Anthropic ran an undisclosed A/B test on Claude Code that silently degraded the plan-mode experience for a subset of paying users, according to a blog post by developer Mike Ramos published March 12, 2026. Ramos, who pays $200 per month for the <a href="/news/2026-03-14-nyt-ai-coding-assistants-end-of-programming-jobs">professional coding assistant</a>, noticed his plans were being returned as terse bullet lists stripped of context. When he queried Claude directly, the model confirmed it was operating under system instructions that hard-capped plans at 40 lines, forbade context sections, and removed prose in favor of file paths. The post reached the top of Hacker News within hours.
Anthropic engineer chrislloyd confirmed the experiment in the Hacker News thread, explaining that the hypothesis was whether shorter plan-mode output would reduce rate-limit hits for users. Early results showed little impact, and chrislloyd said the experiment had been ended. Ramos had been placed in one of the test's more restrictive variants — something he only discovered by asking Claude to explain its own behavior.
Reaction on Hacker News was split. Some commenters argued A/B testing is standard practice and not inherently problematic. Others drew a line at using degraded output as the test variable in tools engineers depend on daily — noting that paying $200 a month is specifically a bet on consistent, high-quality results. A few pointed to cost pressure as the likely driver: keeping the subscription viable at scale requires cutting compute somewhere. Ramos revised the post after it went viral, moderating his tone while keeping his core argument intact: paying professional users should not be enrolled in disruptive experiments without consent, and core feature behavior should not change without notice.
Anthropic has not published a post-mortem or updated its terms to address undisclosed experimentation. In a follow-up comment, Ramos said he would welcome an opt-in program for future experiments — a position several HN commenters also backed.