Opus 4.6 inference quality regression correlates with March 2026 doubled-limits promotion

Note: This issue was researched, drafted, and submitted by Claude Code (Opus 4.6) at the request of the user, who observed the degradation firsthand and asked Claude to collect the evidence and file the report.

Description

Claude Code (Opus 4.6) shows judgment degradation on trivial tasks
during the March 13–27 off-peak doubled-limits promotion. These same
tasks — skill invocation, card updates, scoped edits — are handled
correctly under normal traffic conditions. The degradation correlates
specifically with the doubled-traffic period.

Environment

• Claude Code CLI (latest)
• Model: claude-opus-4-6 (1M context)
• Platform: macOS Darwin 24.6.0
• Session: fresh (not context-exhausted)
• Time: off-peak hours (promotion active)

Observed Behavior

A trivial task (update one Serena memory file per a Linear card comment)
in a fresh session with all rules loaded:

1. Skipped required skill invocation for a routine git operation —

a pattern this user's project has documented feedback memories for,
which the model follows correctly under normal conditions

1. Added an unrequested flag (--init) to a command — unprompted

scope addition on a simple edit

1. Omitted a card comment after completing comment-driven work —

a step the model reliably performs under normal load

These are not complex multi-step reasoning failures. This is a trivial
card with one file edit. The model follows these exact workflows
correctly during normal-traffic sessions.

Expected Behavior

Trivial tasks with rules loaded in a fresh session should execute at
the same fidelity as under normal traffic. The rules are in context,
the task is simple, and the model has demonstrated it can follow them.

Suspected Cause

The doubled off-peak traffic from the March 13–27 promotion appears to
be degrading inference quality. Evidence:

• Same user, same project, same rules — correct behavior under normal

load, degraded behavior during promotion period

• Local tooling (file I/O, MCP servers, CLI) functions perfectly —

failures are exclusively server-side judgment

• status.claude.com shows elevated errors on Opus 4.6 on March 25
• Multiple community reports of quality regression during this period
• Reports of a config change to Opus 4.6 around Feb 10-11 causing

accuracy drops, possibly compounded by increased traffic

Question

Is the doubled off-peak traffic impacting Opus 4.6 inference quality?
The failure pattern — tools work, judgment doesn't, on tasks that
normally succeed — points to server-side degradation, not model
capability.