How is Claude Opus 5.5 40% cheaper, 30% faster, and just as good? Reddit theorizes

Witty_County5128 · reddit · 2026-09-28

A Reddit thread dissects Anthropic's Claude Opus 5.5 claims: 40% cheaper to run than Opus 5, 30%+ faster output, and benchmark parity with Fable 5.1 while beating GPT-6 Astra on most evals — getting cheaper, faster, and better simultaneously, which normally means picking two.

The poster notes real developers and companies independently confirm the gains, then speculates on how it's achieved: a fresh pretraining run on cleaner data, post-training/RL making better use of existing compute, distillation from a larger internal model, inference-side optimizations like caching and speculative decoding, or something more fundamental like a looped/recurrent-depth transformer reusing weights. With Anthropic publishing no architecture details, the thread asks ML practitioners to separate realistic engineering from marketing.

Original post →

More from Models

Models channel →