Coding-agent vendors under fire: effort levels, downgrades, and usage caps

SlopCodeBench's October results (Opus 5.5, Astra winning overall, "xhigh considered harmful") drew a practical reaction: people want a representative subset to run themselves, and one user finds medium effort gives the best output for most tasks. Another grumbled the report reads full of Claudisms.

Anthropic's own post on using Claude to reduce code attack surface got a hostile reception. Commenters say Claude silently drops to a less capable model on anything security-adjacent, even for members of the Anthropic Cyber Verification Program. An older thread shows the same friction: Anthropic stopped letting Claude Code subscriptions run through OpenClaw, citing outsized strain. Users answered that it is the capacity they paid for.

The pattern: developers depend on these tools daily and now treat the vendor as an opaque variable. Silent model swaps, rate limits that hit every afternoon, and harness restrictions are the complaints. The Rails thread adds a calmer angle: conventional frameworks work well with agents because they limit reinvention.