Claude Pro Get How Many Tokens?

Claude Pro include token pikin?

Claude Pro no get fixed token allowance, and as of September 2026 Anthropic no publish any. Claude subscription dey measure usage by messages inside one rolling time window, no be tokens wey dem dey remove from balance. How far that window fit reach depend on the model wey you choose, how long your conversation don grow, and how much text each turn carry.

Na this be the answer wey this question rarely get, because people expect plan to work like API key. API key dey spend balance wey dem count in tokens, so you fit calculate the invoice by yourself. Plan dey sell access to one cap wey Anthropic set and adjust, and dem write the cap in messages, so no fixed token number dey to quote. Anybody wey give you one don make am up.

So ask another question. Wetin dey make my messages expensive? That one get real answer, and you fit act on am today.

Why subscription no fit quote token number

The API dey bill every request by tokens. Input tokens cover everything wey you send, while output tokens cover everything wey model write back. The two get different rates, na why input and output tokens cost different amounts before you estimate anything.

Consumer plans no get balance wey you fit spend. Anthropic dey limit how many messages you fit send inside a rolling window. The window open when you send your first message and close after a set number of hours. Paid plans get another, longer limit wey dem measure across one week. Message no get fixed size, so the same message count fit represent very different amount of work for two different people. Anthropic help pages describe the limits as approximate message counts for short conversation. Then dem warn say long conversations and large attachments go use the allowance faster. That warning na the complete explanation, but dem almost never explain the mechanism behind am.

Wetin make turn 20 for chat cost pass turn 1 by far

The model no dey keep memory between requests. E no get state, so every turn go send the complete conversation again: the system prompt, every earlier message wey you write, every earlier reply from Claude, and every file wey you attach. Your new question fit get twenty words. But the request wey carry am na the whole transcript.

So, cost of one turn dey increase as chat length dey increase, and e dey increase in a straight line. The rows below show the calculation with one set of round assumptions: system prompt near 1,000 tokens, your messages near 1,000 tokens each, and replies near 1,200 tokens each. These examples show how the process work. Dem no be measurement of your account.

[
  {
    "label": "Turn 1",
    "sent_this_turn": "2,000",
    "cumulative_input": "2,000"
  },
  {
    "label": "Turn 5",
    "sent_this_turn": "10,800",
    "cumulative_input": "32,000"
  },
  {
    "label": "Turn 10",
    "sent_this_turn": "21,800",
    "cumulative_input": "119,000"
  },
  {
    "label": "Turn 20",
    "sent_this_turn": "43,800",
    "cumulative_input": "458,000"
  }
]

Turn one dey send 2,000 input tokens. By turn 20, that same short question dey send 43,800, wey pass twenty times the amount, even though you type the same amount. Across the full conversation, you don send 458,000 input tokens, and most of am na the same text wey dey send again and again.

Twenty separate one-turn chats for send twenty times the first row above, and nothing else. Na the same questions with only small fraction of the load. Attachments make the difference bigger, because PDF wey you attach for turn two go send again for turn three, turn four, and every turn after that, whether the conversation still concern am or not.

Na this make two people wey dey use the same plan fit report completely different mileage. One person keep one thread open throughout the week and reach the cap by Wednesday. The other person open fresh chat for every task and hardly see any limit. Nobody do anything wrong. Their token usage differ by one order of magnitude because their habits differ.

For API, you fit reduce the cost of repeating content with prompt caching. E dey store the unchanged beginning of request, so later calls go bill that part with lower input rate. For subscription, you no control caching, so conversation length na the lever wey you actually fit control. The same mechanism dey dominate coding sessions, where file contents and tool definitions dey go with every turn. So make context for coding session small go help your limits pass any setting, and where tokens for Claude Code session really dey go na something wey you suppose read before you blame the plan.

Wetin each plan tier dey give you, compared with the others

No tier publish one fixed allowance. Wetin dem publish na relative comparison, and that comparison dey enough to help you choose. Free dey for the bottom, and free capacity fit reduce when demand for the service high. Pro dey above am. Max dey available for two sizes, described as multiples of Pro usage: roughly five times and roughly twenty times. Team and Enterprise dey priced per seat and get their own caps.

Take those multipliers as indication of wetin dem intend, no be contract. Anthropic dey adjust the caps, and dem no dey announce the changes ahead of time. Na another reason why no honest page fit give you one exact token number. For the money side of the same comparison, wetin Pro cost and where e get cap and the difference between the two Max tiers explain am. wetin the free plan really allow explain the lowest level.

Model choice na another multiplier wey dey add on top of the tier. The same question go use more of your window for the biggest model than for the smallest one, because heavier model cost more tokens to run. If you move routine work from Opus to Sonnet or Haiku, you often get more working time than upgrade fit give you. So how to match model with the job na the first thing to try when you dey run out of capacity.

How to check your own usage instead of guessing

Your own account na the only reliable source, and you fit reach am with two clicks. For claude.ai, open Settings, then Usage. E go show wetin you don use for the current window and when that window go reset. Check am once when replies start dey refused, and once after a long session. You go understand your own pattern faster than any published figure fit teach you.

For Claude Code, two slash commands dey do the same work from the terminal. /usage dey report plan usage and reset time. /context dey show wetin dey fill the context window now, and separate system prompt, tool definitions, files, and conversation history. When /context show say old conversation don take most of the space, /clear go start fresh session, and the cost per turn go drop close to the first row of the chart above.

For the API, the count exact, and you fit request am before you send anything:

curl https://api.anthropic.com/v1/messages/count_tokens \
  --header "x-api-key: $ANTHROPIC_API_KEY" \
  --header "anthropic-version: 2023-06-01" \
  --header "content-type: application/json" \
  --data '{
    "model": "claude-sonnet-5",
    "messages": [{"role": "user", "content": "Summarise the attached report."}]
  }'

Healthy response na one field, and na the number billing go use:

{"input_tokens": 14}

Every request wey complete dey report the same accounting when e finish:

{"usage": {"input_tokens": 21430, "output_tokens": 512}}

Send one short question as fresh request. Then send that same question at the end of one long thread, and compare the two input_tokens values. The difference na exactly the effect wey your subscription dey meter, written as a number.

Wetin to do when limit land for middle of task

1. Wait for the window. App go show the reset time, and the wait often short pass the workaround. Na the correct move be this when nothing dey due.
2. Move go lighter model. Smaller model cheaper per turn for every plan, and for some plans e dey use separate allowance, so work fit continue while heavy model dey rest.
3. Start fresh chat and carry only wetin the next step need. Paste the conclusion, no be the transcript. Cost per turn go drop immediately, and answers usually improve, because model no longer dey weigh thousands of tokens from discussion wey don settle.
4. Move the job go API, where tokens dey count per request and dem bill per token. No window dey wait for, and you fit see the price of each call before and after you make am.

Which option fit depend on whether na clock or workload be your problem. One-off crunch na timing problem. Wall wey you dey hit every Wednesday na workload problem, and e need API or bigger plan instead of better timing. full checklist for limit wey land for middle of task go guide the recovery step by step, and how rolling windows dey open and reset explain why clock dey behave the way e dey. If na API you dey head to, price am first: API measured against subscription for the same workload and wetin one million tokens really cost go give you the arithmetic before you move.

FAQ

How many tokens Claude Pro include?

Anthropic no dey publish token allowance for Pro, and as of September 2026 no such number dey wey we fit quote. Pro dey measure through messages inside a rolling window, with another longer cap wey dem measure across one week. How many messages fit enter depend on the model and how long your conversation be. Any page wey state specific monthly token figure for Pro dey guess. Open Settings, then Usage for claude.ai to see your own consumption and when your usage go reset.

Starting new chat go reset my usage?

No. Dem dey count usage for your account across the window, so new chat no go return wetin you don already spend. Wetin e change na the cost of every turn from that point. New chat no get transcript to send again, so e start for the low end of the chart above instead of continuing to rise. Opening fresh chat when topic change na the cheapest habit wey you fit use for any plan.

Why I hit the limit faster today than yesterday?

Most times na because today's work dey inside one long thread. Every turn dey send the whole conversation again, together with any attached files. So cost per message dey increase steadily, even though the message count wey you see no change. The other common cause na the model. Heavier models dey use the allowance faster for each turn, so afternoon with the largest model fit finish earlier than morning with smaller one.

I suppose move to the API to get real token counting?

Move if you need the numbers or better predictability. API dey count input and output tokens for every call. E also expose token counting endpoint wey you fit run before sending, and e no get window wey you need wait for. E dey bill per token, so heavy use fit cost more than flat subscription, while light use fit cost much less. Price your real workload against your current plan before you switch, because the answer depend on how many tokens you dey send.