Anthropic tightened Claude’s session-based usage limits during peak hours this week, the first time the company has publicly acknowledged that agent-driven compute demand is straining its infrastructure.
The change, announced on X by Anthropic’s Thariq Shihipar on March 26, adjusts how quickly Free, Pro, and Max subscribers burn through their five-hour session limits during weekday peak hours (5:00 AM to 11:00 AM PT / 1:00 PM to 7:00 PM GMT). Weekly limits remain unchanged, but users will exhaust session allowances faster during those windows.
“We’ve landed a lot of efficiency wins to offset this, but ~7% of users will hit session limits they wouldn’t have before, particularly for pro tiers,” Shihipar wrote. He recommended that users running “token-intensive background jobs” shift them to off-peak hours to stretch their session limits further.
The OpenClaw Factor
Business Insider reported that Anthropic is experiencing capacity constraints driven in part by OpenClaw, the open-source AI agent framework. The outlet noted that tools like OpenClaw allow “users to tap into the full potential of AI models like never before,” but the resulting explosion of automated, multi-step agent loops means users are consuming compute at rates that traditional chat-based usage never approached.
The pattern is distinct from normal chatbot usage. Agent frameworks run Claude in autonomous loops where a single user session can trigger dozens of API calls with no human in the loop between them. A developer running an OpenClaw agent to debug a codebase, for example, might generate more token throughput in 30 minutes than a manual user produces in a full day.
How the New Limits Work
The Register’s coverage detailed the mechanics. Anthropic does not publish exact token counts for its subscription tiers. Instead, it uses an opaque “session limit” system where usage is calculated based on “the length and complexity of your conversations, the features you use, and which Claude model you’re chatting with,” per Anthropic’s documentation.
Under the new regime, session limits during peak hours drain faster than the clock. A five-hour session window during peak time may exhaust in under five hours of actual use. Anthropic compensated by expanding capacity during off-peak hours, so the total weekly allocation remains flat.
Subscription tiers affected: Free, Pro ($20/month), Max 5x ($100/month), and Max 20x ($200/month). API customers paying per-token are unaffected by the session limit changes.
Broader Compute Pressure
Anthropic is not the only lab feeling the squeeze. OpenAI announced earlier this week that it is discontinuing Sora, its AI video generation app, as it reallocates compute toward core services. The underlying dynamic is the same: frontier model providers are discovering that the agent era burns through GPU cycles at a rate their infrastructure wasn’t built for.
The timing adds context to Anthropic CEO Dario Amodei’s comments at the New York Times Dealbook Summit in December, where he questioned competitors’ “ambitious spending plans for mega data centers.” The implication at the time was that Anthropic planned to be more capital-efficient. The peak-hour throttling suggests that efficiency has limits when your user base starts running autonomous agents around the clock.
For OpenClaw users specifically, the practical impact depends on their tier and usage patterns. Pro subscribers who run agent loops during US business hours are most likely to hit the new constraints. The workaround is straightforward but inconvenient: schedule heavy agent workloads for off-peak windows, or move to API-based access where billing is per-token rather than session-based.