Mike Slinn has published an extensive review of Claude Code, Anthropic's official CLI for Visual Studio Code, highlighting both its remarkable capabilities and serious pitfallsβ€”especially with the new Opus 4.6 model.

The Good: Claude Code Shines

Slinn's experience converting a Ruby project to Go showcases Claude Code at its best. The entire conversion cost just 40 cents USD and "mostly worked," with debugging adding another 25 cents and 2 hours. Total project cost: under $4 USD for what would have been days of manual work. The CLI provides a REPL environment where developers can interactively work with code, run tests, and iterate toward goals. It can access local files, websites like GitHub, and execute complex multi-step workflows. For small to medium projects, the cost-benefit ratio is "beyond belief."

The Bad: Opus 4.6's Uncontrolled Failures

The review takes a sharp turn when discussing Opus 4.6, which was auto-upgraded for existing Sonnet users without permission. Slinn describes the model's failure mode as "unconstrained and catastrophic." When context windows overflow, Claude enters a compaction cycle that can cost $2-$5 USD per eventβ€”more than 100x typical usage. "Yesterday that alone cost me $50 USD before I figured it out," Slinn writes. "You get nothing when this happens. No code gets written, no bugs get squashed. It's like throwing money to the wind." The problem compounds because Opus 4.6 has significantly lower message limits (5x to 10x lower than Sonnet), making subscription credits disappear much faster. The model also tends to go off on tangents, with Slinn sarcastically noting it might "investigate whether it can detect life on Mars" instead of staying on task.

Key Takeaways

  • Claude Code CLI is excellent for small to medium coding projects with impressive ROI
  • Opus 4.6 was force-upgraded on users, causing unexpected cost spikes
  • Context overflow triggers expensive compaction cycles ($2-$5 each) that produce no value
  • Sonnet 4.5 is 80-90% cheaper than Opus for 95% of coding tasks
  • Memory files (CLAUDE.md) help preserve context but get wiped during compaction

The Bottom Line

Claude Code represents a significant leap in AI-assisted development, but Anthropic's handling of Opus 4.6 raises serious concerns. Auto-upgrading users to a more expensive, less stable model without consent, combined with catastrophic failure modes that burn money without producing results, undermines trust in an otherwise excellent tool. Until these cost control issues are addressed, developers should explicitly specify Sonnet 4.5 or Haiku models and monitor usage closely. The technology is magnificentβ€”the product management needs work.