Deepseek Unveils v3.1 Model with Hybrid Reasoning and Competitive Pricing

DeepSeek's v3.1 Model: A Leap in AI Efficiency and Affordability

The Chinese startup DeepSeek has announced a significant update to its AI model, claiming it outperforms the widely recognized R1 across several key benchmarks. In a recent WeChat post, the company confirmed that the new v3.1 model delivers faster responses and marks their foray into AI agent development. DeepSeek emphasizes that the v3.1 model incorporates a hybrid reasoning architecture, featuring both 'thinking' and 'non-thinking' modes. This, combined with improved agent capabilities, results in enhanced tool use and task execution.

'Deep Thinking' Button: Switching Between Reasoning Modes

The DeepSeek official app and website have already been updated to v3.1, allowing users to seamlessly toggle between 'thinking' and 'non-thinking' modes via the 'Deep Thinking' button, mirroring the functionality of hybrid models like Anthropic’s Opus and Sonnet. This allows users to tailor the model's behavior to specific tasks, optimizing for speed or depth of analysis. Reportedly, the v3.1 model demonstrates superior performance on benchmarks such as SWE and Terminal-Bench, exhibiting improved thinking efficiency compared to R1. According to Artificial Analysis, the model scored 60 points on its intelligence index in reasoning mode, slightly exceeding R1's score of 59. While the underlying architecture remains consistent, with 671 billion total parameters and 37 billion active, the enhanced reasoning capabilities point to significant optimizations. Despite its higher efficiency, v3.1 also utilizes marginally fewer tokens than R1 in reasoning mode. However, the model lags slightly behind Alibaba’s latest model and OpenAI’s open-source reasoning model, GPT-OSS, in overall performance. Furthermore, it lacks function calling in reasoning mode, a crucial feature for many agentic workflows. The initial announcement of the v3.1 model occurred earlier in the week, with its initial availability limited to Hugging Face. A subsequent statement revealed that this version was specifically optimized to run on next-generation Chinese-made AI chips, indicating a commitment to supporting domestic technology.

Competitive Pricing Strategy

DeepSeek has also introduced a new pricing plan for its upgraded v3 models. While some charges have increased and evening discounts have been eliminated, the plan significantly reduces costs for certain applications, effective September 6th. This suggests a strategic move to capture market share by offering a more affordable AI solution. Specifically, DeepSeek has set pricing for its Input API at $0.07 per million tokens for cache hits and $0.56 for cache misses, with output tokens priced at $1.68 per million. These rates are substantially lower than those of competitors: Gemini 2.5 Pro charges $10 per million output tokens ($15 for longer prompts), OpenAI’s GPT-5 is also priced at $10, and Anthropic’s Claude Opus 4.1 commands a premium price of $75.

Analysts' Expectations and the R2 Delay

Analysts had anticipated the release of R1's successor earlier this year. DeepSeek's launch of the low-cost and powerful R1 model in January had previously disrupted the AI landscape. The company has since remained a key player in China's rapidly expanding AI sector, posing a challenge to US-based firms like OpenAI. Market observers are keenly awaiting the follow-up to R1, potentially named R2, which many expected to be launched earlier in the year. Reports suggest that the delay stems from the founder's dedication to perfecting the model. Concurrently, Liang Wenfeng continues to manage his successful High-Flyer Asset Management business, dividing his attention between the financial and AI sectors. As previously reported, DeepSeek postponed the release of its R2 AI model due to persistent technical difficulties with Huawei's Ascend processors. Following the success of R1, Chinese authorities encouraged DeepSeek to adopt Huawei chips over US-made Nvidia products. However, the company encountered significant challenges during the training phase of R2. Sources indicate that DeepSeek was compelled to rely on Nvidia chips for training, utilizing Huawei's Ascend processors solely for inference. Industry insiders note that Chinese chips, including Huawei’s, often trail behind Nvidia in inter-chip connectivity, software support, and overall stability. Huawei dispatched engineers to DeepSeek's offices to assist in adapting the model. Despite on-site support, the startup was unable to complete a successful training run on Ascend hardware. Originally scheduled for a May release, the launch of R2 has been delayed due to these hardware limitations. While some Chinese media outlets speculate about a potential launch in the coming weeks, DeepSeek founder Liang Wenfeng has expressed internal frustration over its progress. He has urged his team to prioritize the development of a model that maintains the company's competitive advantage, even if it requires additional time. Meanwhile, industry giants like Alibaba and Tencent continue to release frequent updates, with Alibaba's Qwen models gaining considerable traction and popularity. This competitive landscape underscores the rapid pace of innovation in the AI sector.