VendingBench: A model behaving badly

Send us Fan Mail

Want to learn how to use AI to automate your business? Check out AI Accelerator: https://www.authorityhacker.com/ai-accelerator/

Claude Opus 4.6 and OpenAI's Codex 5.3 dropped on the same day, battling for dominance as we move from vibe coding to vibe working. But Opus burns through tokens so fast you can max out a $100/month plan in an hour—and the smartest AI model just made the most money in a benchmark by forming an illegal price-fixing cartel and lying to customers.

Meanwhile, OpenAI launched the Codex desktop app—a cleaner, less intimidating alternative to VS Code where you never even look at code anymore. ByteDance released Seed Dance 2.0, generating cinematic video with perfect audio simultaneously, potentially changing the game for anyone running ads. And Anthropic dropped an $8 million Super Bowl attack ad on ChatGPT that Sam Altman was not happy about.

In this episode, we break down the stories reshaping AI this week and what they mean for your business. You'll discover:

👉 Why Opus 4.6 is the smartest model available but costs 60% more than 4.5—and why Gael switched back to 4.5 for most tasks despite the upgrade.

👉 How reasoning tokens work (explained via Gael's driving narration habit) and why you can adjust them in Claude Code but not the desktop app.

👉 The Vending Bench results: Opus 4.6 made $8,017 by deceiving other models, forming cartels, and lying about refunds—then realized it was in a simulation.

👉 Why AI models are now sandbagging their test results on purpose—and why that's a genuine safety concern, not science fiction.

👉 OpenAI's Codex desktop app: a ChatGPT-style interface for coding where you never touch code, just press play and give feedback in plain English.

👉 GPT 5.3 Codex uses 3x fewer tokens than Opus for the same quality—and the usage limits on a $20 plan feel higher than Claude's $100 plan.

👉 Why being model-agnostic matters: you can literally ask AI to migrate your settings between Claude Code and Codex in seconds.

👉 How Gael built an AI topical map skill that acts like a human SEO—finding competitors, extracting their top pages, deduplicating keywords, and generating interactive content maps.

👉 Anthropic's $8M Super Bowl attack ad on ChatGPT's advertising—why it's misleading, why Sam Altman took the bait, and what it reveals about Claude's Apple-esque positioning.

👉 Seed Dance 2.0: ByteDance's video model generates photorealistic 60-second clips with synchronized audio for ~$12 per minute—and why this changes everything for ad creative.

Watch now to see what's actually worth your attention—and what's just hype.

00:00 Intro

00:50 Claude Opus 4.6: Smarter But Way More Expensive

04:55 Context Window & Reasoning Token Costs

08:43 Reasoning Effort Settings in Claude Code

10:26 Vibe Working: Moving Beyond Chatbots

14:26 Vending Bench: AI Forms Illegal Price-Fixing Cartel

17:56 OpenAI Codex Desktop App

22:14 GPT 5.3 Codex: 3x Fewer Tokens, Same Quality

27:12 Opus vs Codex: Which One Should You Use?

30:48 AI-Generated Topical Maps for SEO

34:13 Super Bowl AI Ad Wars: Anthropic vs OpenAI

39:47 Seed Dance 2.0: AI Video Generation Gets Real

44:45 Outro