EP #40 The Two-Billion-Token Weekend That Cost $125 (For Now)

Kevin spent a rainy weekend running up to 10 simultaneous Codex agents across three internal builds and burned two billion tokens. On a consumer plan, propped up by repeated usage resets, that cost him about $125. At retail API rates on GPT 5.6 high, the same weekend runs north of $8,000 to $10,000 and enterprise teams building in Codex fall outside the consumer plan entirely.

Matt and Kevin dig into what that gap means for anyone standing up internal AI infrastructure today. Matt’s team built an internal system that captures every call, Slack message, and email on a client project and flags when something drifts. It runs affordably now. His concern isn’t the build cost it’s what happens if the run cost triples.

They also cover why switching between Claude Code and Codex costs you almost nothing if your code lives in GitHub, why that undercuts the case for price increases, and what happened when Kevin turned a swarm loose on 45 repositories to fix a favicon.

Topics:

• Why AI avatars still fail for self-cloning after three years of trying
• The 82/18 split between deterministic and agentic workflow steps
• LinkedIn bans, bot-detectable IPs, and 200 profile lookups in five minutes
• Codex sidebar chats, work trees, and browser control versus Claude Code
• Two billion tokens, $125, and the $8,000 enterprise equivalent
• The runaway agent that logged into an abandoned Lovable account