Opus 4.5

Boris Cherny, creator of Claude Code at Anthropic, uses «Opus 4.5» exclusively, and his reasoning inverts the industry's usual latency calculus. Michael Nuñez reported on VentureBeat in January 2026 that Cherny runs five Claude agents in parallel across numbered iTerm2 tabs, plus five to ten more sessions on claude.ai, all on Anthropic's heaviest and slowest model. His justification: the model needs fewer corrections and handles tool use well, which makes it faster in practice than a lighter model. The bottleneck he identifies is not generation speed but the human time spent fixing the AI's mistakes.

That claim is what keeps «Opus 4.5» recurring here. It is a data point in an argument about where the real cost of agentic coding sits. If the constraint is human correction time, then model weight is a throughput decision rather than a latency one, and running many slow agents in parallel becomes coherent.

The «Claude Code» curriculum places Opus alongside Haiku and Sonnet as the three models available in the CLI, matched to different levels of task complexity. «Opus 4.5» carries the complex end: multi-step planning, and delegation to subagents. Cherny pairs it with specialized subagents (a code-simplifier for architecture, a verify-app for end-to-end tests) and with a verification loop that iterates until the code works and the interface behaves.

What the sources do not settle is whether Cherny's parallel-agent economics hold outside a team building the tool itself.