On the evening of October 17, 2026, Chinese company DeepSeek released an updated version of its neural network—DeepSeek-V4.1-Flash-1017. This is formally a "minor" update to the V4.1-Flash model released earlier that month, but given the scale of changes and claimed results, it is quite notable: developers have enhanced the model's capabilities in programming, logical reasoning, and executing multi-step agentic tasks, while also reducing response lengths and query execution times. The model weights have been released under an open MIT license, and in DeepSeek's pricing sheet, the name deepseek-flash now points specifically to version 1017—previous names have been deprecated, meaning the update has been rolled out automatically across existing tiers.
What Changed in Version 1017
According to data published by the developers, the key focus of the rework was optimization for AI agents: the model has been trained to independently run code, identify errors and fix them, as well as respond more concisely. To achieve this, DeepSeek reinforced reinforcement learning—the neural network is rewarded for precise solutions without unnecessary explanations. Fewer tokens in the response mean potentially faster and cheaper queries. The company's head of post-training, Wu Yu, called this the beginning of an era of programming accessible to everyone, emphasizing the focus on development speed and working with code.
Benchmark Results
According to results published by DeepSeek, the compact model outperformed Claude Opus 5 in several tests. In DeepSWE 1.1, the new model scored 77% compared to 74% for Opus 5.0 (with GPT-5.6 Sol at 73%). In AutomationBench, the result was 60.5% against 50.3% for Opus 5 and 45.8% for GPT-5.6 Sol, while in Apex Bench it reached 72.1% compared to 65.8%. The initial publication also claimed superiority over a number of other models, including OpenAI solutions. Importantly, all these figures are provided in the company's own materials: details on the comparison methodology and independent verification of the results were not published by the developers.
Architecture and Training
The architectural foundation of the previous version has been preserved: the model uses Causal Encoder-Decoder, Compressed Sparse Attention 2, FP4 quantization, and SWA Bounded Replay. Changes primarily affect the post-training stage and objective functions rather than the fundamental network structure. This approach maintains inference speed and low output cost while simultaneously improving quality on agentic and programming tasks.
Pricing and Availability
The model is available across multiple channels: on the official DeepSeek website, in the mobile app, and via API, and its weights have been published on Hugging Face under an open MIT license—any developer can download, run, and integrate it into a commercial product without permission or royalties. According to the pricing sheet, the model contains 763.2 billion parameters (about 510 GB of files), and the tariff is 15 cents per million input tokens during off-peak hours and 60 cents per million output tokens; during peak hours, the price is twice as high. According to the same pricing tiers, this is approximately 4.1 times cheaper than the vendor's flagship Pro model. In the first day and a half after publication, the model card received over 1,100 likes, and the weights were downloaded more than 10,000 times.
Caveats and Independent Verification
The main caveat to the claim of superiority over Claude Opus 5 and GPT-5.6 Sol is that the results were obtained in DeepSeek's internal tests and have not yet been confirmed by independent laboratories. The comparison methodology, prompt composition, and run conditions are not disclosed in the open materials, so these figures should be treated as developer-claimed rather than verified facts. If the results are confirmed in practice, this means developers can automate programming and agent tasks using a noticeably cheaper open model, which could shift the balance of power in the affordable LLM segment.
