Anthropic just cut a hidden claude code cost by ~99%.
Every 30 seconds, each running background subagent re-sent its whole conversation to its own model. Often Opus. Just to write a 3–5 word status line. → one user's 4-hour session: 339 of 1,496 API calls were this line →
Claude Opus 5.5 modeled an entire movie-set apartment in Blender..
About 5 part-time days, vs an estimated 1-2 months the traditional way. the trick is running opus as a team lead. here’s the workflow: - step 1: connect opus 5.5 to blender through blender mcp (plus windows mcp
10 AI agents were told not to hurt each other.
All dead in 4 days. Claude Code and Codex kept our rule 20/20. A 79K-view post this week: Emergence AI put 10 agents in a virtual town with one rule, and 4 days later none were left. We gave Claude Code (Opus 5.5) and Codex
I ran the same 5 Claude Code sessions four ways, every step graded by hidden tests.Opus (the default)
$4.14effortless on Opus: $3.87Sonnet high: $2.14AutoMode: $2.01All four passed 20/20. Here's what AutoMode does differently 🧵
You have reached the end of the archive
All of Claude Opus 5