DeepSeek has moved its 1.6-trillion-parameter flagship model, V4 Pro 0813, from preview to general availability. The release is available through Expert Mode in the DeepSeek app and through the API. It follows a testing period that began in April.

The company also released an open-source agent framework called Harness and introduced peak and off-peak pricing for its V4 model family. Together, the changes move DeepSeek away from its earlier ultra-low API pricing and toward a broader model-and-agent offering.

DeepSeek V4 Pro Agent Benchmark Results

DeepSeek attributes V4 Pro’s improved agent performance to post-training changes rather than a new architecture. Its reported benchmark results show gains across several agent-focused tests:

 

Benchmark

 

 

Previous Score

 

 

V4 Pro Score

 

 

Terminal-Bench 2.1

 

 

72.1

 

 

87.9

 

 

CyberGym

 

 

52.7

 

 

83.3

 

 

DeepSWE

 

 

12.8

 

 

62.7

 

On Terminal-Bench 2.1, DeepSeek’s reported V4 Pro result sits 0.1 points behind Anthropic’s Fable 5. These vendor-reported scores have not been independently verified.

DeepSeek Harness Open-Source Agent Framework

Alongside V4 Pro, DeepSeek released Harness, an MIT-licensed open-source framework for AI agents. It uses a plugin-first architecture and is designed as an inspectable, self-hostable option for developers building coding agents.

Harness entered the market alongside proprietary coding-agent products such as Anthropic’s Claude Code and OpenAI’s Codex. DeepSeek’s approach gives developers access to the framework’s underlying implementation rather than positioning the tool solely as a closed service.

The project reached 20,000 GitHub stars in roughly one hour on August 13. Chinese and Western technology press described that pace as the fastest ever recorded on GitHub. By August 17, Harness had accumulated more than 141,000 stars.

DeepSeek V4 Pro API Pricing Changes

DeepSeek introduced peak and off-peak billing for the V4 model family starting August 16 at 16:00 UTC. For V4 Pro output tokens, the former flat rate of $0.87 per million tokens changed to:

  • $3.96 per million output tokens during peak hours
  • $1.98 per million output tokens during off-peak hours

The peak price represents an increase of roughly 4.5 times the previous V4 Pro output-token rate. Some token categories increased by more than 1,100% from the earlier flat pricing structure.

This marks a clear break from the period when DeepSeek was closely associated with unusually low-cost API access.

Peak Hours and DeepSeek User Geography

DeepSeek’s peak pricing windows are set from 9:00 to 12:00 and 14:00 to 18:00 Beijing time. The break falls during the typical Chinese lunch period.

When converted to UTC, the peak windows sit outside U.S. business hours. The timing suggests that a large share of DeepSeek’s paying users may remain in Asia.

Similarweb data from March 2026 showed that China accounted for 44% of DeepSeek desktop visitors, while the United States accounted for 5.68%.

DeepSeek’s Position Against Western AI Rivals

Even after the pricing changes, DeepSeek V4 Pro remains less expensive per million output tokens than Anthropic’s Fable 5, which charges $50 per million output tokens. Still, the cost difference is much smaller than it was under DeepSeek’s earlier pricing.

DeepSeek raised approximately $7.4 billion in its first outside funding round in June. The company is also building a gigawatt-scale data center in Inner Mongolia.

V4 Pro’s reported agent benchmark gains, the release of Harness, and the updated pricing structure point to a broader product strategy. DeepSeek is pairing its flagship model with an open-source agent framework while moving away from competing primarily on the lowest API prices.