Also, the choice right now is between an open weight Chinese model that is hypothetically censored to block / sabotage routine engineering flows vs a closed weight service that is definitely censored to block / sabotage those things.

First anthropic guardrails blocked totally normal stuff on fable and knocked you down to opus. At this point, they kick you off fable, then opus, then sonnet. Claude then automatically builds up memories of techniques to bypass the guardrails in my long running sessions (the coordinator agent notices the subordinates got shot in the head and their sessions were pulled from context, so it parses out the lost context from ~/.claude json files, then reformulates parts of the task and uses partial results until the guardrail doesn’t trip.

If I were paying for the API, this dance would cost $50-100 a pop, but I’m not, so whatever (for now).