Claude's Pro and Max subscription tiers and the pay-as-you-go API each handle Opus rate limiting differently. The four windows — five-hour rolling, seven-day rolling, tokens-per-minute, and requests-per-minute — get reshuffled with each tier. Here is how each plan optimizes its limits, where the felt pain lives, and what pooling accounts looks like across tiers.
Four limit windows you're sharing across one product family
Anthropic's published rate-limit model is not a single number. It is four windows operating in parallel:
| Window | Measures | Resets | Typical bottleneck |
|---|---|---|---|
| 5-hour rolling | Aggregate usage in trailing five hours | Continuously, as old usage rolls off | Burst sessions, long days |
| 7-day rolling | Aggregate usage in trailing seven days | Continuously, as old usage rolls off | Heavy weeks, multi-day workloads |
| TPM (tokens per minute) | Throughput within the current minute | Every minute | Large completions, heavy streaming |
| RPM (requests per minute) | Request count within the current minute | Every minute | Sub-agent fan-outs, parallel tool calls |
The mental model most users carry is that rate limits are "one number" — "I get five hours of Claude per day." The four-window reality is more nuanced. A heavy session with small completions can sail through TPM but trip RPM. A session with one enormous completion can do the opposite. A workload that streams steady traffic for hours can run out the 5-hour window without ever tripping TPM or RPM at any individual moment.
Senior developers using Claude Code as a daily driver tend to hit all four eventually. Knowing which one is binding at any given moment is the first step toward not hitting it. The routing layer that reads those windows and picks the account with the most headroom is free to download and try, no credit card needed.
The Pro plan: a single user's daily envelope
Anthropic's Pro plan is sized for a single individual using Claude across a workday. The qualitative properties:
- 5-hour window is the dominant felt constraint. Most Pro users hit it once or twice a day. Recovery is automatic and fast — the window rolls off as wall-clock time passes.
- 7-day window rarely binds for normal use. Casual to moderate coding leaves headroom. Heavy autonomous workloads will saturate it within a few days, at which point the weekly ceiling becomes the operative limit.
- TPM and RPM are generous enough that solo work rarely trips them. A single conversation streaming completions at human-reading speed does not approach the per-minute ceilings. The exception is sub-agent fan-out, which concentrates request rate against the per-account limit.
- One model selection per turn. Pro permits Opus, Sonnet, and Haiku usage. The 5-hour budget is shared across models — using Sonnet does not preserve Opus headroom.
The Reddit pattern that matches Pro most often is the "one or two days" complaint:
burn through the whole damn quota in ONE OR TWO DAYS
That is the 7-day window kicking in on Pro after a heavy Monday and Tuesday. The 5-hour window keeps rolling off, but the weekly cap has accumulated enough usage that Wednesday's session opens with most of the day's headroom already spoken for.
For current numerical ceilings on Pro, the Anthropic rate-limit documentation is authoritative.
The Max plan: more 5-hour headroom, same architectural ceiling
Max raises the published ceiling on each of the four windows substantially. The structural model is identical — rolling windows, per-account, no calendar alignment. The change is the size of the numbers.
What Max optimizes:
- Larger 5-hour bucket. Heavy coding sessions can run two to three times longer before hitting the throttle, depending on workload.
- Larger 7-day window. Sustained heavy use across a full work week is more feasible. The "unusable rest of the week" pattern still happens under autonomous loops, just at a later point in the week.
- Higher TPM and RPM caps. Fan-out workloads that trip Pro's per-minute ceilings have more headroom on Max. The ceilings still exist — they just kick in later.
What Max does not change:
- The window shape (rolling, not block-based).
- The per-account nature of every limit.
- The interaction between 5-hour and 7-day windows.
- The "this account has no idea about your other account" property.
For users whose workload structurally exceeds Pro's ceilings but fits inside Max's, the upgrade is the right move. For users whose workload structurally exceeds even Max's ceilings — autonomous overnight loops, multi-agent fan-outs across full projects — Max just delays the wall by some hours. The structural ceiling is still per account.
Anthropic's plan documentation has the current Max numbers. They have changed before and may change again.
The API tier: pay-as-you-go, different bottleneck
The Anthropic API is structurally different from the Pro and Max subscription tiers. The relevant changes:
- No 5-hour or 7-day rolling windows in the subscription shape. API usage is billed per token, with no aggregate session budgets.
- TPM and RPM still exist as concurrency caps. API users still have per-minute ceilings, often higher than subscription tiers, sized to expected developer workloads.
- Spend is the binding constraint, not throttling. Where subscription users feel the cap as a wall ("you cannot use Claude until 11:47am"), API users feel the cap as a bill ("your previous month's API spend was four hundred dollars").
The pain shape shifts. Pain point number five in every Claude complaint thread is the surprise invoice:
- A widely-cited write-up captured a $47,000 month from a misconfigured production deployment.
- A dev.to post described an $80 overnight surprise from a single iterative script.
- The recurring observation that "90% of all tokens in Claude Code sessions are cache reads" — cheap per token, expensive in aggregate, invisible to most users until the bill arrives.
API users do not stare at countdowns. They reconcile invoices. The bottleneck is cost transparency, not capacity.
For mixed workflows — a subscription for daily interactive coding plus API fallback for autonomous loops — the limits stack independently. The subscription window does not deplete API headroom. The API spend does not consume subscription budget. They live in separate accounting domains.
Pooling for parallelism — what each tier looks like pooled
Pooling accounts within a tier is the structural fix. Pooling across tiers is a second-order option that some workflows benefit from.
Two Pro accounts pooled. The user owns two Pro accounts (their own, no sharing). A router routes each request to whichever account has more remaining 5-hour and 7-day budget. The user-visible ceiling roughly doubles. The 5-hour throttle still exists per account; the rotator just keeps both accounts cool enough that no single one is binding.
One Max + one Pro pooled. A heterogeneous pool. The router prefers Max for heavy workloads (more headroom per request, less risk of throttling the active account) and falls back to Pro when Max is cooling. The Pro account absorbs the overflow, extending the user's working day past where Max alone would have hit the wall.
Max + API fallback. The subscription accounts handle interactive workloads. The API key handles autonomous overnight loops. The router routes by intent: interactive prompts to subscription accounts (no per-token cost), batch loops to API (no per-window throttling, accepted spend). The result is that the user controls both their wall-clock availability and their monthly spend cap explicitly.
The structural property worth holding onto: pooling raises the effective ceiling across every account you own without raising any individual account's published ceiling. It also does not violate Anthropic's terms — each account in the pool is independently owned, billed, and held to its own contract. The pool is a routing layer on the user's side, not a multiplexing arrangement on Anthropic's side. If you want to see the math on your own mix of tiers, the 14-day Premium Trial is $0 today and you can cancel anytime before day 15.
Where the Anthropic docs live
The numerical ceilings for each tier change from time to time. Three references are worth bookmarking:
- The Anthropic Pro and Max plan page — current per-tier features and limits, the authoritative source for what each tier currently includes.
- The Anthropic API rate limit documentation — the per-tier TPM and RPM ceilings, plus the header reference for
X-RateLimit-*response metadata. - The Claude Code documentation — Claude Code's interaction with the underlying rate limits, which can change with the editor extension's release cadence.
The three together are the canonical source for what your account can do today. Anything cited in this post is structural, not numerical — the published numbers are not reproduced here because they move.
Frequently asked questions
Why does my Pro account hit the limit faster than the published "X messages per 5 hours" number suggests?
Because the published number describes a quiet conversational workload, and senior Claude Code workflows are not quiet. Each turn carries the full system prompt, tool definitions, project context, and conversation history. A "message" in a real Claude Code session is several to many times the size of a "message" in a casual chatbot conversation. The 5-hour budget is the same. The per-turn weight against it is larger.
Does Max give me a higher TPM cap, or just a higher 5-hour cap?
Both. Max raises the ceilings on all four windows. The 5-hour and 7-day buckets are larger. The TPM and RPM caps are higher. The structural model is identical to Pro — rolling, per-account — just sized for a more intensive workload. Current numbers are on the Anthropic plan page.
Can I mix a Pro and a Max account in the same pool?
Yes. A pool of mixed-tier accounts works the same as a pool of same-tier accounts at the routing level. The router treats each account as a budget-and-state pair and routes by projected remaining capacity. Mixed pools are useful when you want Max's higher headroom for normal traffic plus a Pro account as a low-cost overflow buffer.
If I pool an API key, do API requests count against my subscription's 5-hour window?
No. API and subscription accounts are independently accounted. A request routed to an API key is billed per-token at API rates and does not draw down any subscription 5-hour or 7-day budget. Pooling subscription + API is one way to get both wall-clock availability (subscription) and overflow capacity (API) without each interfering with the other.
Do prompt-cache reads count against any of the four limits?
They count toward TPM accounting under most current pricing. The token cost of a cached read is lower than a fresh read for billing purposes, but the rate-limiter applies the read at full count for throughput accounting. Caching saves money over a session. It does not let you fan out further before TPM trips.
Will Anthropic merge my multiple Claude.ai accounts if they share a billing card?
Each account is billed independently regardless of which payment method is on file. Anthropic's terms permit each individual to hold multiple accounts so long as each one is owned and controlled by that person. Multiple accounts on the same card are common and not a TOS issue, but always check the current terms rather than relying on a prior reading — these policies do change.
Closing
Pro, Max, and API each optimize for a different workflow, but all three share the four-window structure underneath. Knowing which window is binding on your current account is the first step toward fixing it; pooling is the second. The rotation extension we ship handles the routing layer across whatever mix of tiers you bring to it.
To pool your own accounts, the Premium Trial runs 14 days at $0 today — cancel anytime before day 15 — and adds Balanced Mode, the watchdog, and usage analytics on top of rotation. Not ready to enter a card? Install Power Claude Free instead: the 7-day free trial gives you full Pro access, no credit card.