# DeepSeek’s Cheap AI Is Spooking the Industry > Source: > Published: 2026-10-28 07:07:09+00:00 ## The price tag is the first shock DeepSeek’s V4.1-Flash-1017, an October 17 update to its rapid, cost-optimized V4.1-Flash family, arrived just five weeks after its predecessor. This iteration boasts a **one-million-token context window** and supports both text and image inputs, positioning it as a versatile, high-throughput model. The model’s API rates are exceptionally low: $0.30 per million input tokens and $1.20 per million output tokens. This stands in stark contrast to flagship models like [Claude Opus](https://www.stork.ai/en/anthropic-workbench), which can cost 15 to 20 times more. For businesses deploying coding agents or managing high-volume AI workflows, these savings compound dramatically, turning operational costs from prohibitive to practical. DeepSeek also publishes **open weights** for V4.1-Flash-1017 under the permissive MIT license on Hugging Face. This allows developers to download, modify, and integrate the model into commercial products without incurring licensing fees. While local deployment is permitted, the model’s substantial size (reported between 550 billion and 700 billion parameters) demands significant hardware, creating a barrier for many users despite the open-source nature. ## Strong agent scores, with asterisks DeepSeek's reported benchmarks paint an impressive picture for its V4.1-Flash-1017 model. The company claims strong performance on agentic tasks, including 77% on **DeepSWE v1.1** (a software engineering problem-solving test) and 60.5% on **Automation-Bench** (evaluating end-to-end business workflows). These figures, alongside strong results in other coding and agent evaluations, position it competitively against established models. However, DeepSeek’s published comparison table, while highlighting its strengths, also reveals areas where it trails competitors. For instance, the model scores lower than Claude Opus on tests requiring complex reasoning, such as Humanity’s Last Exam and Terminal-Bench, which assesses in-terminal coding. It also lags in repository-building tests, suggesting that excelling at tool-driven, discrete tasks doesn't automatically translate to leadership in general reasoning or holistic project creation. Crucially, these are company-reported benchmarks, not independent verdicts. DeepSeek’s charts compare V4.1-Flash-1017 against selected rival models and previous generations, such as [Claude Opus 5](https://www.stork.ai/en/anthropic-workbench) and GPT-5.6-Sol. This selective comparison means the results don't necessarily reflect performance against the absolute latest, most capable models available today. While impressive, these scores do not claim DeepSeek beats today’s best models across the board. ## The bargain comes with real trade-offs Bargains often carry hidden costs, and [DeepSeek V4.1-Flash-1017](https://www.stork.ai/en/deepseek) is no exception, despite its impressive agent scores. Independent evaluations paint a more nuanced picture, revealing trade-offs in raw intelligence and occasional factual accuracy. Artificial Analysis, a respected independent benchmark, places DeepSeek V4.1-Flash-1017 on its Intelligence Index at a score of 45. While this is a solid showing across diverse tasks—including office work, agentic coding, and scientific programming—it trails leading frontier systems like [Claude Opus 5.5](https://www.stork.ai/en/anthropic-workbench) (58) and GPT-6-Astra (53). However, its cost per task remains exceptionally attractive, often less than a quarter of a dollar. A significant reported weakness lies in its tendency toward **hallucination**. The Omniscience evaluation found that DeepSeek V4.1-Flash-1017 frequently provides unsupported answers, even when performing well in structured agent loops. This necessitates rigorous verification for factual research or high-stakes decision-making, regardless of the model’s apparent efficiency. More details on DeepSeek's models can be found on the [DeepSeek Official Website](https://www.deepseek.com). User demonstrations highlight the model's cost efficiency but also task dependency. One coding comparison showed DeepSeek producing similar output to a more expensive model at a fraction of the cost. Conversely, a simple animation test saw a different low-cost model outperform DeepSeek. This variability underscores that optimal model choice depends heavily on task complexity and evaluation methodology. Enjoying this? Get one like it in your inbox each morning. one email a day · unsubscribe in two clicks · no third-party tracking ## Why Washington—and the market—are watching Strategic pressure mounts as DeepSeek’s capable, inexpensive open model challenges the premium, closed-API economics favored by US AI companies. Developers gain greater control over deployment and data, shifting the competitive landscape. Video reports claim the US government ordered DeepSeek removed from federal civilian computers, and that a hacker used it to breach over 100 companies. These claims, if independently verified, suggest a growing policy concern regarding vendor origin, distinct from direct evidence of model-induced intrusion. DeepSeek may be a compelling option for **cost-sensitive coding and agent workloads**. Buyers should test its performance on specific tasks, rigorously assess privacy and security requirements, and compare total hosting costs—not just token prices. This model exemplifies the market's accelerating shift toward accessible, high-performance AI. ## Frequently Asked Questions ### What is DeepSeek V4.1-Flash-1017? It is an updated, open-weight DeepSeek model designed for fast, lower-cost text, image, coding, and agent tasks. ### How much does DeepSeek V4.1-Flash cost to use through its API? The cited pricing is $0.30 per million input tokens and $1.20 per million output tokens, with additional off-peak discounts. ### Can anyone run DeepSeek locally? The MIT-licensed weights can be downloaded and modified, but the model is large and typically requires substantial multi-GPU hardware to run. ### Is DeepSeek better than Claude Opus? It performs strongly on some coding and agent benchmarks, but independent results and other tests show it does not consistently match top proprietary models.