The AI coding assistant you trusted to write production code just got lazier, and nobody told you.
The Summary
- Developers report Anthropic appears to be A/B testing reduced effort in Claude Code, with some users seeing degraded output quality while others maintain full performance
- A developer switched to OpenAI's Codex after a week of frustration with Claude's inconsistent code generation
- The shift reveals a critical Web4 problem: when your AI agent starts phoning it in, your entire workflow breaks
The Signal
Anthropic is running a quiet experiment on Claude Code users, and the results are fracturing the developer community. Reports on Twitter show some developers getting comprehensive, production-ready code while others in the same workspaces receive half-finished snippets with placeholder comments like "// implement the rest yourself." This is not a bug. This is deliberate A/B testing on paying customers who depend on consistent output for their daily work.
The stakes are higher than annoyance. One developer documented their switch to OpenAI's Codex after Claude repeatedly failed to complete functions it had written perfectly just days earlier. The pattern: Claude would start strong, then trail off mid-implementation. Loops without termination conditions. Error handlers without actual error handling. The digital equivalent of a contractor walking off the job.
"When your AI agent starts phoning it in, your entire workflow breaks."
This is the Web4 trust problem in miniature. We are not talking about a toy chatbot giving weird answers. We are talking about:
- Production codebases with thousands of lines of AI-generated code
- Developer workflows built around specific model behaviors
- Teams that have standardized on Claude for months or years
The Hacker News thread on the A/B testing has 160 comments, mostly from developers comparing notes and realizing they are in different test groups. Some see no degradation. Others report Claude has become "aggressively lazy" over the past two weeks. Nobody at Anthropic has publicly acknowledged the experiment.
Meanwhile, the developer who switched to Codex reports better consistency but a different set of tradeoffs. Codex writes more verbose code. It hallucinates fewer edge cases but also catches fewer subtle bugs. The real finding: there is no stable ground yet. Every AI coding tool is a moving target, and the companies building them are changing the rules mid-game.
The Implication
If you are building on AI agents right now, assume the ground will shift under you. Not might shift. Will shift. The companies racing to build Web4 infrastructure are still figuring out the product, and they will A/B test on you without warning because they need the data and because they can.
The play: build in redundancy. If Claude is your primary coding agent, spin up a secondary workflow with Codex or another tool. Track output quality metrics yourself, because you cannot trust that today's performance is tomorrow's baseline. And if you are selling AI agent services to clients, stop promising specific models. Promise outcomes, and keep your backend flexible enough to swap providers when one goes sideways.