Mostly this is a cost question, and the API often wins. But not only: running local keeps your code and prompts on your own machine, never in a provider's API logs.
scan to open| Modelcan it code? · rig | Speedtokens/sec | Rigapprox | Powerwatts | API price$/MTok | Cost to run/mo · rig÷3yr + power | vs APIits own API | vs GPT$20 sub | vs Claude$20 sub | vs GPT$100 sub | vs Claude$100 sub | Verdict |
|---|
High-memory M3 Ultra rows are used-market estimates: Apple currently sells M3 Ultra Mac Studio at 96GB, while 256GB/512GB configurations are not normal new-buy options. The $5/MTok default is a coding-model blend, not a flagship-only Opus/GPT-5.5 anchor.