DeepSeek Releases Minor Update to V4.1-Flash, Emphasizing Performance Gains and Cost Efficiency
Beijing, China – October 17, 2026 – Chinese AI firm DeepSeek launched a minor update, V4.1-Flash-1017, for its AI model on Saturday, touting improved benchmark scores, faster processing, and shorter responses, all while maintaining its existing API pricing. The release comes just a day after Beijing announced plans to potentially regulate AI technology exports, with DeepSeek proceeding with its open-source model as usual.
The V4.1-Flash-1017 update, announced on X (formerly Twitter), is described as a refinement rather than a new generation of the model. DeepSeek stated that the underlying architecture and scale remain unchanged, but performance has been enhanced across various tasks, including reasoning, coding, and agent-based simulations. The updated model is now available on its website, app, and API, with its weights also made available for open-source use.
The company’s official model card indicates significant advancements in inference, programming, agent tasks, and visual tests. DeepSeek highlighted the use of self-feedback reinforcement learning in this version, which it claims enhances the model’s ability to generate complete projects from scratch, citing examples ranging from game development to 3D modeling. The model is licensed under MIT, with weights accessible on Hugging Face and ModelScope.
DeepSeek also released a cost-performance comparison chart based on the DeepSWE benchmark, which assesses the model’s ability to fix real-world coding issues. The chart plots average cost per problem against success rate. Preliminary readings from the chart show the new version achieving a top success rate of 77.0% at approximately $0.20 per problem, with a lower tier at around 72% success for about $0.10 per problem. In contrast, the previous V4.1-Flash version had a top success rate of approximately 74% at about $0.50 per problem. This indicates a substantial improvement in both performance and a reduction in cost, with the per-problem expense dropping to about 40% of its previous level.
The reduction in response length is directly linked to the cost savings. While the per-token API pricing remains unchanged, the model’s more concise outputs and faster task completion lead to lower overall bills for users.
The official benchmark table includes 11 tests, all of which reportedly saw improvements compared to the previous version. The most significant gains were observed in Terminal-Bench 4.0 (from 31.2 to 42.3), ProgramBench (from 20.3 to 29.1), and HLE (Human Last Exam, a test of complex math, science, and humanities problems, from 36.8 to 42.6).
When compared to Claude Opus 5, the new DeepSeek version reportedly outperformed it in four categories: DeepSWE (77.0% vs. 74.0%), Automation-Bench (60.5% vs. 50.3%), Agents’ Last Exam (35.4% vs. 28.6%), and Apex Bench (72.1% vs. 65.8%). These wins are concentrated in areas related to software engineering and agent automation. However, DeepSeek’s model lagged behind Claude Opus 5 in seven categories, with the largest discrepancies in HLE (42.6% vs. 56.3%), Terminal-Bench 4.0 (42.3% vs. 51.8%), and ProgramBench (29.1% vs. 37.0%). Performance in three visual tests showed only a minor difference of 1 to 2 points.
Against other Chinese competitors, DeepSeek’s V4.1-Flash-1017 reportedly leads Kimi K3 and GLM-5.3 across most benchmarks, with a slight edge lost only in HLE against Kimi K3 (42.6% vs. 43.5%).
However, two points warrant attention. Firstly, Anthropic released Claude Opus 5.5 on September 22nd, claiming it matches Fable 5.1 performance on most tasks with a 40% reduction in execution costs compared to Opus 5. DeepSeek’s benchmark data, however, compares its model to Opus 5 and GPT-5.6 Sol, and the cost chart does not include Opus 5.5, meaning the latest flagship competitor was not part of the comparison. Secondly, all performance figures are self-reported by DeepSeek, and independent third-party evaluations are not yet available.
The announcement of potential AI export regulations by China’s Ministry of Commerce and Ministry of Science and Technology on October 16th proposed adding 23 AI-related technologies to the restricted export list for 2026. These include large AI models, agents, computing clusters, and information processing. The notice explicitly stated that technologies already in the public domain through open-source releases or publications would not be subject to these controls. Public feedback on the proposal is being accepted until November 15th. Bloomberg reported that the draft targets top-tier models that meet certain thresholds in performance, usage, and safety, potentially including those ranked in the global top 10, with controls covering model weights, checkpoints, source code, and architecture.
DeepSeek’s decision to proceed with its open-source release the following day aligns with statements from a Ministry of Commerce spokesperson, who indicated that open-source AI technologies would benefit from “automatic licensing” and expedited approval processes. The spokesperson also clarified that providing API, app, or web services using restricted large models does not constitute technology export; only deploying restricted technologies locally overseas would fall under the export control. Based on this interpretation, Taiwanese and overseas developers using DeepSeek’s services via API, web, or app would likely not be affected by the proposed regulations. However, the draft is still open for public comment, and the final version may see adjustments.
Leave a Reply