Guide

Claude Code Usage Limits: How They Work and How to Work Within Them

The quota numbers change constantly; the mechanics don't. Here is how the limit system actually works, one pool, rolling windows, model-dependent burn, and the strategies that stretch a plan.

By Kylian Migot · Updated July 2026 · 8 min read

Quick answer

Claude Code subscription limits are one shared pool per account, shared with claude.ai chat, metered on rolling windows plus longer-period caps. Heavier models and parallel sessions burn the pool faster; API-key usage bills per token instead. The biggest lever is model routing: Sonnet or Haiku for routine work, Opus or Fable where it pays.
The pool
One per account, shared between Claude Code and claude.ai chat
The meter
Rolling windows plus longer-period caps; /status shows your position
Burn rate
Scales with model tier; Fable 5 fastest, Haiku 4.5 slowest
Parallel sessions
All draw from the same pool, 4 agents is roughly 4x burn
Exact quotas
Change often, check Anthropic's plan docs for current numbers
01

The Stable Mechanics

Most pages about Claude Code limits lead with quota figures, and most of those figures are wrong within weeks, because Anthropic tunes them continuously. The underlying architecture, though, has been stable, and understanding it is what actually lets you plan work. Four mechanics do almost all of the explaining.

One pool per account

Subscription plans (Pro and the Max tiers) grant a single usage pool that Claude Code and claude.ai chat both draw from. There is no separate coding budget: a long chat session and a long agent session spend the same currency.

Rolling windows plus longer caps

Usage is metered over a short rolling window that continuously recovers, with longer-period caps layered on top for sustained heavy use. Hitting a limit pauses new work until the relevant window resets; nothing about your session is lost.

Burn scales with the model

The pool is consumed in proportion to the compute you use, so heavier models drain it faster per message. Model choice is not a taste question on a subscription, it is the throttle on how long your plan lasts.

Subscription pool vs API metering

An API key bypasses the pool entirely: usage bills per token at published rates instead. Same CLI, different economics, a flat monthly cost with a ceiling versus an uncapped cost with no ceiling.

Two practical corollaries. First, every parallel session on the account draws from the same pool, so running four agents at once burns it roughly four times as fast, parallelism buys wall-clock time, never capacity. Second, you can always see where you stand: run /status inside Claude Code and it reports your current usage position in-session, no dashboard hunting required.

Ship your first agent today

Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.

Download AIDEN free

Free to start · macOS 12+ · No credit card required

02

Model Burn Rates: The Price Ratio Is the Proxy

Anthropic doesn't publish a per-model "pool cost" table for subscriptions, but the API token prices are the honest public proxy for relative burn, they reflect what each model costs to run. The ratios below are what matter: on a subscription, a Fable 5 session spends your pool in the neighborhood of ten times as fast as the same tokens through Haiku.

ModelAPI price (burn proxy)Relative burnRoute here
Claude Haiku 4.5$1 in / $5 out per MTok~1x (baseline)Mechanical edits, triage, sub-agents
Claude Sonnet 5$3 in / $15 out per MTok~3xRoutine features, tests, refactors
Claude Opus 4.8$5 in / $25 out per MTok~5xComplex multi-file agentic work
Claude Fable 5$10 in / $50 out per MTok~10xThe hardest refactors and long autonomous runs

Prices verified against vendor docs on July 18, 2026. The takeaway is not "never use Fable", it is that defaulting to a heavy model for routine work is the fastest way to hit a wall mid-week. Which tier earns its burn for which task shape is covered in the best Claude model for coding.

03

Four Strategies for Working Within the Limits

Once you accept the pool as a budget, stretching it becomes an engineering problem with four good levers, roughly in order of impact.

1. Route models by task, not by habit

Sonnet or Haiku for routine implementation, tests, and mechanical edits; Opus or Fable only where the task is genuinely hard enough to pay their burn. This one habit can multiply how far a plan goes. The decision rules are in best Claude model for coding.

2. Budget parallel agents deliberately

Four parallel sessions are roughly 4x burn, so parallelize on the cheap tiers and keep the heavy model for the one story that needs it. AIDEN makes this a per-story routing decision on the board rather than a per-terminal default. How to run parallel sessions sanely is its own guide: managing multiple Claude Code sessions.

3. Treat a second ecosystem as a second pool

A Codex plan is a genuinely independent budget, its limits don't touch Anthropic's. Engineers who run both route work across two pools and hit walls far less often. See Claude Code and Codex together for the routing, and Codex usage limits for how that pool behaves.

4. Know when API billing beats the subscription

Per-token API billing wins in two situations: bursty, occasional use that never justifies a monthly fee, and sustained heavy use that keeps slamming into plan caps where uncapped pay-per-token, budgeted deliberately, is cheaper than the frustration. Steady daily use in between is where subscriptions shine.

FAQ

Do parallel Claude Code sessions share usage limits?
Yes. Every session signed into the same account draws from one shared usage pool, so four parallel agents burn that pool roughly four times as fast as one. Parallelism buys wall-clock time, not extra capacity. If you run agents in parallel routinely, budgeting which model each session uses matters more than how many sessions you open.
Does Claude Code share limits with claude.ai chat?
Yes. On subscription plans, Claude Code and claude.ai chat on the same account draw from the same usage pool. A heavy morning of chat conversations leaves less room for afternoon coding sessions, and vice versa. They are one budget, not two.
Which Claude model burns through limits fastest?
Heavier models consume the pool faster, and API token prices are the best public proxy for the ratio. Fable 5 ($10/$50 per MTok) is the fastest burn in the current lineup, roughly double Opus 4.8 ($5/$25), which in turn runs well above Sonnet 5 ($3/$15) and Haiku 4.5 ($1/$5). Routing routine work to Sonnet or Haiku is the single biggest lever for stretching a plan.
What happens when I hit the limit?
Claude Code tells you and stops accepting new work on the subscription pool until the relevant window resets; rolling windows recover on their own schedule, and longer-period caps reset less frequently. Your session state and files are untouched, you are rate-limited, not logged out. You can wait for the reset, switch to a lighter model earlier next time, or fall back to API-key billing, which is pay-per-token and not subject to the plan pool.
Are Pro and Max limits the same?
No. Higher tiers come with substantially larger usage pools, that is most of what you are paying for. The exact multiples and quotas change often enough that any number printed here would go stale, so check Anthropic's official plan documentation for current figures, and use /status inside Claude Code to see where your own account stands right now.

Keep reading

Codex usage limits & credits

How Codex plans and credits work, and how to budget agent work across them.

Best Claude model for coding

Fable 5 vs Opus 4.8 vs Sonnet 5 vs Haiku, which to use for which task, verified data.

Claude Code and Codex together

The dual-agent workflow: which tool takes which task, and how to run both on one project.

Manage multiple Claude Code sessions

From terminal tabs to worktrees to a board: scaling past one agent without losing track.

Opus vs Sonnet vs Haiku

The Claude tier picker: capability, cost per story, and when each tier earns its price.

Route the model per story, not per mood.

AIDEN runs your Claude Code sessions as cards on a board, with the model chosen per story, so heavy models go only where they pay. Free for one project.

Download AIDEN, freeSee pricing

macOS 12+ · Bring your own Claude Code or Codex · Your code stays local