What are people actually building with the newest frontier models?

People are building revenue-generating applications with frontier models: automated trading bots ^601,609,102, legal entity setup ¹⁰¹, faceless YouTube channels ^92,467, AI-generated ad creation ^90,210,599, and an OnlyFans persona earning $43,000 in 30 days ¹⁶⁶. The dominant pattern is pairing a strong planner (e.g., Opus 4.8) with a fast executor (e.g., Fable 5) for cost-performance gains ^88,351. Fable 5 is the most cited model, used for complex multi-step tasks ^95,538, autonomous self-improvement ⁹⁸, and rapid game development ^164,311,439. Newer practices include using GPT-5.6 Sol for procedural 3D modeling ¹¹⁶³, solving math problems ¹¹⁶², and generating Minecraft-like games ¹¹⁰⁴; Grok 4.5 for slide decks ¹⁰⁴⁹ and browser tasks ⁹⁹⁴; and local models like Qwen3 for voice assistants ¹¹⁶⁸ and memory systems ¹¹³⁸. Many claims remain emerging with single-source corroboration.[25]

For Fable 5, using /loop for autonomous multi-step work and giving objectives instead of step-by-step instructions yields better results (corroborated by 5 sources) ⁸⁰. Fable 5 can autonomously self-improve and outperform Opus 4.8 on complex tasks at lower cost (corroborated by 15 sources) ⁹⁸. Fable 5 enables rapid full-game development from idea to code in about one hour (corroborated by 4 sources) ¹⁶⁴. A solo operator built an AI-generated OnlyFans persona using Claude 4.8 Opus, earning $43,000 in 30 days with $400 monthly compute cost (corroborated by 2 sources) ¹⁶⁶. Combining multiple frontier models can automate ad creation from product photos to video (corroborated by 3 sources) ⁹⁰. Using Fable 5 and ClaudeKit's EngineerKit, a single slash command can generate a professional website without a designer or developer (corroborated by 8 sources) ¹⁷¹. Using native Obsidian skills instead of generic HTML parsing reduces context usage by over 99% for AI agents searching vaults (corroborated by 3 sources) ¹⁷⁰. A leaked system prompt can recreate a model's personality on a different base model (corroborated by 2 sources) ⁸⁵. Fable 5 can produce high-quality, animated websites by combining reference screenshots, pre-built components, and specific animation tools with iterative prompting (corroborated by 6 sources) ²⁵⁴. Fable 5 can transform a folder of notes into a connected knowledge graph that answers in your own words and accelerates learning (corroborated by 2 sources) ²⁵². Running Claude Code on an e-ink tablet (reMarkable Paper Pro) creates a distraction-free, eye-comfortable coding environment (corroborated by 2 sources) ²⁴⁶. Claude Opus 4.8 enables reliable autonomous systems that run unattended for hours, using trigger-agent-verification pattern (corroborated by 2 sources) ³⁵⁸. Claude Code with Higgsfield MCP can produce a fully animated, scroll-driven website with cinematic effects for a fraction of traditional cost (corroborated by 2 sources) ³⁰⁸. Fable 5 is designed for complex, multi-step tasks that take humans hours or days (corroborated by 3 sources) ⁹⁵. Fable 5 can be given a goal, readiness criteria, and limits to autonomously plan, write, test, fix, and deliver a result (corroborated by 2 sources) ⁵³⁸. Sonnet 5 offers Fable-style agent orchestration at a cheaper tier but may not replace Opus based on testing (corroborated by 2 sources) . Claude Opus 4.8 can generate fully interactive 3D websites from a simple prompt (corroborated by 2 sources) . Different frontier models produce distinct creative outputs when given the same product brief in an agentic pipeline (corroborated by 2 sources) . By adjusting effort, prompts, caching, batching, budgets, and routing, Fable 5 costs can be halved (corroborated by 4 sources) . Fable 5 is free until July 7; use it with Opus 4.8 planning to clone software, audit Claude Code setup, review codebases, or build long-horizon apps (corroborated by 4 sources) . With careful configuration, Qwen3.6-27B at Q8 quantization can achieve near 100K context on 32GB VRAM (corroborated by 3 sources) . Using a frontier model like Fable 5 for planning and reasoning, then delegating execution to cheaper models, yields near-identical results at lower cost (corroborated by 4 sources) . A pipeline combining Claude Opus 4.8 for research, NotebookLM for analysis, and Obsidian for storage can automatically populate a knowledge vault with structured markdown in minutes (corroborated by 2 sources) . Gemini Omni Flash can power a tool that automatically adapts video clips into short-form formats optimized for engagement (corroborated by 3 sources) . Sonnet 5 is designed as a fast, cheap implementer in a multi-agent stack, not a standalone benchmark winner; paired with Fable 5 as advisor, it outperforms solo models (corroborated by 3 sources) . Fable 5 can write, backtest, and deploy a Polymarket arbitrage bot that achieves 98% win rate (corroborated by 2 sources) . Anthropic's free workshops provide a hands-on playbook for building self-improving Claude agents with memory, autonomy, and tool use (corroborated by 2 sources) . A CLAUDE.md file should contain project-specific rules earned from mistakes, not generic advice copied from GitHub (corroborated by 2 sources) . Grok 4.5 can replace Sonnet 5 as an implementer agent in a multi-agent setup, leveraging Grok Build CLI to avoid API fees (corroborated by 2 sources) . Programmers can replace Opus and Sonnet with Grok 4.5 in their systems, achieving 2-3x speed and 50-80% cost reduction (corroborated by 2 sources) . Grok 4.5 is cheapest and fastest but writes more code; Fable 5 produces leanest code but is 22x more expensive (corroborated by 3 sources) . Qwen3.5 122B MoE model outperforms smaller dense and MoE models for complex tool-calling tasks like data extraction from PowerPoints (corroborated by 2 sources) . GPT-5.6 Sol as an agent in ChatGPT Codex can autonomously complete complex multi-step tasks like comparing insurance quotes across companies (corroborated by 2 sources) . GPT-5.6 Sol is an excellent implementor and can be used as a subagent for Fable to achieve superior coding results (corroborated by 2 sources) . GPT-5.6 can generate a functional Minecraft-like game from a single prompt (corroborated by 2 sources) . GPT-5.6 Sol Ultra produces weaker physics simulations than GPT-5.5 despite higher cost and token usage (corroborated by 2 sources) .[36]

Frontier models: GPT-5.6 Sol (reasoning levels Light to Ultra) ^1303,1191, Claude Fable 5 (agentic orchestrator) ^213,503, Claude Opus 4.8 (planner) ^358,466, Claude Sonnet 5 (fast implementer) ^351,437, Grok 4.5 (cheap/fast coder) ^991,1024, Gemini Omni Flash (video adaptation) ⁴⁶⁴, Qwen3.5 122B MoE (tool-calling) ¹⁰²⁵, Gemma4 (local) ³⁵⁷, Hy3 (open-weight) ⁷¹⁵, GLM 5.2 (cost-effective) ⁶⁵⁰, Muse Spark 1.1 (low-cost agentic) ^1463,1462. Local hardware: Mac Mini with Ollama ¹⁶⁹, Strix Halo ¹²⁴⁸, two RTX 3090s ²⁵¹, Ugreen NAS ²¹². Agent frameworks: Claude Code ^308,990, Codex ¹⁰⁴⁷, Hermes Agent ^885,884, n8n ⁹², OpenSquilla ⁷¹⁷, ATG ^1690,1534. MCP servers: Higgsfield ^249,308, Arcads ¹¹⁶¹. Knowledge management: Obsidian ^{170,789,847,1137}, GBrain ¹¹³⁸. Image/video: Nano Banana 2 Lite ⁹²², Seedance 2.0 ^209,474, Kling 3.0 ²¹⁰.[42]

Most claims are emerging with single-source corroboration. Notable unverified claims: GPT-5.6 can autonomously build a complete training pipeline and train a custom language model from a single prompt on a Mac ^{1628,1468,1308,1542}. A home AI server with Hermes Agent and GPT 5.5 can interpret camera input and control physical outputs via Arduino ^{1627,1541,1467}. Voodoo Quant achieves 95% KLD improvement over Unsloth Dynamic ¹⁶²⁶. A loop-based AI workflow can automate research, drafting, and self-critique requiring only one human approval per final article ¹⁶⁹⁸. Using GPT-5.6 Sol with a verifier loop can produce sites indistinguishable from agency builds ¹⁶⁹⁷. Antigravity skill allows non-coders to use advanced AI for history analysis with plain English ^{1696,1464,1247}. Using AI to build a living knowledge graph in Obsidian that updates daily improves memory and learning ¹⁶⁹⁵. GPT-Live's voice mode can turn a 10-minute voice dump into a structured brief ¹⁶⁹⁴. Hunyuan3D models can be ported to MLX to run image-to-3D on Apple Silicon ¹⁶⁹³. App founders can scale to $10k/mo by using AI to generate slideshow content ^1692,1461. Before losing access to a frontier model, conduct an exit interview to extract tacit knowledge into reusable skill files ^{1691,1620,1537}. Atomic Task Graph (ATG) framework enables small open-source models to outperform GPT-4 on complex agent benchmarks without fine-tuning ^{1690,1534,1459,1404}. Using Claude Fable 5 as a spec writer and reviewer with Grok 4.5 as the implementer creates a cross-vendor review workflow ^1689,1394. Loop engineering, where one AI model acts as a manager spawning subagents, is more effective than single-prompt interactions ¹⁶³⁰. Using a tiered AI team with Fable 5 as manager, Sol as senior worker, and Luna as intern army is more effective than relying on a single model ^1629,1544. GPT-5.6 can analyze raw sales data, score campaigns, and build an interactive dashboard for budget allocation ¹⁶²⁵. Frontier models can one-shot complex game builds from a single prompt ¹⁶²³. Fable 5 can generate a detailed 3D globe with textures and atmospheric glow from a flight dataset . A Jacobian lens visualizer and steerer for GGUF models can be built using Fable 5 . Following OpenAI's official prompting guide for GPT-5.6 can make the model 10x more efficient . Using Fable 5 and GPT image2, a turn-based battle game scene can be built in one day . For coding work with GPT-5.6, the 'Sol on Medium' reasoning effort may be a good default . GPT-5.6 can generate custom fonts for branding . Using Claude Code, a developer created a web-based iOS simulator that streams a live iOS app built on a headless Mac Mini . Claude Fable 5 can be structured into an Agentic OS with knowledge graphs and AI swarms, cutting token usage by 15x . Two frontier models collaborating can produce a complex browser-based combat simulator with terrain, radar, and HUD in a few hours . Successfully extended Gemma4-31B to 40.5B by inserting new layers . GPT-5.6 wrote the full codebase to train a custom 1.38M parameter Transformer on iMessage history, all locally on a Mac . Meta Muse Spark 1.1 is a multimodal reasoning model optimized for agentic tasks . NVIDIA's TwoTower method adapts pretrained autoregressive models for fast parallel generation . Andrew Ng's course teaches how to build agentic skills with Claude . A single Claude-powered AI can replace human coaching for form analysis, scaling to hundreds of users . Using Claude to code Pine Script strategies and refine them, one can build a profitable trading bot . GPT-5.6 can assist in solving previously unsolvable math problems when used by an expert mathematician . Using open interest data from options chains, one can build a regime classifier that predicts volatility clustering and price direction . GPT-5.6 Sol excels at video creation and editing, outperforming Fable . GPT-5.6 Sol can be used to build premium cinematic websites .