The artificial intelligence industry has changed considerably since DeepSeek first attracted worldwide attention.
What began as a relatively little-known Chinese AI company became one of the most closely watched names in the industry after its reasoning model R1 demonstrated that a capable AI system could be developed and operated at a much lower cost than many of its Western competitors.
The reaction was immediate, with investors questioning whether the enormous spending on AI chips and data centers was sustainable. On January 27, 2025, Nvidia lost about $593 billion in market value in a single day as markets reacted to the possibility that more efficient AI models could reduce the demand for expensive computing infrastructure.
That episode became one of the defining moments of the current AI race.
DeepSeek did not eliminate the advantage of Silicon Valley companies, but it challenged an assumption that had become increasingly common in the industry: that progress in AI would primarily depend on building ever larger models with ever greater amounts of computing power.
DeepSeek instead drew attention to efficiency, open model releases and lower inference costs, while forcing companies and investors in the United States to reconsider how much they were willing to spend to achieve incremental improvements.
More than a year later, DeepSeek is operating in a very different environment.
The company is no longer competing simply by presenting an inexpensive alternative to American models. Chinese companies including Alibaba, Tencent, Moonshot AI, Zhipu AI and ByteDance have continued to develop increasingly capable systems, while the leading American labs have continued investing heavily in computing infrastructure and increasingly capable models.
DeepSeek itself has also expanded its operations, including plans involving additional staff, computing capacity and chip development.
And now, DeepSeek has now formally released V4 Pro, nearly four months after introducing the model as a preview.
When the model first appeared, DeepSeek introduced the V4 family in two versions, V4 Pro and V4 Flash.
The preview introduced a one-million-token context window and placed particular emphasis on coding, reasoning and agentic tasks.
V4 was also notable for its support for Huawei's Ascend hardware, reflecting China's wider effort to reduce its dependence on advanced foreign AI chips.
The April release was explicitly a preview, meaning the models were made available while DeepSeek continued developing and evaluating them. The company has now moved V4 Pro into general availability under the version name DeepSeek-V4-Pro-0813.
The model is available through DeepSeek's API, web interface and app, marking the transition from the April preview to the production version.
The four-month gap between the preview and the general release is significant because the V4 family changed considerably during that period.
DeepSeek released the official V4 Flash version at the end of July, and the company reported that Flash had surpassed the earlier V4 Pro preview on several agent benchmarks.
The result was unusual because Flash was designed as the faster and less expensive member of the family, while Pro was intended to deliver the highest level of performance.
That development also raised expectations for the final V4 Pro.
The August release is focused particularly heavily on agentic performance, which measures how effectively a model can complete multi-step tasks involving tools, software and code rather than simply answering questions. DeepSeek's reported results for the new model include an 87.9 score on Terminal-Bench 2.1, alongside strong results on other coding and agent evaluations. Some comparisons place the new model close to leading proprietary systems on particular tests, although these results should be treated as benchmark claims rather than proof that V4 Pro is universally superior.
Benchmarks also suggest that it performs quite well, with results higher than Anthropic's Claude Opus 4.7 and SpaceXAI's Grok-4.5 (both released in July), but lower that OpenAI's GPT-5.6 Sol, and the Anthropic's Claude Fable 5.
This emphasis on agents represents an important change in the AI market.
The competition is increasingly moving beyond chatbots that generate text and toward systems that can perform work. Coding agents can inspect repositories, modify files, run commands and test software. More general agents can use browsers, call external tools and work through a sequence of actions.
For DeepSeek, V4 Pro is being positioned for this environment rather than simply as another model designed to answer questions in a chat window.
DeepSeek is also introducing changes to how developers pay for V4 models. New peak and off-peak pricing is scheduled to take effect on August 16 at 16:00 UTC, with off-peak rates set at half the peak rates. The change reflects the increasing cost of running large models at scale and represents a notable shift for a company whose reputation was partly built around unusually low inference prices.
The pricing change illustrates how DeepSeek's position has evolved since the R1 shock.
In early 2025, its low-cost approach helped convince investors that AI capabilities might not require the enormous infrastructure spending associated with the leading American labs. Now DeepSeek itself is dealing with the same pressures that affect the broader industry. More capable models require substantial computing resources, and expanding access to those models brings higher operational costs.
This means the V4 generation is significant for reasons beyond model benchmarks.
The release of V4 Pro is consequently less about repeating the disruption of 2025 and more about showing whether DeepSeek can turn that disruption into a sustainable position in the AI industry. The April preview established the direction, while V4 Pro 0813 is the production version intended to compete in a market increasingly centered on coding agents, tool use and autonomous workflows.
