Gate News message, April 23 — Anthropic’s engineering team confirmed that the Claude Code quality degradation reported by users over the past month stemmed from three independent product-layer changes, not from API or underlying model issues. The three problems were fixed on April 7, April 10, and April 20 respectively, with the final version now at v2.1.116.
The first change occurred on March 4, when the team reduced the default reasoning effort level for Claude Code from “high” to “medium” to address occasional extreme latency spikes in Opus 4.6 under high reasoning intensity. After widespread user complaints about reduced performance, the team reverted the change on April 7. The current default is now “xhigh” for Opus 4.7 and “high” for other models.
The second issue was a bug introduced on March 26. The system was designed to clear old reasoning records after conversation inactivity exceeded one hour to reduce session recovery costs. However, a flaw in implementation caused the clearing to execute repeatedly on every subsequent turn rather than once, causing the model to progressively lose prior reasoning context. This manifested as increasing forgetfulness, repeated operations, and abnormal tool invocations. The bug also resulted in cache misses on every request, accelerating user quota consumption. Two unrelated internal experiments masked the reproduction conditions, extending the debugging process to over a week. After fixing on April 10, the team reviewed problematic code using Opus 4.7 and found that Opus 4.7 could identify the bug while Opus 4.6 could not.
The third change launched on April 16 alongside Opus 4.7. The team added instructions to the system prompt to reduce redundant output. Internal testing over several weeks showed no regression, but post-launch interaction with other prompts degraded coding quality. Extended evaluation revealed a 3% performance drop in both Opus 4.6 and 4.7, leading to a rollback on April 20.
These three changes affected different user groups at different times, and their combined effect created widespread and inconsistent quality decline, complicating diagnosis. Anthropic stated it will now require more internal employees to use the same public build version as users, run full model evaluation suites for every system prompt modification, and implement staged rollout periods. As compensation, Anthropic has reset usage quotas for all subscription users.
Artikel Terkait
Meta Platforms Rencana Pengurangan Tenaga Kerja 10% pada 20 Mei, Berdampak pada Kira-kira 8.000 Posisi
Pemerintahan Trump mengumumkan rencana penindakan terhadap pemurnian AI, menuduh perusahaan Tiongkok melakukan pencurian sistematis kemampuan model
DeepSeek meluncurkan V4 versi pratinjau sumber terbuka, penilaian teknis 3206 melampaui GPT-5.4
Cambricon Menyelesaikan Adaptasi Day 0 DeepSeek-V4, Menandai Tonggak Penting untuk Ekosistem Chip AI Tiongkok
Tencent merilis Hy3 versi pratinjau sumber terbuka, tolok ukur pengujian kode meningkat 40% dibanding pendahulunya
Investasi Portofolio FTX Senilai 158 Triliun Won Jika Tidak Bangkrut