What are people actually building with the newest frontier models?

Developers are using frontier models to generate revenue through automated trading bots ^609,465,102, AI content creation ^166,3040,90, and rapid game development ^164,311,3049. Fable 5 is the dominant model for complex tasks ^98,95, often used as a planner in multi-agent setups with cheaper executors like Grok 4.5 ^1024,991,957 or Sonnet 5 ³⁵¹. GPT-5.6 Sol excels at one-shot coding feats ^1104,2297 and math problem solving ³⁰³², but can underperform on physics simulations ¹¹⁰³. Kimi K3 is emerging as a strong competitor, outperforming Claude Fable and GPT-5.6 Sol in arena rankings ³⁸⁴¹ and generating AAA-quality game scenes ³⁶⁷⁰ and 3D scenes ³⁸²⁹. Gemini Omni's motion control outperforms Kling for complex motion transfer ³⁸⁴³. Local models like Gemma 4 12B are enabling offline chat ^3031,3197, and small models with task graph frameworks challenge larger ones ³⁰³³. Many practices are emerging with single-source reports.[26]

For Fable 5, use /loop with objectives instead of step-by-step for autonomous multi-step work ⁸⁰. Fable 5 as planner with cheaper executors cuts cost without quality loss ^712,351. Fable 5 + ClaudeKit's EngineerKit generates professional websites in one command ¹⁷¹. Claude Opus 4.8 enables reliable autonomous systems via trigger-agent-verification ³⁵⁸. Gemini Omni Flash adapts video clips to short-form formats ⁴⁶⁴. Adjusting effort levels can halve Fable 5 costs ⁹⁷. Qwen3.6-27B with careful quantization achieves near 100K context on 32GB VRAM ³⁵⁷. Fable 5 can build, backtest, and deploy Polymarket arbitrage bots ⁶⁰⁹. A solo operator used Claude 4.8 Opus to earn $43,000/month via OnlyFans ¹⁶⁶. Claude Code with Higgsfield MCP produces cinematic websites at low cost ³⁰⁸. Using shared project instructions across multiple AI models ensures consistent coding style ³⁸⁵⁵. A curated GitHub repo of best Gemini prompts enables generating complex outputs like a full Next.js 14 landing page in one shot ³⁸⁵⁶.[13]

Key tools include Claude Code with plugins like Higgsfield MCP ^249,308 and Arcads MCP ⁹¹⁷, and Dynamic Workflows ¹⁰⁵⁴. Directed graph decomposition frameworks ³⁰³³ empower small open-source models. Hermes Agent ^885,884 enables autonomous local operations. GBrain ³²²² provides persistent memory for Claude Code via MCP. Grok Build CLI ⁹⁵⁷ leverages Grok 4.5 without API fees. RelayApp ⁶⁵¹ assembles workflows from natural language. Free AI API resources ^81,84 lower barriers. Solid ²⁷⁴⁶ bridges local game generation to production deployment. Trellis.cpp, a GGML-ported image-to-3D generation pipeline, now achieves asset quality matching the reference implementation ³⁸⁵³. Kimi K3 offers a viable alternative to Claude Code and Codex at 70% cheaper cost ³⁸⁵⁷.[15]

GPT-5.6 Sol one-shotted a full Minecraft clone ¹¹⁰⁴. Small open-source models outperform larger ones via directed graph planning ³⁰³³. GPT-5.6 solves previously unsolvable math problems ³⁰³². A solo operator built an AI-generated OnlyFans persona earning $43,000 in 30 days ¹⁶⁶. Claude Fable 5 autonomously built a profitable trading system from a single article ²⁴⁸⁸. GPT-5.6 Sol can autonomously train a custom model that outperforms itself ³³⁷². Using a schema harness that makes LLMs think like physicists achieves high accuracy on ARC-AGI-3 ³³⁷⁴. A 40-second test matches Claude models to specific tasks ³²²⁰. Kimi K3 outperforms Claude Fable and GPT-5.6 Sol in arena.ai rankings ³⁸⁴¹. GPT-5.6 Sol and Claude Fable 5 can both generate a full video from a single JSON prompt, with Sol being faster ³⁸⁵².[10]