AI startup DeepSeek founder Liang Wenfeng. Photo: VCG

Chinese AI company DeepSeek has unveiled the official version of DeepSeek-V4-Pro late Wednesday, featuring significantly enhanced agent capabilities and support for the Responses Application Programming Interface (API) and Codex integration, the company said in a post on its website.
 
“It is now fully available across the web, mobile app, and API; we welcome your testing and feedback,” according to DeepSeek.
 
DeepSeek updated its API documentation, replacing the preview edition of its V4-Pro model with a new version, DeepSeek-V4-Pro-0813, marking the formal rollout of the API version of DeepSeek-V4-Pro.
 
According to the company, the updated model offers a 1 million-token context window, a maximum output length of 384K tokens, and supports both thinking and non-thinking modes.
 
The launch of new model comes as artificial intelligence (AI) developers increasingly shift from simple chat use cases toward longer, multi-step workflows that rely on tool use, code execution, and persistent task completion. Those workloads consume far more tokens than traditional conversational apps, making pricing and inference efficiency increasingly important.
 
DeepSeek also released benchmark charts showing strong performance across several agent-focused tests, including terminal operation, code engineering, tool use and security-related tasks. In some cases, the company said, the V4-Pro-0813 approached or even surpassed leading overseas models.
 
Based on information released by the company on Thursday, the DeepSeek-V4-Pro-0813 delivers performance approaching that of Anthropic’s Claude Fable 5 on benchmark tests. Benchmark results showed V4 Pro-0813 scored 87.9 on Terminal-Bench, a test designed to measure how well AI agents perform tasks in terminal environments, narrowly trailing Fable 5's score of 88.
 
The pricing has also drawn attention. DeepSeek has priced DeepSeek-V4-Pro-0813 at 3 yuan ($0.44) per million input tokens and 6 yuan per million output tokens, with cached input priced at 0.025 yuan per million tokens.
 
The update quickly drew attention from industry analysts. Research firm SemiAnalysis posted on X congratulating DeepSeek on the release of V4-Pro-0813 and said the model “massively beats Nemotron 3 Ultra on agentic tasks.”
 
In a near-simultaneous move, Elon Musk’s AI company xAI on Wednesday released Grok 4.6, a new frontier model designed for long-running agents and visual work. Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens.  
 
According to xAI, it achieves frontier intelligence across several agentic coding and knowledge work benchmarks, and matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index.
 
The near-simultaneous release of DeepSeek-V4-Pro-0813 and Grok 4.6 shows that the AI industry is entering a new phase of competition. Frontier models are no longer defined only by parameter scale or conversational ability, but increasingly by the ability of AI systems to independently use tools, write and execute code and complete multi-step tasks, according to Tian Feng, former dean of SenseTime's Intelligence Industry Research Institute.
 
According to media reports, DeepSeek has previously indicated that it may raise API prices overall, saying on August 6 that it planned a “relatively significant” increase, though it has not disclosed the exact timing or scope of any changes. The statement has drawn attention to the company’s broader commercial strategy. Analysts said DeepSeek may eventually adopt a more granular pricing structure to balance rising inference costs with commercial returns.
 
Analysts noted that DeepSeek’s price remains highly competitive among frontier models, particularly for agentic and coding workloads that often require large token consumption. Compared with overseas models, even if the company raises its prices, it still has substantial room for price adjustment while maintaining competitive capabilities.  
 
One comparison cited by market observers puts DeepSeek-V4-Pro at roughly $0.88 per million output tokens, compared with $50 per million output tokens for Fable 5, underscoring the company’s price-performance positioning.
 
It also comes as demand for coding and agent applications continues to rise. Data from China’s National Data Administration in March showed that daily token usage in the country had surpassed 140 trillion, more than 1,000 times the level seen at the beginning of 2024, the Shanghai Securities News reported.
 
As application demand expands, model providers face growing pressure to maintain technical advances while keeping prices attractive, Tian told the Global Times on Thursday, adding that for frontier AI companies, the next competitive edge may depend less on who answers better in chat, and more on who can complete complex real-world tasks at lower cost and with higher reliability.