually exclusive. As the focus of the agent economy shifts from raw parameter counts to the granular economics of [inference](/glossary/inference/) and memory management, the ability to deliver high-performance reasoning at a fraction of the cost will define the next phase of the [open-weight](/glossary/open-weights-models/) model landscape.", "url": "https://wpnews.pro/news/deepseeks-new-architecture-slashes-agentic-costs-by-80", "canonical_source": "https://forkast.news/deepseeks-new-architecture-slashes-agentic-costs-by-80/", "published_at": "2026-09-10 09:23:37+00:00", "updated_at": "2026-09-10 09:53:18.762111+00:00", "lang": "en", "topics": ["large-language-models", "ai-infrastructure", "ai-agents", "ai-products", "ai-chips"], "entities": ["DeepSeek", "DeepSeek V4.1 Flash", "DeepSeek V4-Pro", "DeepSeek V4-Flash", "DeepSeek V1", "Causal Encoder-Decoder", "Terminal-Bench 2.1", "GPQA Diamond"], "alternates": {"html": "https://wpnews.pro/news/deepseeks-new-architecture-slashes-agentic-costs-by-80", "markdown": "https://wpnews.pro/news/deepseeks-new-architecture-slashes-agentic-costs-by-80.md", "text": "https://wpnews.pro/news/deepseeks-new-architecture-slashes-agentic-costs-by-80.txt", "jsonld": "https://wpnews.pro/news/deepseeks-new-architecture-slashes-agentic-costs-by-80.jsonld"}}