QWE AI Academy You’re paying Anthropic $20, $100, or $200 per month for Claude. As of March 26, 2026, your usage burns faster between 5-11am Pacific on weekdays – not because you changed anything, but because Anthropic decided peak hours cost more tokens per minute.
Free users? Same restrictions.
Nobody at Anthropic seems willing to answer this: Why throttle everyone during peak hours instead of restricting free accounts to off-peak only? If capacity’s the problem, the people paying nothing should bear the constraint.
What Changed
Starting March 26, Anthropic adjusted how Claude’s five-hour session limits work during weekday mornings. Peak hours: weekdays 5-11am PT (1-7pm GMT). Your session budget depletes faster than the clock. Five-hour window? Might last you three actual hours. Ninety minutes if you’re running Claude Code.
Affects Free, Pro ($20/month), Max 5x ($100/month), Max 20x ($200/month). All equally. About 7% of users will hit limits they wouldn’t have before – Pro tier subscribers get hit hardest (according to Anthropic engineer Thariq Shihipar’s X post).
Weekly limits stay the same. Morning work just costs more of your quota. Off-peak got a temporary 2x boost through March 28 – after that? Straight throttling, no upside.
API customers? Exempt. They pay per token, no session limits.
Opaque Pricing Meets Moving Targets
Anthropic doesn’t publish token counts for session limits. Your usage depends on “conversation length and complexity, features used, which Claude model you’re chatting with.” You track a percentage on the dashboard. Can’t predict when you’ll hit the wall until lockout.
Max 20x users ($200/month) report usage jumping 21% to 100% on a single prompt during peak. Or exhausting quota in 90 minutes. ChatGPT Plus: $20, no time-of-day throttling.
Think about that pricing model for a second. You’re paying 10x more than ChatGPT Plus users, but Anthropic treats your morning sessions the same as someone paying zero. The subscription tiers exist to segment the market, but the throttling ignores them completely. It’s like buying first-class and getting economy legroom during rush hour.
Alternative: Restrict Free Users to Off-Peak
What if free accounts could only access Claude outside peak – say, after 11am PT weekdays, plus all weekend?
Free users see “Claude available after 11am PT” during restricted hours. Paid subscribers get 24/7 access, no peak throttling. API customers continue as-is.
People paying $100-200/month wouldn’t subsidize free users’ morning compute. Anthropic reserves peak capacity for revenue-generating subscribers. Free users still get generous access – just not during the six hours of highest demand.
But: Anthropic risks slower free-to-paid conversion. They’re already throttling free users during peak anyway. Current approach just spreads the pain without rewarding paying customers.
Current Reality: Throttle Everyone
Every tier burns quotas faster during weekday mornings. “Weekly limits stay the same, just redistributed.”
Three problems:
1. The 2x off-peak promo ends March 28. Right now off-peak usage is doubled (through the promotion), offsetting peak throttling. After Friday? Off-peak returns to normal. “Weekly limits unchanged” becomes a de facto cut once users lose the 2x buffer.
2. Paid users can’t differentiate. Max 20x hitting limits in 90 minutes gets the same experience as free. What’s the value prop?
3. Opacity breeds frustration. You don’t know session cost until you’ve burned it. API pricing shows exactly what you’re spending per request. Web users don’t.
The Notification Strategy
Peak-hour change: announced via an engineer’s personal X account. The 2x off-peak promo? Official Claude account post. Bad news (throttling paying users) vs. gift (temporary bonus). The messaging mismatch tells you everything.
Workaround 1: Schedule Heavy Work After 11am PT
Peak hours: 5-11am Pacific (8am-2pm Eastern, 1-7pm GMT). Running Claude Code, deep research, large files? Push to late morning or afternoon.
Your local peak window:
- Pacific: 5-11am
- Eastern: 8am-2pm
- GMT/UK: 1-7pm
- Central Europe (CET): 2-8pm
- Asia (Tokyo JST): 10pm-4am next day
In Asia? Your daytime is Anthropic’s off-peak.
Workaround 2: Use the API for Morning Sessions
API customers aren’t subject to session throttling. Morning work predictable and token-intensive? Claude API might cost less, avoid surprises.
Use API if:
- You’re on Pro, regularly hit limits during peak
- Use case is scriptable (batch document processing, code review loops)
- You can estimate tokens (use API usage tracking to baseline)
Skip API if: You need web UI features (Artifacts, Projects, conversational back-and-forth). Or usage is sporadic – subscription’s more predictable than pay-per-token.
Workaround 3: Track Dashboard Like Fuel Gauge
Claude dashboard shows progress toward five-hour daily limit, weekly cap. Watch it during peak. 40% after one task? You’ve got maybe two more before lockout.
Manual, annoying. But it’s the only way to predict behavior when Anthropic won’t publish token counts. Percentage moving fast? Wrap up, wait for off-peak.
Workaround 4: Split Work Across Free + Paid
Anthropic hasn’t banned multi-account for personal use. Free account for lightweight queries during peak, save Pro/Max quota for complex tasks off-peak. Not elegant, but if you’re on tight budget and need day coverage, it works.
Caveat: Don’t create dozens to dodge limits. That’s abuse, gets flagged. One free + one paid for personal workflow? Fine.
Why This Matters Beyond Claude
Anthropic isn’t alone. AI usage rationing is becoming standard as compute costs collide with flat-rate subscriptions. Google Gemini publishes hard daily prompt limits (5 free, 100 Pro, 500 Ultra). ChatGPT throttles to GPT-3.5 when you exhaust GPT-4. Claude’s just more opaque.
Bigger question: restrict by tier (free vs. paid) or time (peak vs. off-peak)? Anthropic chose time. Punishes paying users to protect capacity for free ones during same window. It’s a choice, not an obvious one.
Here’s the thing most AI companies won’t say out loud: flat-rate subscriptions were never sustainable for compute-intensive workloads. Traditional SaaS (Dropbox, Notion, Slack) has predictable costs per user – storage, database queries, API calls. AI inference? Cost varies 100x depending on what you ask. A simple chat: pennies. Multi-turn coding session with file context: dollars. Anthropic’s trying to fit a variable-cost product into a fixed-price model, and peak-hour throttling is the hack that makes the math work.
After March 28
2x off-peak promotion ends Friday. Off-peak returns to standard. Been running heavy sessions outside peak all month? You’re about to feel the real cost. Weekly limits technically “unchanged,” but losing 2x buffer will feel like a cut.
Anthropic says they’re “investing in scaling efficiently” (no timeline for lifting restrictions). Translation: new normal until they add more GPUs.
FAQ
Does upgrading to Max fix the peak-hour throttling?
No. Max 20x users report hitting limits in 90 minutes during peak despite $200/month.
If I switch to the Claude API, do I avoid session limits entirely?
Yes – API customers pay per token, no session limits. But you lose Artifacts, Projects, conversational UI. Say you’re batch-processing 50 documents for compliance checks. API: perfect. You script it, get exact token counts, done. But debugging a complex codebase where you’re going back and forth, testing ideas, refining prompts? Web interface wins. The API doesn’t remember context well across separate calls unless you manually manage conversation history.
Can I see how many tokens a session used after it’s done?
Not if you’re a subscription user. You see percentage (“47% of weekly limit”), but Anthropic hides the math. API users get exact counts in response headers. This opacity is deliberate – if users knew the token math, they’d optimize around it, which defeats the point of session limits as a capacity management tool. But it creates a trust problem: you’re flying blind on what actually costs quota.
Next step: Open Claude dashboard now. Note current usage percentage. Run one typical task (coding session, document summary, whatever you normally do). Check again. That’s your new cost baseline for peak vs. off-peak. Adjust schedule – or switch to API if subscriptions no longer make sense for your workflow.