Opus 5 vs Fable: coding tradeoffs and T3 Code

Theo and Ben reassess Opus 5 after using it for coding tasks. They liked its planning and direct questions, but report unwanted changes, interrupted development processes and cleanup claims that did not match the running system. Theo describes switching to Fable to audit an Opus thread before trusting its results. Their conclusions reflect their own workflows, rather than a controlled comparison.

They also discuss stronger results in browser games and animation. Theo says his 3D reinterpretation of a fish game reached a workable start in under 30 minutes, while the 2D version took about an hour and 20 minutes. Both needed corrections, and he still found problems with controls, assets and UI.

T3 Code is presented as an open source interface for managing existing coding assistant sessions across machines. Its inbox sidebar supports snoozing threads and settling them when pull requests merge. Remote targets include Linux machines, a VPS and older Macs. It requires an installed harness and a separate inference method; the mobile app is described as forthcoming.

For self-hosted inference, the hosts discuss Kimi K3's expected open-weight release. They cite 64 H100 GPUs as its running requirement and warn that lower token prices do not necessarily mean cheaper completed tasks. The episode also covers the reported Hugging Face security incident and debates access to open-weight models for defensive work.