Can GPT-6 Supervise Local AI? Qwen & DeepSeek Tested

What happens when a frontier model supervises a local coding agent? I pair GPT-6 Astra with Qwen3.8 Flash Next on my M5 Max, then repeat the task with DeepSeek V4.1 Flash through OpenRouter. Two teams, the same small website brief, and two different outcomes. Watch the project folders appear on my Mac, the agents write separate HTML, CSS and JavaScript files, and Astra send the API worker back for a specific correction. I also show how to create a reusable agent, select a Codex subscription model and reasoning effort, and choose how the team collaborates in MaiaChat. MaiaChat is a work in progress that I've been building over the past several months. With enough clear interest, I intend to make it open source and build it with the community. If you'd use it, like the video, subscribe, and tell me what you'd build with it. CHAPTERS 0:00 The experiment 0:30 Why build this? 1:12 A project on your Mac 1:56 Create a reusable agent 2:42 Subscription, API and local models 3:29 Choose how the team works 4:18 The local setup 5:00 Qwen builds and checks 5:41 The local result and the handoff 6:28 Switching the worker to DeepSeek 7:10 The supervisor asks for a fix 8:00 Testing the corrected application 8:51 The numbers we can report 9:43 What remote access means here 10:28 Work in progress, built with you 11:19 What should we test next? THE SETUP • Supervisor: GPT-6 Astra via Codex subscription, High effort. • Local worker: Qwen3.8 Flash Next GSQ-RCO, IQ3_S, LM Studio, 64,000-token configured context. • Hardware: Apple M5 Max MacBook Pro, 128 GB unified memory. • API worker: DeepSeek V4.1 Flash via OpenRouter. No explicit worker reasoning-effort override selected. • Both workers: temperature 0.7; output limit 8,192 tokens per call. • Task: Weekend Lens, a five-file offline photography-trip planner. RESULTS AND LIMITS Qwen: 30m 31s before I stopped it. It built a working application and passed core browser checks, but did not reach Astra's review. DeepSeek: 5m 36s, including the worker correction and Astra's acceptance. These are individual case studies in this setup, not a universal model ranking. DeepSeek reported 889,258 input and 34,953 output tokens. Its worker-only API estimate is about $0.09 using uncached OpenRouter catalog rates checked on 6 October 2026. That excludes Astra, can differ from billing, and is not an invoice. Attributable Astra usage and Qwen's cancelled-run token total were unavailable. Remote use requires the host to stay awake and online, a connected bridge, appropriate workspace permissions, and the local server when using local models. Cloud supervision can send relevant task context to the cloud provider. This episode demonstrates the connection on this setup; off-site and physical-phone tests are separate. MaiaChat: https://maiachat.com Previous Qwen local-model test: https://www.youtube.com/watch?v=TjhFzscnad4 OpenRouter model/rate catalog: https://openrouter.ai/api/v1/models PRODUCTION Screen recordings show the captured application runs; selected waits are shortened. Narration and brief avatar appearances use my authorised HeyGen voice/avatar. The original background instrumental was generated in Gemini for this video. The thumbnail is an AI-assisted illustration. #LocalAI #AIAgents #LMStudio