Open source: context compression for coding agents, 63% less tool output ⭐ optscale-kompressor on GitHub →
AI Cost Optimization

Claude Usage Limits: How the 5-Hour and Weekly Caps Work and How to Stop Hitting Them

November 10, 2026

~ 7 min

On this page
    Hitting Claude limits across your team?
    Book a Demo

    Claude usage limits cap how much you can use Claude in two windows: a five-hour session limit and a weekly limit. Both scale with your plan, both are shared across Claude on the web, desktop, mobile and Claude Code, and both are consumed by the amount of text Claude processes, measured in tokens, not by a fixed number of messages. Long conversations and large files therefore use far more of each limit than short questions.

    Below: the limits for each plan, how the two windows interact, why Claude Code drains them faster, how tokens are counted, where to check usage and how to stretch each limit. All figures come from Anthropic’s official documentation as of November 2026.

    ‍

    Claude usage limits by plan (as of November 2026)

    Anthropic does not publish subscription limits in tokens or messages; it publishes multipliers against the Pro plan. The table summarizes the official Pro, Max, Ultra and Team plan pages and Anthropic’s Fable model policy.

    Plan and price 5-hour session limit Weekly limit Fable models Claude Code
    Free, $0 Yes; varies with demand Not published Not available Not included
    Pro, $20/month Yes (baseline, 1×) Yes, all models Usage credits only Included
    Max, $100/month 5× Pro 5× Pro, plus a separate Fable limit Up to 50% of the weekly limit Included
    Ultra (formerly Max 20x), $200/month None for verified accounts 10× Pro, plus a separate Fable limit Up to 50% of the weekly limit Included
    Team Standard, 25/seat/month(20 annual) 1.25× Pro Yes, all models Usage credits only Included
    Team Premium, 125/seat/month(100 annual) 6.25× Pro Yes, all models Up to 50% of the weekly limit Included
    Enterprise, seat-based Yes, per seat Yes, per seat Premium seats: included; Standard: credits Included
    Enterprise, usage-based, and Claude API None; pay per token None Billed at API rates Via API key

    Every paid plan can enable usage credits to keep working past a limit at standard API rates, billed separately. Anthropic may also adjust limits to manage capacity and keep access fair, so treat the multipliers as current, not permanent.

    Claude API rate limits are a different system

    Claude rate limits on the API are not session or weekly allowances. They cap requests per minute (RPM), input tokens per minute (ITPM) and output tokens per minute (OTPM) per model, refill continuously, and sit under a monthly spend cap per usage tier: $500 on Start, $1,000 on Build and $200,000 on Scale. Only uncached input counts toward ITPM by default, so prompt caching raises throughput. Exceeding a limit returns HTTP 429.

    ‍

    How the 5-hour window and weekly cap actually work

    The Claude 5-hour limit

    The five-hour limit is a session allowance that starts with your first message and resets after a full five hours. It is often called a rolling 5-hour window, but it neither slides minute by minute nor resets at midnight: start at 9:10 and the allowance returns at 14:10, no matter how much you used. Verified Ultra accounts have no five-hour limit.

    The Claude weekly limit

    The weekly limit caps total usage across all models and resets at a fixed day and time assigned to your account, regardless of when you subscribed. On Max, Ultra and Team Premium seats, up to 50% of that weekly limit can go to Fable models, which consume it faster than other Claude models.

    The two caps stack: a developer can have session headroom on Thursday and still be blocked by the weekly cap. Because every surface shares one pool, an afternoon of document analysis in claude.ai also cuts what is left for coding.

    ‍

    Claude Code usage limits: why coding sessions drain faster

    Claude Code usage limits are the same limits as the rest of your plan: on Pro, Max, Ultra, Team and seat-based Enterprise, Claude Code draws from the same five-hour and weekly allowances as claude.ai, Claude Desktop and Cowork. What differs is consumption. Every turn re-sends the conversation so far, the project context (CLAUDE.md plus every file Claude has read) and the new prompt, and every tool call adds another request carrying its results. According to Anthropic’s Claude Code guide, the conversation history is the fastest-growing part.

    Three behaviors around Claude Code limits catch people out:

    •       Sign-in decides metering. With a subscription, Claude Code uses plan limits; with an API key, there is no plan cap and every token is billed. If an ANTHROPIC_API_KEY environment variable is set, Claude Code uses it instead of your subscription, and charges move to API billing.

    •       Switching models rarely rescues you. Session and weekly limits apply across all models; only a model-specific cap, such as an Opus or Fable limit, can be worked around with /model.

    •       Parallel work multiplies usage. Subagents, workflows and scheduled loops send their own requests, and Anthropic estimates agent teams use about seven times the tokens of a standard session when teammates run in plan mode.

    When a session limit hits mid-response, Claude Code may finish the step in progress under a small, capped wrap-up allowance: up to once per week on Pro and at each session limit on Max.

    ‍

    How Claude counts tokens

    Claude meters usage in tokens, not messages, which is why Anthropic publishes no fixed Claude message limits. For scale, Anthropic estimates an average 10 kB web page at about 2,500 tokens and a 500 kB research PDF at about 125,000. Claude 4.7 and later models use a tokenizer that produces roughly 30% more tokens for the same text than earlier models, per Anthropic’s pricing page.

    Each request counts four token types. Input covers everything sent: instructions, tool definitions, history, attachments and tool results. Output covers the response, including extended thinking. Cache writes store input for reuse at 1.25× the base input price for five minutes or 2× for one hour. Cache reads serve stored input at 0.1× on most models, 0.05× on Opus 5.5 and Sonnet 5.5, and 0.025× on Fable 5.1 and 5.2.

    Prompt caching matches the start of each request exactly, so changing the model, effort level or loaded tools mid-session forces the full history to be processed again. On claude.ai, reused Project content is cached and does not count against your limits. Together these mechanics produce four drains:

    Drain Why it costs First fix
    Context length The whole conversation is re-sent every turn New session per task; compact at natural breaks
    Model and effort Opus and higher effort use more quota; thinking is billed as output Sonnet by default; Opus for hard reasoning
    Cache misses A miss re-processes the full history at cache-write prices Keep model and effort fixed mid-task
    Cache misses A miss re-processes the full history at cache-write prices Keep model and effort fixed mid-task
    Tool output and connectors Logs, dumps, search results and tool definitions fill the context window Filter output; disable unused connectors

    How to check your Claude usage

    On claude.ai or Claude Desktop, open Settings > Usage. Paid plans show a Current session bar with the time until reset, a weekly bar with its reset day and time and, on Max and Ultra, a separate Fable bar. Verified Ultra accounts show no session bar.

    In Claude Code, run /usage. On a subscription, it shows the plan bars and attributes recent usage to skills, subagents, plugins and individual MCP servers, flagging any behavior behind 10% or more of usage; press d or w to toggle between 24 hours and 7 days. It reads local history only, so other machines and claude.ai are excluded. Since v2.1.251, it also reports the share of input served from cache. Run /context to see what currently fills the context window. On the API, the Console Usage page charts tokens and rate-limit headroom.

    Worked example: what one turn costs

    API list prices show how the drains compare. Take a Claude Code session whose context has grown to 150,000 tokens:

    Scenario Calculation Cost per turn
    Sonnet 5.5, context read from cache 150K × $0.10 per million $0.015
    Sonnet 5.5, no caching 150K × $2.00 per million $0.30
    Opus 5.5, context read from cache 150K × $0.20 per million $0.03
    Opus 5.5, cache expired and rewritten 150K × $5.00 per million $0.75

    One expired cache on Opus 5.5 costs as much as 25 cached turns. Subscription allowances aren’t published in dollars, but Anthropic names the same drivers (long context, heavier models, cache misses) as the fastest way to use up a limit.

    ‍

    How to get more out of each limit

    These seven habits target the four drains, roughly in order of impact.

    1.     Start fresh between tasks. Open a new conversation per topic in claude.ai and run /clear when you switch tasks in Claude Code. If your next prompt would make sense in a new session, clear first; /clear cannot be undone, so copy out anything you still need.

    2.     Compact at natural breaks. /compact replaces history with a summary and is cheapest while the cache is warm. Run it between tasks rather than waiting for auto-compaction, and use /rewind to return to an earlier, already-cached turn.

    3.     Match the model and effort to the job. Sonnet handles most coding, Opus is for cross-cutting refactors and hard debugging, and Haiku suits mechanical edits. Lower the effort level for routine work, since thinking tokens count as output.

    4.     Protect the cache. Pick the model and effort level at the start: both are part of the cache key, so changing either mid-task re-processes the whole history. On a subscription within plan usage, Claude Code already caches the main conversation for an hour; with an API key the default is five minutes, and setting promptCacheTtl to 1h pays off if you often step away.

    5.     Point to files instead of pasting them. Name the path and function so Claude reads selectively, trim logs and stack traces to the relevant 20–30 lines, and keep large dumps on disk.

    6.     Keep the fixed context lean. CLAUDE.md loads into every request, a fixed tax on your token budget, so hold it under about 200 lines and move workflow-specific instructions into skills that load on demand. Disable connectors and web search you are not using.

    7.     Filter noisy tool output before it reaches the model. A PreToolUse hook can reduce test output to failures, and subagents keep verbose operations out of the main context. For coding agents, the open-source OptScale Kompressor sits between the agent and the API, deterministically de-duplicates machine output, and never rewrites the cached prefix. In OptScale’s November 2026 benchmark (Terminal-Bench 2.1, 89 tasks, Claude Code on Opus 5.5), it cut the Anthropic API bill by 20%, mostly by keeping the cache warm during long tool runs. We measured that in API billing, where the default cache lifetime is five minutes. The method is in our post on Claude Code prompt caching.

    ‍

    Controlling Claude cost across a team

    Personal habits stop scaling because Team limits are per member and cannot be pooled or transferred, so an idle seat cannot absorb a colleague’s crunch week. Anthropic’s native controls cover the basics. Admins can mix Standard and Premium seats, enable usage credits with spend limits at organization, group or member level, and export per-user spend reports; API organizations can set workspace spend and rate limits in the Console. For budgeting, Anthropic reports that enterprise Claude Code deployments average about $13 per developer per active day and $150–250 per developer per month.

    Native controls leave three gaps:

    •       Teams holding their own API keys, or mixing Claude with other providers, produce spend that never reaches one report.

    •       Finance needs cost by feature, department or agent, not by seat.

    •       A proxy that strips Claude Code’s cache_control markers makes the entire conversation bill as uncached input on every turn, per Anthropic’s documentation.

    OptScale AI is built for that layer. The AI Gateway routes Claude and other models through one control point with role-based access and department-level billing, and keeps provider caches hitting instead of breaking them. AI Cost Optimization & Token Compression applies cache-safe compression and model selection inline, with no client changes; the benchmark behind it is in our token compression report. AI ROI & Impact then joins spend to delivered work, such as cost per merged pull request. For the finance view, see why token bills break the cloud FinOps playbook.

    Keep your team’s Claude spend in check

    Team attribution and cache-safe token compression on every Claude request. Start free with up to 10 seats

    Frequently asked questions

    What happens when I hit the Claude usage limit?

    Claude stops accepting new messages and shows when your limit resets. You can wait for the reset, upgrade (the new plan's limits apply immediately), enable usage credits to continue at standard API rates, or apply a limit reset if Anthropic has offered one on your plan. Claude Code may first finish a step already in progress.

    When do Claude usage limits reset?

    The five-hour session limit resets in full five hours after the first message of that session, not at a fixed clock time. The weekly limit resets at a fixed day and time assigned to your account, unaffected by when you subscribed. Both reset times are shown under Settings > Usage on claude.ai.

    Does Claude count input and output tokens toward the limit?

    Yes. Input includes your messages, files, conversation history, tool definitions and tool results; output includes Claude's replies and its thinking. Caching lowers the cost of repeated input: on claude.ai, reused Project content does not count against limits, and on the API, cache reads cost 2.5–10% of the standard input price, depending on the model.

    Are Claude Code limits separate from Claude Pro limits?

    No. On Pro and every other subscription plan, Claude Code shares one usage pool with claude.ai, Claude Desktop and Cowork, so all activity counts against the same five-hour and weekly limits. The exception is signing in with an API key: Claude Code then bills per token, and plan limits do not apply.

    Does Claude Max have a weekly limit?

    Yes. Claude Max limits are five times Pro's usage per five-hour session and per week, for $100 a month, and up to 50% of the weekly limit can go to Fable models. Ultra, formerly Max 20x, costs $200 a month and offers ten times Pro's weekly usage, with no five-hour limit for verified accounts.

    How do I use less of my Claude limit?

    Keep context small and stable. Start a new chat or run /clear between tasks, use Sonnet for routine work and Opus only for hard reasoning, avoid switching models mid-session, reference files instead of pasting them, keep CLAUDE.md short and disable unused connectors. Each step reduces what Claude re-processes on every turn.