How many Claude Pro tokens? 5-hour + weekly caps

How many tokens does Claude Pro include?

Claude Pro does not include a token allowance, and as of October 2026 Anthropic does not publish one. A Claude subscription is not a balance of tokens that you draw down. Anthropic measures usage in tokens, but Pro caps it with a session limit that resets every five hours and a weekly limit that applies across all models, and neither limit is published as a token figure. How far a session goes depends on the model you pick, how long your conversation has grown, and how much text each turn carries with it.

That is the answer this question rarely gets, because people expect a plan to work like an API key. An API key spends a balance counted in tokens, so you can add the invoice up yourself. An allowance-based plan such as Pro, Max or Team sells access to a cap that Anthropic sets and adjusts. Anthropic does not state that cap as a token number, and its Pro help page says only that the number of messages you can send will vary. So there is no fixed token number to quote. Anyone who hands you one has made it up.

So ask a different question. What makes my messages expensive? That one has a real answer, and you can act on it today.

Why a subscription cannot quote a token number

The API bills every request by tokens. Input tokens cover everything you send, output tokens cover everything the model writes back, and the two carry different rates, which is why input and output tokens cost different amounts before you estimate anything.

The consumer plans have no balance to spend on their included chat usage. Except on verified Ultra accounts, Anthropic caps your usage inside a session that resets every five hours, and the allowance-based paid plans (Pro, Max, Ultra, Team and older seat-based Enterprise contracts) carry a second, longer limit measured across a week. Usage based Enterprise plans work differently: Anthropic says there are no per-seat usage limits and no included token allowance, every token is billed at API rates, and administrators can set spend limits for the organization and for individual users instead. A message is not a fixed size, so the same message count is a very different amount of work for two different people. Anthropic's Pro help page gives no message count at all. It says the number of messages you can send varies with message length, the length of the files you attach, the length of your current conversation, and the model or feature you use. That warning is the whole story, and the mechanism behind it is almost never explained.

Why turn 20 of a chat costs far more than turn 1

The model keeps no memory between requests. It is stateless, so each request supplies the conversation context again: the system prompt, earlier messages and replies, and attached content still in context. Claude Code compacts long conversations into a summary, so that context does not always contain the entire original transcript. Your new question might be twenty words. The request carrying it includes the existing context too.

The input token count of a turn therefore grows with the length of the chat, and in this illustration it grows in a straight line. The rows below show that arithmetic on one set of round assumptions: a system prompt near 1,000 tokens, your messages near 1,000 tokens each, and replies near 1,200 tokens each. These illustrate the mechanism. They are not a 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 sends 2,000 input tokens. By turn 20 the same short question sends 43,800, more than twenty times as much, for the same amount of typing from you. Across the full conversation you have sent 458,000 input tokens, and most of that is the same text sent over and over.

Twenty separate one-turn chats would have sent twenty times the first row above and nothing else. The same questions, a fraction of the load. Attachments widen the gap when their contents stay in context, because that content counts as input again on later turns.

This is why two people on the same plan report completely different mileage. One keeps a single thread open all week and hits the cap by Wednesday. The other opens a fresh chat for each task and rarely sees a limit. Neither is doing anything wrong. Their token spend differs by an order of magnitude because their habits differ.

On the API you can soften the repeat cost with prompt caching, which stores the unchanged front of a request so later calls are billed at a reduced input rate for that part. Claude Code does this for you automatically, on a subscription as well as on an API key, and lets you disable it or choose how long the cache lives. Caching makes the repeat cheaper, but it does not remove it: every turn still carries the current context, so conversation length is the lever you actually hold. The same mechanic affects coding sessions, where file contents and loaded tool definitions add to the context, so keeping a coding session's context small helps preserve your limits, and where a Claude Code session's tokens really go is worth reading before you blame the plan.

What each plan tier buys, in relative terms

No subscription tier publishes an absolute allowance. What is published is relative, and the relative picture is enough to choose with. Free sits at the bottom, and Anthropic says the number of messages a free user can send varies with demand. Pro sits above it. Max costs $100 per month on the web and includes five times Pro's usage every five hours and every week. The former Max 20x plan is now Ultra, at $200 per month on the web, with ten times Pro's weekly usage and no five-hour limit for verified accounts. Team is priced per seat, and its limits apply to each member rather than to the team as a whole. Usage based Enterprise plans are the exception to this whole page: they charge a seat fee that covers platform access only and bill every token separately at standard API rates, with a 1.1x multiplier for US-only inference on Opus 4.6, Sonnet 4.6 and later models. Administrators can set spend limits, and self-serve Enterprise usage also stops for the whole organization when prepaid credits run out, until someone with billing permission purchases more. Every consumer cap belongs to an account rather than a person, so two people splitting one Pro login drain a single window between them, and the consumer terms forbid the arrangement anyway.

Treat those multipliers as a statement of intent rather than a contract. Anthropic tunes the caps, and its help pages reserve the right to limit usage in other ways, such as weekly and monthly caps or limits on particular models and features, at its discretion. That is another reason no honest page can hand you a token number. For the money side of the same comparison, what Pro costs and where it caps you and the gap between Max and Ultra cover it, and what the free plan actually allows covers the floor.

Model choice is a second multiplier sitting on top of the tier. The same question costs more of your window on the largest model than on the smallest, because a heavier model is more expensive per token to run. Dropping routine work from Opus to Sonnet or Haiku often buys more working time than an upgrade does, so matching the model to the job is the first thing to try when you keep running out. If that routine work later moves to the API, check what Haiku 5.5 costs per million tokens first, because prompts over 100K tokens are priced at five times the normal rate, so the long-thread habit that drains a subscription window gets more expensive there too.

How to see your own usage instead of guessing

Your own account is the only authoritative source, and it is two clicks away. On claude.ai open Settings, then Usage. It shows what you have consumed and when your limits reset. Check it once when replies start being refused, and once after a long session, and you will learn your own pattern faster than any published figure could teach you. The app also warns you before that point: a banner saying you've used 90% of your session limit means only a tenth of the current session is left, which is the moment to open a fresh chat or move to a lighter model rather than push a long thread further.

In Claude Code two slash commands do the same job from the terminal. /usage reports your plan usage limits, and on Pro, Max, Team and Enterprise plans it adds a breakdown of what counts against them. /context shows what is filling the context window right now, with a per item breakdown and suggestions for what to trim. When /context shows that old conversation dominates, /clear starts a new conversation with empty context and the per turn cost drops back toward the first row of the chart above.

On the API you can ask for a count 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."}]
  }'

A healthy response is a single field, shown here with an illustrative count:

{"input_tokens": 14}

Anthropic labels that number an estimate. The count says nothing about the output the model has not written yet. Every completed request reports the actual accounting when it finishes:

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

Send one short question as a fresh request, then send that same question at the end of a long thread, and compare the total input tokens, adding cache_read_input_tokens and cache_creation_input_tokens to input_tokens when caching is used. The difference shows how much more context the long thread carries; cached content counts less against subscription limits when reused.

What to do when the limit lands mid-task

1. Wait for the window. The app shows the reset time, and the wait is often shorter than the workaround. This is the right move when nothing is due. If your account has a limit reset available, Anthropic's Pro help page says you can use it to bring your five hour session limit or your weekly limit back to full, and how limit resets work covers when that is worth doing. If you would rather not wait, Pro subscribers can enable usage credits to keep working beyond the plan's included limits.
2. Move to a lighter model. A smaller model is cheaper per turn on every plan. On Max, Anthropic's Fable guidance lets you spend up to half of your weekly limit on Fable models at no extra cost, drawn from the same weekly limit as every other model, so once that share is used you can switch to another model or keep using Fable on usage credits. On Pro, Fable is not part of the plan's limits at all and runs on usage credits from the first message.
3. Start a fresh chat carrying only what the next step needs. Paste the conclusion, not the transcript. Per turn cost drops immediately, and answers usually improve, because the model is no longer weighing thousands of tokens of settled discussion.
4. Move the job to the API, where tokens are counted per request and billed per token. There is no five hour session to wait out, but the API has windows of its own. Rate limits are measured per minute in requests, input tokens and output tokens, and a request over any of them returns a 429 error, usually with a retry-after header saying how long to wait. The Start, Build and Scale usage tiers also carry a monthly spend cap of $500, $1,000 and $200,000, and once you reach it, API usage pauses until 00:00 UTC on the first day of the next month unless you request a higher limit sooner. You can estimate a call's input before you send it and read its exact usage after.

Which option fits depends on whether your problem is the clock or the workload. A one-off crunch is a timing problem. A wall you hit every Wednesday is a workload problem, and it wants the API or a larger plan rather than better timing. the full checklist for a limit that lands mid-task walks the recovery step by step, and how the rolling windows open and reset explains why the clock behaves the way it does. If the API is where you are heading, price it first: the API measured against a subscription for the same workload and what a million tokens actually costs give you the arithmetic before you move.

FAQ

How many tokens does Claude Pro include?

Anthropic does not publish a token allowance for Pro, and as of October 2026 there is no such number to quote. Pro usage is measured in tokens against a session limit that resets every five hours and a weekly limit across all models, and how many messages fit varies with the model, your attachments and the length of your conversation. Any page stating a specific monthly token figure for Pro is guessing. Open Settings, then Usage on claude.ai to see your own consumption and your reset time.

Does starting a new chat reset my usage?

No. Usage is counted for your account across the window, so a new chat does not give back what you have already spent. What it changes is the price of every turn from that point on. A new chat has no transcript to re-send, so it starts at the low end of the chart above instead of continuing to climb. Opening a fresh chat when the topic changes is the cheapest habit available on any plan.

Why did I hit the limit faster today than yesterday?

Usually because today's work lived in one long thread. Earlier messages and attached content still in context count as input again, so the input per message grows until context is trimmed or compacted. The other common cause is the model: heavier models draw down the allowance faster per turn, so an afternoon on the largest model ends sooner than a morning on a smaller one.

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

Move if you need the numbers or the predictability. The API reports input and output tokens on every call, exposes a free token counting endpoint you can run before sending, and has no five hour session to wait out. It still has windows: per minute rate limits that return a 429 error, usually with a wait time, when you exceed them, and on the Start, Build and Scale tiers a monthly spend cap that pauses usage until the first day of the next month once you reach it. It bills per token, so heavy use costs more than a flat subscription and light use costs much less. Price your real workload against your current plan before switching, because the answer flips depending on how many tokens you send.