DeepSeek has officially released DeepSeek V4 Pro 0813, moving its flagship 1.6-trillion-parameter artificial intelligence model from preview status to general availability. The new release is accessible through the company’s app, website and API under the existing deepseek-v4-pro name, while DeepSeek has also made the model weights available under the MIT License.
DeepSeek V4 Pro Moves to General Availability
The release represents an important step for DeepSeek V4 Pro, which was previously available as an April preview. Although the basic model architecture remains largely unchanged, DeepSeek says the latest version delivers significant improvements in agent performance.
Agentic AI systems are designed to do more than simply answer questions. They can interact with tools, navigate software environments, work through complex tasks and perform multiple steps toward a desired outcome. Improvements in these areas could make the new model particularly useful for developers working on coding agents and autonomous workflows.
DeepSeek’s own testing indicates substantial gains across several agent-focused benchmarks. Terminal-Bench 2.1 reportedly increased from 72.1 in the preview version to 87.9, while CyberGym rose from 52.7 to 83.3.
DeepSWE showed one of the largest improvements, climbing from 12.8 to 62.7. Other reported results include 61.5 on NL2Repo, 74.1 on Toolathlon-Verified, 31.8 on AutomationBench Public and 67.2 on DSBench-Hard.
These results suggest that the company has focused heavily on practical software development and tool-use capabilities.
Benchmark Results Should Be Viewed Carefully
While the reported improvements are significant, the DeepSeek V4 Pro benchmark figures come from the company itself and should therefore be considered vendor-reported results.
Independent evaluations provide a different perspective. Artificial Analysis gave V4 Pro 0813 an Intelligence Index score of 53, compared with 60 for Kimi K3.
This does not necessarily diminish DeepSeek’s progress, but it highlights why model comparisons should involve multiple independent evaluations rather than relying exclusively on a single company’s testing.
Real-world performance can also vary depending on the task, prompting strategy, available tools and deployment environment.
API Pricing Takes a Major Jump
One of the biggest changes accompanying the release is the cost of using DeepSeek V4 Pro through its API.
During peak hours, the model costs $1.32 per million uncached input tokens and $3.96 per million output tokens. The previous rates were $0.435 for input and $0.87 for output.
Off-peak pricing is lower, at $0.66 per million input tokens and $1.98 per million output tokens. Peak periods currently run from 01:00–04:00 UTC and 06:00–10:00 UTC.
The increase is particularly noticeable for generated output. At peak rates, output pricing is roughly 4.5 times higher than the previous price.
That represents a significant change for developers who previously viewed DeepSeek as an especially aggressive low-cost option in the AI market. Businesses using large volumes of generated tokens will need to consider the new pricing carefully when calculating operating costs.
Million-Token Context for Large Projects
The new DeepSeek V4 Pro also supports a 1-million-token context window and can generate up to 384,000 tokens of output.
A large context window can be particularly valuable for software developers working with extensive codebases, technical documentation or long-running agent sessions. Instead of repeatedly removing older information to make room for new material, an AI system can retain substantially more project context in a single interaction.
The model supports both thinking and non-thinking modes, along with low, high and maximum reasoning-effort settings. This gives developers greater flexibility to balance response quality, reasoning depth and resource consumption depending on the task.
New Developer Tools
DeepSeek has also expanded its developer-focused ecosystem alongside the model release. DeepSeek V4 Pro now supports the OpenAI Responses API format, potentially making integration easier for developers already using compatible tooling.
The company has also introduced specific Codex integration, further emphasizing the model’s focus on software development and autonomous coding workflows.
Another major addition is DeepSeek Harness, an open-source agent framework released under the MIT License. The framework uses a plugin-based architecture that allows developers to replace components such as models, tools, storage systems and agent loops.
DeepSeek Harness remains in developer preview, but its open architecture could give developers considerable flexibility when building customized AI agents.
A More Powerful but More Expensive Model
The release of DeepSeek V4 Pro 0813 demonstrates how quickly the AI market is evolving. DeepSeek has significantly improved its reported agent benchmarks while expanding the model’s context capabilities and developer integrations.
However, the higher API prices represent an important trade-off. Developers now have access to a more capable model for demanding workloads, but organizations must determine whether its performance improvements justify the increased cost.
With open model weights, a massive context window, improved agent capabilities and a growing developer ecosystem, DeepSeek V4 Pro 0813 is positioned as a serious option for advanced AI and software development applications.
