SSD Nodes Learn Hosting plans →
How to do am Matt ConnorBy Matt Connor

Claude limit reach? Wetin you fit do next

Claude show “usage limit reached”? See whether na 5-hour session or weekly window, then choose smaller model, smaller context, credits, or API.

Which Claude limit you reach

A Claude limit reached message dey name the window wey you dey wait for, and na that window go decide wetin you fit do next. For Pro and Max, session and general weekly limits dey cover every model at once, so changing model no go restore your access. Ultra general weekly limit too dey cover every model at once. If the limit name only one model, na exception be that: change model and continue work while that model remain blocked. Anthropic still advise Enterprise seat users wey reach usage limit to try lighter model with /model.

Claude Code dey print the window and the time e go reset for the same line:

You've hit your session limit · resets 3:45pm
You've hit your weekly limit · resets Mon 12:00am
You've hit your Opus limit · resets 3:45pm

The session window dey roll over every five hours. Weekly windows dey reset for one fixed time every week wey assign to your account, so the day and hour remain the same from one week to the next. Pro plan get the five-hour session window plus one weekly limit wey apply across all models. Max plan add second weekly limit wey apply to one model family by itself. Ultra too get general weekly limit and separate weekly Fable limit, but verified Ultra accounts no get five-hour limit. Accounts wey Anthropic never verify yet get temporary limit until Anthropic verify dem automatically. The actual ceilings dey inside the complete breakdown of Claude usage limits. This page start from where that one stop.

Na this part dey catch plenty people. Where session limit apply, every request count against the session window and weekly window at the same time. So one heavy afternoon fit finish the weekly allowance while the five-hour window still get space. Check the reset time before you plan the rest of your day. Reset for 3:45pm na small break. Reset on Monday at midnight no be so.

Read the meter before you guess

For claude.ai, open Settings, then Usage. Progress bars show how much of the five hour session window and each weekly window you don use. This go show you which one don empty, instead make you dey guess. Verified Ultra accounts no get Current session bar.

For Claude Code, run /usage. E go show the same plan bars, plus breakdown of where recent usage go. E attribute usage to skills, subagents, plugins, and individual MCP (model context protocol) servers as percentage of the total. Press d for the last 24 hours, and w for the last 7 days. E also flag any pattern wey account for 10% or more of recent usage. The two patterns you go see most na long context and cache misses.

Two things dey important to note. The breakdown dey calculate from session history wey dey stored for that machine. So work wey you do for another laptop or for claude.ai no go dey inside am, even though e use the same allowance, because Claude Code and your claude.ai chats dey share one subscription window. The dollar figure for the Session block on top na estimate wey API (application programming interface) users fit use. Claude Code calculate am with standard list rates, unless administrator set a modelPricing table for your organization's managed settings. For that case, e use the contracted rates, and the Total cost line go say at your organization's configured rates. For both cases, this no be your bill when you dey use subscription.

Smaller model dey help?

Yes, e dey help when the message name one model. Anthropic also recommend say Enterprise seat users wey reach usage limit make dem try am. After You've hit your Opus limit, run /model, choose Sonnet or Haiku, and the session go continue with everything wey dey context. For Pro, Max and Ultra, after You've hit your session limit or You've hit your weekly limit, /model no dey restore the shared allowance. If na Opus alone dey blocked, wetin spent weekly Opus allowance mean explain when Opus go return and whether usage credits or moving to Max fit bring am back earlier.

The switch get much more value before you hit the limit than after. Sonnet fit handle most coding work at a fraction of the Opus rate. So, leaving Opus as your default na one of the two habits behind most unexpected usage. The other one na session wey you never clear. Give mechanical edits and log triage to Haiku. Keep Opus for the design decision wey genuinely hard for you. how to choose between Opus, Sonnet and Haiku explain where each one justify its price.

Stop paying for context you stopped using

Claude Code dey send your complete conversation with every request. Every time Claude use tool, e send another request wey carry that batch of tool results. Prompt caching mean say system go read the history again with cached rate instead of full input rate. This make am cheaper, but e no make am free. So one-line question for session wey you open since morning still dey use the whole conversation.

/context
/clear
/compact Focus on the failing tests and the files we changed

/clear no cost anything, because e throw the context away instead of reading am. /compact must read the conversation wey e dey summarize, so compacting very large context na large request by itself. Use /clear when you move go unrelated work. Use /compact only when you dey inside task and you need the thread to continue. Run /rename before you clear am, so you fit find the session again with /resume.

Cache lifetime matter pass wetin plenty people expect. For subscription, while you dey within the usage wey your plan include, the main conversation cache dey live for one hour. Subagents and most other background requests get five minutes, even for subscription. The main conversation go reduce to five minutes once you start use usage credits. Five minutes also be the default for API key. The promptCacheTtl and subagentPromptCacheTtl settings fit override those defaults if you want different lifetime. Your first message after long lunch go miss the cache. So system go process the whole context again and write am back to the cache. With API list rates, cache write cost pass plain input: 1.25 times the base input price for five minute cache, and 2 times for one hour cache. Na why usage fit jump without any visible reason. Where prompt caching pays for itself explain the calculation.

Three other controls dey the same place. /mcp list the servers wey you configure, so you fit switch off the ones wey this task no need. Long CLAUDE.md dey load when session start, whether the work need am or not. So move specialized instructions enter skills, wey only load when you invoke dem. Extended thinking dey bill as output tokens, and na the expensive side. So reduce am with /effort, switch am off for /config on model wey allow am, or set MAX_THINKING_TOKENS=8000 for model wey get fixed thinking budget. Claude Code no fit switch thinking off for Opus 5.5, Sonnet 5.5, Haiku 5.5, or the Fable models. For those models, /effort na the control. This na Claude Code rule, no be model rule. For API, Haiku 5.5 accept thinking turned off at high effort or lower. Sonnet 5.5 fit reduce the thinking wey e do upfront with between_tools setting. But Opus 5.5 and the Fable models reject thinking off everywhere.

Split the job so e fit enter one window

Press Shift+Tab to enter plan mode before any big change. Claude go inspect the codebase and suggest one approach for you to approve. If the direction no correct, catching am for there go cost one plan instead of two hours of edits plus rework.

Delegate noisy operations to subagents. If you run test suite or read 10,000-line log inside subagent, the raw output go remain for that subagent context. Na only the summary go return to your main conversation.

Press Escape immediately when Claude start going the wrong way. /rewind go restore both the conversation and the code to an earlier checkpoint, so you no go pay to undo work wey you never want.

You suppose buy usage credits or use the API?

Usage credits dey let Pro, Max or Ultra account continue work after e don pass the plan allowance, and billing go follow standard API rates. Turn am on under Settings and then Usage for claude.ai with payment method wey you don attach. Then add funds, because credits na prepaid and dem only cover usage after you buy some. You fit also run /usage-credits for Claude Code, wey go open your billing settings for browser. Credits apply to Claude conversations and Claude Code terminal usage. The cap na monthly spend limit wey you set yourself with Adjust limit, or remove with Set to unlimited. The command need subscription login, so e no dey available when you authenticate with API key. Free plan no offer credits at all. So the first step na Pro at $20 every month, and wetin Pro add and where e limits fit stop you explain how far e fit carry you. If you dey add credits most weeks, bigger allowance usually cheaper. The difference between $100 Max and $200 Ultra mainly na the allowance: Max give 5x Pro usage every five hours and every week, while Ultra give 10x Pro usage every week with no five hour limit for verified accounts. Both na monthly web prices, but mobile pricing fit change, and both plans offer monthly billing only. Dem also include different monthly Claude Platform API credits: $100 for Max and $200 for Ultra. These credits no increase your limits for Claude or Claude Code.

The other option na plain API key from Claude Console. Dem bill am per token, and e no get session or weekly window at all. One thing you need watch na stray ANTHROPIC_API_KEY for your environment fit take over from your subscription. For interactive session, Claude Code go ask once whether to use the key and remember your answer. So quick yes wey you give months ago fit easy to forget. The Use custom API key toggle for /config fit change am back. For non-interactive -p run, e no ask anything. E go use the key whenever you set am. You go know from the bill, or from this:

API Error: Request rejected (429) · this may be a temporary capacity issue. If it persists, check https://status.claude.com.

That 429 na rate limit for the key, no be for your plan. Run /status and confirm say the active credential na the one wey you mean to use. When login and API key both dey configured, /status mark the one wey no dey use.

API dey cost how much for one afternoon?

The worked example below na one afternoon of agent-style coding: 200,000 uncached input tokens, 1.8 million cache reads, and 150,000 output tokens. Prices na Anthropic list rates as of October 2026. Every one of these rates dey quoted per million tokens. So, if that unit still dey feel abstract, wetin one million tokens really worth go turn am into reading and writing amount wey you fit picture.

ChartOne afternoon of agent work priced at API list rates, October 2026
Di data wey dey behind dis chart
[
  {
    "label": "Opus 5",
    "input_usd_per_mtok": 5,
    "cache_read_usd_per_mtok": 0.5,
    "output_usd_per_mtok": 25,
    "afternoon_usd": 5.65
  },
  {
    "label": "Sonnet 5",
    "input_usd_per_mtok": 2,
    "cache_read_usd_per_mtok": 0.2,
    "output_usd_per_mtok": 10,
    "afternoon_usd": 2.26
  },
  {
    "label": "Haiku 4.5",
    "input_usd_per_mtok": 1,
    "cache_read_usd_per_mtok": 0.1,
    "output_usd_per_mtok": 5,
    "afternoon_usd": 1.13
  }
]

The same afternoon cost $5.65 for Opus 5 and $1.13 for Haiku 4.5. Output na wetin dey push the total: for Opus 5, $25 per million tokens, compared with $5 for Haiku 4.5. So, the 150,000 tokens wey Claude write cost pass the 2 million tokens wey e read. Na why lowering the thinking budget dey save pass trimming files. If the question na which provider to pay instead of which model to run, three jobs wey dem price for both Claude and ChatGPT APIs show where each bill cheaper.

One date pass without any change. Sonnet 5 launch for $2 and $10 per million input and output tokens. Dem announce am as introductory pricing through 31 August 2026, with change to $3 and $15 scheduled for 1 September 2026. Anthropic cancel that increase and make $2 and $10 the standard price. So, the same afternoon still cost $2.26 for Sonnet 5.

The newest models bring the same afternoon cost down again. Opus 5.5, wey release on 22 September 2026, cost $4 for input and $20 for output. Cache reads cost 5% of the input rate, so the afternoon cost $4.16. Sonnet 5.5 keep the $2 and $10 rates but cut cache reads by half to $0.10, giving $2.08. Haiku 5.5 start at $0.10 and $0.50, and e charge five times that amount for any request wey prompt pass 100,000 tokens. So, the total depend on how large your context grow.

How dem calculate each afternoon figure

Opus 5: 200,000 uncached input tokens at $5 per million na $1.00. 1.8 million cache reads at one-tenth of the input rate na $0.90. 150,000 output tokens at $25 per million na $3.75. Total na $5.65.

Sonnet 5: $0.40 for uncached input, $0.36 for cache reads, and $1.50 for output. Total na $2.26.

Haiku 4.5: $0.20 for uncached input, $0.18 for cache reads, and $0.75 for output. Total na $1.13.

Opus 5.5: $0.80 for uncached input, $0.36 for cache reads at $0.20 per million, and $3.00 for output. Total na $4.16. Sonnet 5.5: $0.40 for uncached input, $0.18 for cache reads, and $1.50 for output. Total na $2.08.

All 3 chart rows assume the same token counts, so the gaps between dem na only price difference. Real afternoon fit vary based on how much of your context remain cached.

For scale, Anthropic own documentation put average Claude Code spend across enterprise deployments near $13 per developer per active day, and below $30 per active day for 90% of users, as of October 2026. Per-token billing no automatically cheaper than subscription. E win when your workload dey come in spikes, because quiet days cost nothing. E lose when you work every day. API against subscription, compared with real numbers get the crossover.

Run the long job on a VPS overnight

Some jobs no need make you dey watch dem: big refactor, test migration, or documentation sweep. Those jobs no need your subscription window too. Put dem for VPS (virtual private server) with API key, and dem go use per token billing while your plan allowance remain untouched for tomorrow interactive work. Cloud session dey run Claude Code for Anthropic servers by default, without server wey you need manage, but e dey share your plan limits. So wetin cloud session credits cover and when dem go finish go decide whether e really keep the job away from your window.

Install Claude Code for Ubuntu box and confirm say e dey run:

curl -fsSL https://claude.ai/install.sh | bash
claude --version

claude --version suppose print version string like 2.1.211 (Claude Code). If you see command not found here, e mean say ~/.local/bin never dey your PATH, so open new shell and try again. Na this same installer you go run for Linux desktop, where the CLI dey work together with beta desktop app and editor plugins.

Put the key for file instead of your shell history. Create ~/.claude-env with editor and put one line inside, export ANTHROPIC_API_KEY=sk-ant-your-key-here. Then lock the file down and load am:

chmod 600 ~/.claude-env
. ~/.claude-env
printenv ANTHROPIC_API_KEY | cut -c1-10

That last line go print the first ten characters. E enough to confirm say the variable dey set without putting the complete key for screen.

Now start the job inside tmux so e go survive when you close your laptop:

tmux new -s overnight
claude --bare -p "Convert every test under ./tests from unittest to pytest, run the suite, and fix what fails" \
  --allowedTools "Read,Edit,Bash" \
  --append-system-prompt-file ./CLAUDE.md \
  --output-format json > ~/overnight.json

Detach with Ctrl-b then d, and close the connection. Reattach later with tmux attach -t overnight. Claude Code for tmux session on VPS explain the session handling fully.

--bare important here. E skip auto-discovery of hooks, skills, plugins, and CLAUDE.md. E no ever read OAuth credentials too, so the run go use the API key and nothing else. Na this one keep the job away from your subscription. E also mean say the project rules don disappear, na why --append-system-prompt-file dey return dem deliberately. A -p run dey use set ANTHROPIC_API_KEY without asking either way. So the difference go show when the key fail to load: without --bare, the run go fall back to your subscription login; with am, the run go fail authentication and leave your plan alone.

For morning, read the result and cost from the JSON:

jq -r '.result' ~/overnight.json
jq -r '.total_cost_usd' ~/overnight.json

Claude Code dey exit 0 when e succeed and non-zero when the run fail. So wrapper script fit branch based on exit status and send the failure to you by mail. Set workspace spend limit for Claude Console before you leave am running. Agent wey dey retry loop for 3am fit spend real money. Cost control for agents wey you leave running go explain the caps wey you suppose set first.

One extra benefit if you dey use Max, Ultra, or Team. Those plans now include monthly API credits for Claude Platform: $100 every month for Max, $200 for Ultra, and for Team $20 per Standard seat plus $100 per Premium seat, pooled and capped at $500. Interactive Claude Code no dey covered, but claude -p run wey you start yourself with API key from your linked Console organization dey covered. So this job fit use dem. Unused credits expire at the end of every billing cycle. After dem finish, the organization API requests go stop until next month credits arrive, unless e still get purchased credits. If the job must finish, keep some purchased Console credits available. Otherwise e go stop with Credit balance is too low.

Give the job its own unprivileged user and its own git branch. Agent wey dey edit repository unattended go sometimes do something wey you for stop. Branch go make that a git checkout instead of incident. How to run coding agent safely on VPS cover the isolation.

For bulk work wey no interactive at all, like classifying thousands of records, Batch API dey process requests asynchronously with 50% discount on both input and output tokens. Most batches dey finish within one hour, but processing fit take up to 24 hours.

Wetín no go bring your access back

Apart from upgrading, two things fit restore your allowance: the clock, and a limit reset if Anthropic give you one. Dem dey give eligible plans limit resets sometimes, and if you use one, e go immediately restore either your five hour session limit or your weekly limit to full, depending on the offer. You no need wait until you reach the limit before you use am. Look for Reset for free inside the Resets section of Settings, then Usage, for the web or Claude Desktop, or check the limit message itself. The button no dey inside Claude Code for the terminal, but the reset still apply there because the limits dey shared.

Starting a new conversation no restore your allowance. Your account dey carry the usage, no be the conversation. But e make every message wey follow cheaper, because fresh conversation no get old history wey e need read again, so e still make sense to do am. Na cost reduction, no be reset.

Deleting old conversations no go refund anything. Those tokens don spend when dem process am.

For Pro, Max and Ultra, switching models after a session or general weekly message no restore the shared allowance. A message wey name a model dey respond to /model, and Enterprise seat users fit also try a lighter model after dem reach usage limit.

Retrying inside loop no dey do anything apart from keeping your terminal busy. The reset time wey the message print na the real answer, and /usage go show you the same clock without the error. For subscription, Claude Code 2.1.234 or later fit wait inside the open session and continue the interrupted task shortly after the reset, so you no need sit down dey watch am.

Upgrading na the one step wey go raise the ceiling. So if the bars no move after you pay, check why new plan fit leave the old limits in place before you conclude say the allowance spoil.

FAQ

How long until my Claude limit resets?

The message dey tell you. Session limits dey run for rolling five hour window, and the message go show clock time, like resets 3:45pm. Weekly limits dey reset for fixed time every week wey dem assign to your account, and the message go show the day, like resets Mon 12:00am. To see the bars and reset times wey apply to your plan before you hit any limit, run /usage for Claude Code, or open Settings, then Usage for claude.ai. Verified Ultra accounts no get five hour session limit. If Anthropic give your account a limit reset, the same page go show Reset for free button wey go return the window to full immediately.

Switching to Sonnet go make me continue work?

Yes, when the message name a model, and Anthropic also recommend say Enterprise seat users wey reach usage limit should try a lighter model. After You've hit your Opus limit, run /model, choose another model, and the session go continue with your context intact. For Pro, Max and Ultra, after You've hit your session limit or You've hit your weekly limit, switching models no go restore the shared allowance. Switching to cheaper model before you hit the limit na wetin really dey stretch the window.

Starting a new chat go reset my usage limit?

No. Usage dey count against your account, so new chat go start with the same empty allowance. But e go make each next message cheaper, because new conversation no carry history wey Claude need read again. /clear for Claude Code dey do the same thing and e no cost anything, while /compact need read the conversation wey e dey summarize, so compacting very large session na itself an expensive request.

API cheaper pass upgrading my plan?

E depend on how steady your usage dey. The worked afternoon for this page cost $2.26 on Sonnet 5 and $1.13 on Haiku 4.5 based on October 2026 list rates. Subscription na flat fee for capped allowance, so e dey better when you work most days. Per token billing dey better when your load dey come in spikes, because quiet days no cost you anything.

I fit keep a job running while I dey wait for the reset?

Yes, if you move am away from your subscription. Run am on VPS with ANTHROPIC_API_KEY set for the environment, and billing go happen per token instead of drawing from your plan window, so both no go compete. claude --bare -p inside tmux go survive when you close your laptop, and --output-format json go write total_cost_usd field wey you fit check in the morning with jq.

Dis article help you?