Anthropic Admits Its Own Bugs Broke Claude Code After Weeks of Denial

Anthropic admitted that engineering missteps caused a monthlong decline in Claude Code performance, affecting paying users who reported the tool as less capable and more expensive to use. The company's April 23 postmortem detailed three separate bugs introduced between March 4 and April 16, including a reasoning effort change, a caching bug, and a system prompt instruction, all resolved in Claude Code version 2.1.116. Fortune reported that a senior AI executive at AMD called Claude 'unusable for complex engineering tasks' during the worst stretch, and Anthropic also faced criticism for initially pushing back on user complaints before publishing the postmortem.

Anthropic did eventually own the Claude Code regression, but the real damage was the weeks in which paying users were told, directly or indirectly, that the problem was mostly theirs. You pay $20 a month for Claude Pro expecting the tool to work the same way it did last week. For a chunk of Anthropic's user base this spring, it didn't. Claude Code felt less capable, more forgetful, and more expensive to use because mistakes were burning through limits. Users were right. According to reporting from Fortune and Yahoo Finance, Anthropic said engineering missteps were behind a monthlong decline in Claude Code performance that sparked what Fortune called a user revolt. The company first pushed back on complaints, then pointed to changes made for latency and token use, and only later published a detailed engineering postmortem. That's the part customers remember. Nobody enjoys being told the dashboard is fine while the product in front of them is plainly wobbling. Fortune reported that a senior AI executive at AMD called Claude "unusable for complex engineering tasks" during the worst stretch. That's not a fringe complaint. You don't need to agree with every angry post on X to see the problem: Claude Code has become part of how real developers work. Not a toy sitting in a side tab. When it starts reading less context, forgetting its own reasoning, or stopping too early, the failure lands inside pull requests, deadlines, and security reviews. The Bugs Were Ordinary Anthropic's own April 23 postmortem is the strongest source here, and the details are oddly mundane for something that cost the company weeks of trust. On March 4, Anthropic changed Claude Code's default reasoning effort from high to medium to reduce long latency that could make the interface look frozen. Anthropic later called that the wrong tradeoff and reverted it on April 7. Speed went up. Quality went down. On March 26, the company shipped a caching change meant to clear older thinking from sessions that had been idle for more than an hour. A bug made that clearing happen on every turn for the rest of the session. The result was a model that looked forgetful and repetitive while using more of a customer's limit than expected. That fix shipped April 10. Then came April 16. Anthropic added a system prompt instruction telling Claude to keep text between tool calls to 25 words or fewer, with final responses capped at 100 words unless more detail was required. It sounds tiny. It wasn't. Anthropic said the instruction hurt coding quality and removed it on April 20. All three issues were marked resolved in Claude Code version 2.1.116, and Anthropic said the API was not affected. The postmortem also said Claude Agent SDK and Claude Cowork were touched by the broader product-layer issues, so this wasn't quite as narrow as frustrated Claude Code users might have assumed. Three separate changes, across six weeks, each weakening the same work surface in a different way. That's a rough stretch for any engineering team. But bugs are survivable. The slow admission is harder to explain away. The Bill Kept Moving While Anthropic was trying to recover from the quality issue, it was also changing how subscribers could use Claude around the edges. Fortune reported in April that Anthropic tested showing Claude Code as unavailable on the $20 Pro plan for about 2% of new sign-ups, though the current Claude pricing page says Claude Code is included in all paid plans. That distinction matters. A test is not a full removal, and writing it as one would overstate the case. The firmer change came through third-party tools. TechCrunch reported that, starting April 4, Anthropic told subscribers they could no longer use subscription limits for third-party harnesses including OpenClaw. Boris Cherny, Anthropic's head of Claude Code, wrote on X that subscriptions were not built for those third-party usage patterns. Frankly, that is a fair business problem. It is also exactly the sort of change users notice when they built workflows around the old bargain. Anthropic tried another billing adjustment in June. Its Help Center said Claude Agent SDK usage, the claude -p command, Claude Code GitHub Actions, and third-party apps would move into a separate monthly Agent SDK credit pool starting June 15. Then it paused the plan that same day, saying nothing had changed for now and that the monthly credit was not available. It backed down. For now. The company has also tried to put capacity back into users' hands. On May 6, Anthropic said it was doubling Claude Code's five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans after new compute deals, including one with SpaceX. A separate Claude Code weekly-limit promotion has kept eligible users 50% higher through August 19, according to Anthropic-linked support coverage. Those moves cost Anthropic real capacity. They also show the pressure was landing. A Trust Problem Here's the part that should worry Anthropic more than any single bug report. The company markets itself as the careful AI lab, the one enterprise buyers can trust with sensitive work. That pitch depends on users believing what Anthropic says about its own product. Weeks of denial followed by a technical apology leaves a mark. You can forgive a bad deployment. You can even forgive a pricing change if it is direct, dated, and explained before it hits your workflow. Stack quality regressions, usage-limit confusion, third-party tool restrictions, and paused billing plans together, and you get something worse than a bug. You get uncertainty. For developers, uncertainty is expensive. They don't just want Claude Code to be clever on a demo stage. They want to know whether the same command line tool will behave next week, whether the same subscription still covers the same work, and whether complaints will be treated as signal before the cancellation posts pile up. Anthropic fixed the bugs by April 20. The trust repair is still open. Also read: Amazon Shuts Its AGI Lab and Cuts Jobs to Chase Enterprise AI Instead https://startupfortune.com/amazon-shuts-its-agi-lab-and-cuts-jobs-to-chase-enterprise-ai-instead/ • NXP is chasing Ambarella for $3.3 billion because it needs eyes, not just brains https://startupfortune.com/nxp-is-chasing-ambarella-for-33-billion-because-it-needs-eyes-not-just-brains/ • California's AI Transparency Act Takes Effect With Fines That Compound Daily https://startupfortune.com/californias-ai-transparency-act-takes-effect-with-fines-that-compound-daily/