Spending $5,000 Vibe Coding With Claude Fable 5.1

Summary
Matthew Miller, founder of BridgeMind, hosts a multi-hour live vibe-coding stream testing Anthropic's Claude Fable 5.1 model alongside newly released Gemini 3.8 Flash. Throughout the stream, Miller runs dozens of parallel coding sub-agents within the BridgeMind desktop app to automate customer support pipelines, develop voice-driven agent tools, and generate full 3D browser games.

What is shown

• Multi-Agent Orchestration & Infrastructure [00:10, 44:00, 73:45]: Miller utilizes BridgeMind's multi-pane interface to coordinate background agents (Claude Fable 5.1, Cursor Agent, Grok) reading Discord bug reports and programmatically filing and triaging tickets in Linear via an MCP integration.
• Gemini 3.8 Flash vs. Fable 5.1 Benchmark Tests [16:00, 33:20, 88:00]: Miller inputs identical prompts into Gemini 3.8 Flash and Claude Fable 5.1. Gemini 3.8 Flash rapidly compiles a Mario Kart clone and a Minecraft clone in under 10 minutes, but produces broken geometry, black screens, and corrupted void worlds [91:00], contrasted against Fable 5.1's coherent 3D tracks and voxel rendering [91:38].
• Subway Surfers Browser Clone [58:45]: A functional, one-shot 3D Subway Surfers endless runner built with Three.js by Fable 5.1, featuring procedurally generated tracks, coin collection, train obstacles, and synthesized audio.
• FIFA Soccer Game [103:10]: A 3D soccer exhibition match built with Three.js, featuring team selection (Spain vs. Argentina), stadium geometry, crowd audio, and animated player models, though hindered by sluggish keyboard controls.
• BridgeMind Voice Orb [115:00, 180:05]: Testing a voice-control system enabling full-duplex conversational interaction to navigate workspaces, inspect active terminal panes, and issue coding prompts to sub-agents via speech.
• Apex Formula F1 Game [138:05, 140:10]: A detailed 3D Formula 1 racing simulator built with Three.js, featuring a menu system, track selection (Kingsmere Circuit), engine audio, pit crew radio commentary, collision physics, and AI opponents.
• GTA 6 Web Clone ("Leonida Vice City") [258:10, 261:20]: An open-world urban driving and character game built in Three.js featuring narrative dialogue sequences, city block rendering, pedestrian spawns, entering vehicles, and driving mechanics, consuming significant RAM (17 GB in Chrome).

Claims & numbers

• API and Subscription Limits: Miller states Claude Fable 5.1 consumes limits extremely fast, exhausting a $200/month Claude Max subscription session cap in under 30 minutes [03:57]. He notes he is burning through roughly $15,000 in Cursor API credit allocation.
• Token Output Speeds: Miller cites Artificial Analysis benchmarks showing Gemini 3.8 Flash reaching ~305 output tokens per second [15:03, 30:50].
• CursorBench Scores: On CursorBench, Fable 5.1 Max is listed at 73.4% ($6.96/task), Grok 4.6 Extra High at 70.8% ($2.81/task), Fable 5.1 Extra High at 70.5%, and Gemini 3.8 Flash High at 69.2% ($2.38/task) [161:20].
• LM-Arena Rankings: On the Code Arena WebDev leaderboard, Claude Fable 5.1 Max is shown ranked #1 with an arena score of 1703 [119:40].
• BridgeMind Metrics: Live telemetry shows BridgeMind ARR fluctuating around $196,800 to $197,742 during the broadcast [81:50, 245:10].
• Search Trends: Miller highlights VidIQ analytics showing search volume for "Claude Code" peaked around 8 million in April 2026 and dropped to ~3.2 million [200:05].

Notable quotes

• [02:14] "Today we are going to be spending $5,000 on the newly released Fable 5.1, but buckle up, it's going to be a good one."
• [88:01] "Gemini 3.8 Flash was able to do in 8 minutes what took Fable 5.1 90 minutes."
• [140:15] "Did Fable 5.1 cook or what? Guys, I need Ws in the chat, this is insane!"

Assessment
This is a live, unedited developer stream showcasing raw coding workflows and agent generation capabilities. The presenter demonstrates real successes in complex UI and 3D web game generation, but openly displays and critiques failures, including game control bugs, severe browser memory bloat, and rendering failures produced by both models tested.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.