Theo's decisions

  • He rejects Gemini and standard Muse Wonder as sources of useful T3 code contributions, while explicitly distinguishing Muse Wonderful Full. He proposes automatically closing PRs made with the rejected models; this is an intention, not a policy shown as implemented. 36:30
  • He says reviewing Gemini's code cost more time and tokens than using a stronger model from the start. His remaining Gemini uses are YouTube data retrieval and possibly an early pass on an attractive marketing page. 37:06
  • He reports merging none of Gemini's output from this testing. His closing preference is primarily Opus 5.5, despite joking that he will stick with Muse. 37:30 38:14

What the review leaves open

  • Theo does not test the limits of the $200 subscription plan. His comment that usage might feel effectively unlimited is frustration, not a measured quota result. 36:12
  • Specialized vision and multilingual work remain incompletely tested; the favorable Skatebench result is concrete, while broader specialist recommendations require task-specific evaluation. 9:47 16:10 16:37
  • Rumors of a better "Carbon" snapshot are unconfirmed, and Theo does not know how much better it would be. They do not change his verdict on the reviewed model. 37:56
  • His conclusion rests on his private audits, game demonstration, implementation experience, and workflow preferences. It supports a strong negative judgment about his code work while retaining the specific exceptions documented elsewhere in the notes. 27:33 33:33 37:06