I Asked ChatGPT, Claude and DeepSeek to Build Tetris

2025-12-22 · Source: KDnuggets · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, medium

Summary

An editorial analyst conducted a practical test comparing the code generation capabilities of three flagship large language models: Claude Opus 4.5, GPT-5.2 Pro, and DeepSeek V3.2. The objective was to evaluate their ability to build a fully functional Tetris game from a single, detailed prompt, focusing on first-attempt success, feature completeness, playability, and cost-effectiveness. Claude Opus 4.5 delivered a complete, smooth, and playable game in under 2 minutes. GPT-5.2 Pro, despite being OpenAI's most intelligent model and 4x more expensive than Opus 4.5, produced a game with a layout bug on its first attempt, requiring a follow-up prompt to fix, and still resulted in a less smooth user experience. DeepSeek V3.2, the most affordable option, had multiple bugs, including disappearing pieces and scrolling issues, making the game unplayable even after a second iteration. The cost analysis showed Opus 4.5 at ~$0.09 for a playable game, GPT-5.2 Pro at ~$0.41 for a playable but poor UX game, and DeepSeek V3.2 at ~$0.005 for an unplayable game.

Key takeaway

For AI engineers and software developers evaluating LLMs for code generation, prioritize Claude Opus 4.5 for day-to-day coding tasks due to its high first-attempt success and superior output quality, which ultimately saves time and cost. If your project has a tight budget and you have debugging capacity, DeepSeek V3.2 offers a cost-effective alternative, even with multiple iterations. Avoid GPT-5.2 Pro for simple coding, as its strengths lie in complex reasoning, making it over-engineered and less efficient for straightforward development.

Key insights

Model performance for coding tasks varies significantly across LLMs, impacting development cost and user experience.

Principles

• Higher cost does not guarantee superior code generation.
• First-attempt success reduces overall development cost.
• Model suitability depends on task complexity and budget.

Method

Evaluate LLMs for code generation by prompting them to build a complex, interactive application (e.g., Tetris) and assessing first-attempt success, feature completeness, playability, and total cost across iterations.

In practice

• Use Claude Opus 4.5 for reliable daily coding tasks.
• Consider DeepSeek V3.2 for budget-constrained projects.
• Reserve GPT-5.2 Pro for complex reasoning tasks.

Topics

• AI Model Comparison
• Code Generation
• Large Language Model Performance
• Cost-effectiveness
• Game Development

Best for: AI Engineer, Machine Learning Engineer, Software Engineer

Related on AIssential

• Claude Opus 4.6 vs GPT-5.3 Codex: How I shipped 93,000 lines of code in 5 days — Pairing generative AI models with specialized review models optimizes complex software development workflows.
• I found that different models (when used for coding) have different "work morale" — LLMs exhibit distinct, inherent behavioral "personalities" that impact their coding performance.
• DeepSeek V4 + Claude Code = BEST AI Coder! — Combine DeepSeek V4 with Claude Code for a cost-effective, hybrid AI coding workflow.
• How I Choose Between GPT, Claude, Gemini and Open Source Models for Every Task — Optimizing LLM usage requires a decision framework matching tasks to models for cost and quality.
• ChatGPT vs Claude: I Used Both Every Day for a Month. The Winner Surprised Me. — AI tool selection should align with primary daily tasks, not just general capability.
• Bigger Isn't Always Better: A Comparative Evaluation of LLMs for Automated Code Review — Smaller, cheaper LLMs can outperform larger models for automated code review, particularly on real-world data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by KDnuggets.