DeepSeek V4 Pro tops the value-for-money programming benchmark, ranking second overall at just one-fourteenth the cost of the first-place model.

The CLUE team released the SuperCLUE-Terminal Chinese intelligent agent terminal programming evaluation leaderboard, with the new version DeepSeek-V4-Pro-0813 scoring 51.52 points.

This result ranks second among all tested models, trailing only Kimi-K3, while maintaining extremely low invocation costs, giving it a standout cost-performance advantage.

SuperCLUE-Terminal is a practical evaluation standard designed specifically for domestic developers, using entirely local real-world programming tasks and uniformly running on Claude Code as the framework to fairly compare each model's complete capabilities in understanding Chinese requirements, planning tasks, calling tools, debugging, and fixing code.

Each test question requires the model to complete 50 to 110 rounds of interaction, with a single question taking half an hour to an hour and a half to fully run, accurately reproducing the complex workflows of everyday development.

In this leaderboard, Kimi-K3 ranks first with 60.61 points, followed closely by DeepSeek-V4-Pro-0813, whose score surpasses GLM-5.2's 48.48 points and the older DeepSeek-V4-Flash's 46.46 points.

Its most impressive aspect is cost control, with a single-question invocation cost of only 1.42 yuan, roughly one-fourteenth of Kimi-K3's and one-sixteenth of GLM-5.2's.

The evaluation scenarios cover real-world needs such as front-end pages, 3D games, and engineering script modifications. In actual testing, it fully completes the game development logic loop, with collision, scoring, and settlement features all functioning properly, and page color schemes and interaction feedback free from templated AI styling; only a few creative drawing tasks show minor structural flaws, but overall completion remains in the top tier.

For small and medium-sized enterprises and independent developers, this model strikes a balance between performance and cost. Top-scoring models carry high computational costs that put long-term pressure on small teams, while DeepSeek-V4-Pro-0813 approaches top-tier coding capability at a lower price.