One Max plan subscriber noticed something odd after their $100/month Claude subscription lapsed: the model got faster.
The user had been on Claude's Max 5x plan - Anthropic's top consumer tier, which promises five times the usage of the standard Pro plan. Despite paying the premium rate, they consistently ran into queuing delays during heavy Claude Code sessions, sometimes burning through their entire 5-hour usage window in under an hour. Agent steps would stall. Complex multi-step tasks would queue between each action.
Then the subscription expired. For roughly 25 minutes, they ran Opus 4.6 at full speed - no queuing, no delays, no rate limiting.
The likely explanation: Anthropic's throttling isn't purely about usage caps. It's about compute allocation. When you're on the Max plan and running Claude Code hard - the kind of session where an AI agent might take dozens of sequential steps to complete a task - you're hitting shared priority queues alongside every other Max subscriber doing the same thing. When the subscription dropped to a lower tier, the user briefly landed in a different queue that, at that moment, had more available headroom.
This behavior isn't unique to Claude. Running frontier models like Opus 4.6 is genuinely expensive on the infrastructure side, and most AI services throttle heavy users even on paid plans. Anthropic hasn't published specifics on how compute is distributed across subscription tiers.
What the experience points to: if you're using Claude Code for long agentic tasks, you may be hitting infrastructure limits well before you reach your usage cap. The $100/month buys more messages - it doesn't guarantee faster ones during peak load. Anyone who uses Claude Code for complex, multi-step work should factor in real-world latency, not just the advertised usage limits.
Anthropic has not commented on the specific behavior described.