DeepSeek's AI Model Offers Competitive Edge in Coding and Cheaper Web Search

DeepSeek has released a minor update to its AI model, V4.1-Flash-1017, along with public access to its Search and Fetch tools for AI agents. The model includes open weights under an MIT license, which means users can freely download and use it to run or build a paid service. On the DeepSWE v1.1 coding-agent test, the model scored 77.0, outperforming Claude Opus 5 and GPT-5.6 Sol. However, it's crucial to note that the tests were self-administered by DeepSeek, and Opus 5 still leads on other benchmarks.

Cost-Effective Tools and Challenges

DeepSeek's web Search and Fetch tools offer significant cost savings, charging $0.75 per 1,000 successful requests compared to competitors charging $5 and up. This pricing is expected to impact bulk activities like routine coding and agent loops, as open weights allow any cloud to host the model. Despite this, leading models are still preferred for complex reasoning tasks, and no other companies have reduced their prices yet.

While DeepSeek's services offer a budget-friendly alternative, users should remain aware of data security concerns, particularly as several governments restrict the use of DeepSeek's tools on official devices. It's sensible to use these tools for public-web research only, keeping sensitive information secure. Additionally, the free usage period ends on November 1. At that point, verifying DeepSeek's coding test scores independently could provide more clarity on their efficacy.

Source: The Neuron