If Claude Code has felt inconsistent lately, you might not be imagining it.
Reports from developers suggest that Anthropic may be running A/B tests within Claude Code, the company's AI-powered coding assistant. The practical effect: some users get routed to different model configurations or versions behind the scenes, which can produce noticeably different output quality during the same session or across sessions.
Anthropic hasn't publicly confirmed or denied running such tests, which is standard practice for AI companies iterating on their products. A/B testing lets companies evaluate model changes on real traffic before rolling them out broadly. The downside is that users paying for a premium coding tool may not realize they're getting a degraded experience as part of an experiment.
For developers relying on Claude Code for production work, the practical advice is straightforward: if output quality drops suddenly, try starting a new session. If the problem persists across multiple sessions, it's more likely a genuine regression than an unlucky test bucket. Tracking which responses feel off and reporting them through Anthropic's feedback channels is the most direct way to signal that a test variant isn't working.