Introduction
DeepSeek-V4-Pro-0813 has taken second place in the latest SuperCLUE-Terminal Chinese agentic terminal programming benchmark, scoring 51.52 points. It trails only Kimi K3, which leads the ranking with 60.61 points.
What makes the result especially notable is cost. In the benchmark run, DeepSeek-V4-Pro-0813 averaged about RMB 1.42 per task, giving it a strong cost-performance position among the highest-scoring models.
SuperCLUE-Terminal is designed around realistic programming tasks for Chinese developers. The models are evaluated through a common agent framework based on Claude Code, allowing the benchmark to compare how well they understand Chinese requirements, plan multi-step work, use tools, debug problems, and modify code through an end-to-end workflow.
DeepSeek V4 Pro Ranks Second in the Latest Coding Evaluation
The benchmark is intended to reflect real development work rather than short, isolated coding questions. According to the published results, individual tasks can require roughly 50 to 110 interaction rounds, while a complete run for one task may take around 30 to 90 minutes.
In the current ranking, Kimi K3 remains in first place with 60.61 points. DeepSeek-V4-Pro-0813 follows at 51.52, ahead of GLM-5.2 at 48.48 and the earlier DeepSeek-V4-Flash-0731 at 46.46.
| Model | SuperCLUE-Terminal Score | Reported Average Cost per Task |
|---|---|---|
| Kimi K3 (max) | 60.61 | RMB 19.63 |
| DeepSeek-V4-Pro-0813 (max) | 51.52 | RMB 1.42 |
| GLM-5.2 (max) | 48.48 | RMB 23.30 |
| DeepSeek-V4-Flash-0731 (max) | 46.46 | RMB 0.64 |
These figures come from the SuperCLUE-Terminal benchmark environment and should be read as benchmark execution costs, not as fixed per-task prices charged by the model providers.
DeepSeek has separately confirmed that the production API model deepseek-v4-pro now points to DeepSeek-V4-Pro-0813. The model is available through DeepSeek’s web app, mobile app, and API.
RMB 1.42 per Task: The Standout Cost Advantage
The most striking part of the result is DeepSeek-V4-Pro-0813’s reported cost.
At around RMB 1.42 per benchmark task, its average cost was roughly one-fourteenth of Kimi K3’s RMB 19.63 and around one-sixteenth of GLM-5.2’s RMB 23.30 in the same SuperCLUE comparison.
That difference matters because the score gap between high-performing coding models can be relatively small while the cost of running long agentic tasks can vary dramatically. For smaller companies, independent developers, and teams that need to run coding agents repeatedly, execution cost can become just as important as the headline benchmark score.
The evaluation covered practical development scenarios including front-end pages, 3D games, and engineering-script modifications. In the reported tests, DeepSeek-V4-Pro-0813 was able to complete game-development logic involving collision handling, scoring, and end-state behavior, while also producing interface styling and interaction feedback that went beyond a basic template-like result.
Some creative drawing tasks still showed minor structural issues, but the overall completion quality remained competitive with the leading group.
For budget-sensitive development teams, the result gives DeepSeek-V4-Pro-0813 a clear positioning: near-frontier coding-agent performance with a much lower benchmark execution cost than several higher-priced competitors.
常见问题
What is DeepSeek-V4-Pro-0813?
DeepSeek-V4-Pro-0813 is the current production version behind the deepseek-v4-pro API model name. DeepSeek officially released the GA version on August 13, 2026, with improved agent capabilities and support for its web app, mobile app, and API.
Build a showcase site and grow leads in minutes
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
What is SuperCLUE-Terminal?
SuperCLUE-Terminal is an agentic terminal-task benchmark focused on realistic Chinese programming and engineering workflows. It evaluates end-to-end performance such as requirement understanding, planning, command execution, file modification, debugging, and final task completion.
Why does the benchmark use Claude Code?
The benchmark uses a common agent framework so that participating models are tested under a more comparable execution setup. Claude Code provides the terminal-oriented agent environment used to run the tasks, while the underlying model is changed for each evaluation.
Is RMB 1.42 the official price of one DeepSeek V4 Pro task?
No. RMB 1.42 is the average cost observed per task in this SuperCLUE-Terminal benchmark run, based on the benchmark’s token usage and pricing methodology. Actual API costs depend on input tokens, output tokens, cache usage, reasoning settings, and the provider’s current pricing.
How does DeepSeek V4 Pro compare with Kimi K3 in this benchmark?
Kimi K3 scored higher, with 60.61 points versus DeepSeek-V4-Pro-0813’s 51.52. However, the benchmark reported an average task cost of RMB 19.63 for Kimi K3 and RMB 1.42 for DeepSeek-V4-Pro-0813, giving DeepSeek a substantial cost advantage in this particular evaluation.
Does DeepSeek V4 Pro support long-context and agent workflows?
Yes. DeepSeek’s official API documentation lists a 1M-token context length for DeepSeek V4 Pro and supports tool calls, the Responses API, the Anthropic-compatible API, and multiple reasoning-effort settings. DeepSeek specifically positions the GA release as an upgrade for agent workloads.
Is DeepSeek V4 Pro always better than GLM-5.2 for coding?
Not necessarily. In this SuperCLUE-Terminal result, DeepSeek-V4-Pro-0813 scored 51.52 while GLM-5.2 scored 48.48, but one benchmark cannot determine performance across every coding workload. Repository size, language, tool environment, long-context behavior, latency, and cost can all affect model choice.
相关工具
- DeepSeek API: The official API platform for accessing DeepSeek models including
deepseek-v4-pro. - SuperCLUE-Terminal: The benchmark used for the agentic terminal-programming comparison discussed in this article.
- Claude Code: Anthropic’s agentic coding tool used as the common execution framework in the benchmark.
- Kimi API: Moonshot AI’s official developer platform for Kimi K3 and other Kimi models.
- Z.AI GLM-5.2: The official overview of GLM-5.2, another high-scoring model in the benchmark.
Related Links
- DeepSeek-V4-Pro GA Release: DeepSeek’s official announcement for the production V4 Pro release.
- DeepSeek Models and Pricing: Official model versions, context limits, features, and current API pricing.
- DeepSeek API Quick Start: Official instructions for accessing
deepseek-v4-prothrough the API. - SuperCLUE-Terminal Benchmark: The official benchmark page containing the ranking and cost comparison.
- Kimi K3 API Documentation: Official Kimi documentation covering the K3 model and API usage.
- Claude Code Documentation: Official documentation for the coding-agent framework used by the benchmark.
- GLM-5.2 Official Overview: Z.AI’s official introduction to GLM-5.2 and its long-horizon coding focus.
Summary
DeepSeek-V4-Pro-0813 scored 51.52 in SuperCLUE-Terminal, placing second behind Kimi K3 and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731 in this benchmark.
Its main advantage was cost: SuperCLUE reported an average of roughly RMB 1.42 per task, far below the cost recorded for Kimi K3 and GLM-5.2 in the same evaluation.
The result does not mean DeepSeek V4 Pro is universally the best coding model, but it does show a strong balance between agentic coding performance and execution cost for long, tool-heavy programming tasks.
For teams that care about both coding-agent capability and operating cost, DeepSeek-V4-Pro-0813 is one of the most competitive options in this SuperCLUE-Terminal comparison.
