Opus 4.6 regression since ~Sep 16: hallucinating dates and degraded accuracy
Description
Since approximately September 16, 2026, Claude Opus 4.6 has shown a noticeable regression in accuracy and reliability. Before that date, the model was working perfectly for complex agentic workflows. Specific issues observed after the update:
1. Date hallucination
When composing a formal email in September 2026, Claude wrote "20 de septiembre de 2025" instead of 2026. The current date was available in system context, yet Claude defaulted to its training cutoff year (May 2025).
2. Base64 reproduction corruption
When sending email attachments via Gmail MCP, Claude corrupts base64 strings in output tokens. Strings as short as ~1200 chars fail with "Base64 decoding failed". Only ~800 char strings reproduce correctly, and inconsistently.
Timeline
- Before Sep 16, 2026: Opus 4.6 worked flawlessly — accurate dates, reliable tool use, no hallucinations in agentic workflows.
- After Sep 16, 2026: Consistent date hallucinations, degraded output accuracy, base64 corruption in tool calls.
Environment
- Model:
claude-opus-4-6 - Session: Claude Code cloud container
- Tools: Gmail MCP, Google Drive MCP
Expected behavior
- Use current date from system context, not training cutoff year
- Reproduce base64 strings accurately in tool call parameters
Impact
Severe degradation for agentic workflows relying on accurate data and tool interoperability. A user who relied on Opus 4.6 daily for complex multi-step tasks (formal complaints, GitHub issues, email composition in Spanish) now encounters frequent errors that did not exist before