[Regression since v2.1.89] Max subscription severely degraded — external pay-per-token providers unaffected

Environment

• Claude Code: v2.1.101 (regression first noticed: v2.1.89)
• OS: macOS
• Subscription: Max

Description

Since approximately v2.1.89, Claude Code has become significantly slower when using native Max subscription tokens. The same tasks complete 2-3x slower than before:

• Previously: ~5 agentic tasks (file edits, multi-step operations) in ~5 min
• Currently: same 5 tasks take 10-15 min

Two symptoms:

1. Higher first-token latency — noticeably longer wait before any output starts
2. Slower terminal scroll rendering during generation

Key observation

Switching from native Max tokens to an external pay-per-token API provider restores original performance immediately. External providers are unaffected by this regression.

The degradation is specific to the Max subscription code path.

Impact

Agentic workflows depend on fast iteration cycles. At 2-3x degraded speed, the Max subscription has become impractical for professional use. The fact that external pay-per-token providers are unaffected raises a serious question about whether this is an intentional throttling decision targeting flat-rate subscribers.

Combined with undocumented reductions to weekly token limits around the same timeframe, there is a clear pattern of silent service degradation for Max subscribers — with no changelog, no email, no prior notice.

Request

1. Identify and revert the performance regression introduced ~v2.1.89
2. If rate limits or throttling policies have changed for Max subscribers, communicate this explicitly and in advance