What are people actually building with the newest frontier models?

People are building revenue-generating products with frontier models: solo developers earn $100K–$165K from AI-adapted games like GTA India ^5506,5112, trading bots generate $18,800/month at $480 API cost ⁶¹¹⁶, and architectural CAD files sell for $49/month yielding $58,800 in month one ⁵⁵⁰⁷. Fable 5 dominates for complex, multi-step tasks ^98,95 and as planner in multi-agent systems ^{6116,5920,5535}. Kimi K3 provides comparable quality at 3.3× lower cost for 3D prototyping ⁶¹⁰⁸ but its superiority is disputed ⁵³²⁸. GPT-5.6 Sol excels at robustness after adversarial training ^6112,5912 and one-shot cinematic websites ^5317,5316. Agent loops are categorized into four types ^6095,5302 and agent patterns into six reusable workflows ^6105,5916. Local stacks combining Obsidian, n8n, and Claude enable recurring revenue without cloud fees ^6104,5901. New practices include autonomous engineering via Claude+MCP+GitHub Actions ⁶⁵⁰⁹, Hermes agent controlling Unreal Engine 5.8 ⁶⁵⁰⁷, and 4-agent development teams producing PRs ⁶⁵⁰⁵.[24]

Practices confirmed by multiple reports: Fable 5 autonomously self-improves and outperforms Opus 4.8 on complex tasks (98, corroborated 19). Fable 5 with ClaudeKit's EngineerKit generates professional websites with one command ^171,8. Using /loop with objectives yields better autonomous results ^80,6. Fable 5 can build games rapidly ^164,5. Fable 5 planning with cheaper executors cuts cost without quality loss ^712,5. Adjusting effort levels halves Fable 5 costs ^97,5. Qwen3.6-27B with quantization achieves near 100K context on 32GB VRAM ^357,4. Gemini Omni Flash adapts video clips to short-form ^464,4. Fable 5 builds backtests and deploys Polymarket arbitrage bots ^609,3. A solo operator used Claude 4.8 Opus to earn $43,000/month via OnlyFans ^166,3. Claude Code with Higgsfield MCP produces cinematic websites ^308,2. Claude Opus 4.8 enables reliable autonomous systems with trigger-agent-verification ^358,2. New corroborated practices: removing examples from system prompts improves performance ⁶⁵⁰⁸, and 4-agent teams with file-based communication autonomously produce PRs ⁶⁵⁰⁵.[14]

Key tools include Claude Desktop with Higgsfield MCP for virtual tours ⁵⁷⁴³ and marketing campaigns ²⁴⁹. Local stacks using Obsidian, n8n, and Claude automate pipelines ^6104,5925. Codex Security plugin scans code via chat ⁵⁹⁰⁹. PromptQL offers free credits for Opus 4.8 and GPT-5.5 on code review ⁵⁹²⁷. ktransformers runs DeepSeek-V3 with 139K context on a 24GB GPU ⁵¹¹⁴. Hermes Agent enables local autonomous operations ⁸⁸⁴. Google's Agents CLI scaffolds and deploys ADK agents ⁸⁸⁸. Depth map technique improves dance transfer accuracy ⁶¹¹⁰. Shared project instructions ensure multi-model consistency ⁵⁵³⁴. Anthropic's Frontend Design skill forces aesthetic direction before coding ^5724,5118. New tools: MCP servers for Unreal Engine control ⁶⁵⁰⁷, GitHub Actions integration for autonomous engineering ⁶⁵⁰⁹, and Unlimited-OCR for 100-page PDF parsing ⁶⁴⁷⁵.[16]

Performance claims between Kimi K3 and Fable 5 conflict: Kimi K3 produced better game design at lower cost ⁵³¹⁴, but Fable 5 outperformed in direct coding tests ⁵³²⁸. Revenue from GTA India mods varies: $100K ⁵⁵⁰⁶ vs. $165,480 ⁵¹¹². Kimi K3 can rebuild a AAA game scene in 10 hours ⁵⁹¹³ unreplicated. GPT-5.6 Sol one-shotted a Minecraft-like game ¹¹⁰⁴ with limited verification. A solo developer claims $18,800/month from 7 Claude agents ⁶¹¹⁶. Non-coders build realistic games in days ⁵¹²⁶. The depth map method for dance transfer is single-source ⁶¹¹⁰. New contested claims: Claude Code on the web prevents cloning public repos ⁶⁵⁰⁴, and aggressively quantized models below 4-bit degrade agentic performance ⁶⁵⁰⁶.[11]