Same prompt. 15× difference in cost. Four completely different frontend implementations. I compared Claude Fable 5, @Zai_org GLM-5.2, DeepSeek V4 Pro, and Qwen3.7 Max in the @novita_labs Arena Playground using the exact same prompt. Task: Build a single-file HTML/CSS/JS animation where two airplanes collide mid-air with clouds, particles, and a pause button. Results 👇 ✈️ Claude Fable 5 💰 $0.5357 🪙 10,715 tokens ✈️ GLM-5.2 💰 $0.0343 🪙 7,796 tokens ✈️ DeepSeek V4 Pro 💰 $0.0604 🪙 18,885 tokens ✈️ Qwen3.7 Max 💰 $0.0290 🪙 7,729 tokens Same prompt. Different outcomes. Different animation quality Different UX decisions Different code structure Up to 15× difference in cost Benchmark scores are useful. But asking models to build the same real-world app reveals something different: how they design, reason, and implement. Which result would you ship? Try the prompt yourself 👇

3.8× cheaper. Same prompt. Same frontend challenge. Powered by @novita_labs . ⚡ 🟢 GLM-5.2 💰 $0.0518 🪙 11.8K tokens 🔵 GPT-5.5 💰 $0.1969 🪙 13.1K tokens Really impressed by what @ZAI_Org has achieved with GLM-5.2. Comparable frontend rendering, fewer tokens, and significantly lower inference cost.

Starting to build NovitaBox in public. NovitaBox is the local edition of Novita @novita_labs Sandbox: a secure MicroVM-based runtime for AI agents. Run agent code locally with low latency, strong isolation, and full data privacy. Same SDK. Local-first. Cloud-ready. We’ll share the tradeoffs, rough edges, and demos as we build.