Deepseek unveils v3.1 model with hybrid reasoning and lower prices

The Chinese startup DeepSeek introduced a new update, claiming it outperforms the widely recognized R1 across core benchmarks. In a Thursday WeChat post, the AI company confirmed that the new model version, V3.1, provides quicker responses to queries and signals their entry into AI agent development.

DeepSeek added that the model supports a hybrid reasoning architecture, having both thinking and non-thinking modes, improved agent capabilities, and stronger performance in tool use and task execution.

DeepSeek provides a “Deep Thinking” button to switch between modes

So far, DeepSeek’s official app and website have already been updated to V3.1, allowing users to toggle between thinking and non-thinking modes via the “Deep Thinking” button, similar to how Anthropic’s hybrid models like Opus and Sonnet work.

Reportedly, the V3.1 model also performs better on benchmarks like SWE and Terminal-Bench and thinking efficiency than R1. Moreover, according to Artificial Analysis, the model reached 60 points on its intelligence index in reasoning mode, just above the 59 scored by R1. Still, the underlying architecture remains the same, with 671 billion total parameters and 37 billion active.

Despite having a higher efficiency, it also uses slightly fewer tokens than R1 in reasoning mode. The new model, however, is slightly behind Alibaba’s latest model and OpenAI’s open-source reasoning model, GPT-OSS, in performance. It also lacks function calling in reasoning mode, which is considered a major constraint in agentic workflows.

The startup had first announced the new model on Tuesday, though it was only available on Hugging Face at the time. A separate statement added that the version had been tailored to run on next-generation Chinese-made AI chips. 

Now, the company unveiled a new pricing plan for its upgraded V3. The plan raises some charges, eliminates evening discounts, and reduces costs in certain applications, effective Sept. 6.

DeepSeek set pricing for its Input API at $0.07 per million tokens for cache hits and $0.56 for cache misses, with output tokens at $1.68 per million. The rates sharply undercut competitors: Gemini 2.5 Pro costs $10 per million output tokens ($15 for longer prompts), OpenAI’s GPT-5 is also $10, and Anthropic’s Claude Opus 4.1 goes as high as $75.

Analysts expected DeepSeek to release R1’s successor earlier this year

DeepSeek first rattled Silicon Valley with its low-cost and powerful R1 AI model launch in January. The model has since stayed at the forefront of China’s accelerating AI push, challenging US firms such as OpenAI.

Market observers, however, are still waiting for the follow-up to R1, a possible R2 model, which many had expected to launch earlier this year. Local reports have hinted that the delay in the launch stems from founder Liang Wenfeng’s insistence on perfecting the model. At the same time, he also manages his profitable High-Flyer Asset Management business. 

As previously reported by Cryptopolitan, DeepSeek has delayed the launch of its R2 AI model after encountering persistent technical issues with Huawei’s Ascend processors. Following the success of its R1 model in January, DeepSeek was encouraged by Chinese authorities to adopt Huawei chips instead of US-made Nvidia products. However, the company ran into significant problems during the training phase of its R2 model.

Sources familiar with the matter said DeepSeek had to rely on Nvidia chips for training while using Huawei’s Ascend processors only for inference. Industry insiders note that Chinese chips, including Huawei’s, often lag behind Nvidia in inter-chip connectivity, software support, and overall stability.

Huawei sent engineers to DeepSeek’s offices to help adapt the model. Still, the start-up could not complete a successful training run on Ascend hardware even with on-site assistance. Originally slated for a May release, the R2 model’s launch has been postponed due to these hardware challenges.

While some Chinese media outlets speculate that the new model could launch in the coming weeks, DeepSeek founder Liang Wenfeng has voiced internal frustration over its progress, urging the team to take the necessary time to develop a model that preserves the company’s competitive edge.

Meanwhile, industry heavyweights including Alibaba and Tencent continue to release updates briskly, with Alibaba’s Qwen models attracting a particularly strong following.

Sign up to Bybit and start trading with $30,050 in welcome gifts