By Claude Pro in Portuguese: limits and tokens
Claude Pro spends your limit in tokens rather than messages, and Portuguese costs more tokens than English. Measure the gap and make your quota last.
Why does Claude Pro run out faster in Portuguese?
Claude Pro in Portuguese reaches its usage limit sooner than the same work in English, because the limit is spent on tokens and Portuguese turns into more tokens than English that says the same thing. A token is a small piece of text, often a few characters long. The model reads tokens and writes tokens, and your plan meters tokens.
On one short question the difference is small enough to ignore. Over a long conversation it is not, because every turn resends the whole conversation history as input. If each of your messages costs twenty percent more tokens, every later copy of that message costs twenty percent more too. That is why a plan which looked generous in an English-language review runs out somewhere around your thirtieth reply.
Your limit is spent on text volume, not on message count
Anthropic says this on its own Pro plan page. As of September 2026 it reads: "The number of messages you can send will vary based on message length, including the length of files you attach, the length of your current conversation, and the model or feature you use."
Every factor in that sentence is an amount of text. None of them is a count of messages. So "how many messages do I get on Pro" has no fixed answer, and the answerable version of the question is "how much text do I get". The same page says the session limit resets every five hours, and that Pro also carries a weekly limit which resets at a fixed time assigned to your account. For the mechanics of those windows on their own, how Claude counts a usage limit covers them, and what to do when the limit lands in the middle of a task covers the recovery.
Why Portuguese costs more tokens than English
Two separate causes stack on top of each other.
The first is the language itself. Translated Portuguese usually runs longer than its English source. "Configuration file" is 18 characters. "Arquivo de configuração" is 23. More characters means more tokens before any model is involved.
The second cause is the tokeniser, and it is the larger effect. A tokeniser is fitted on a training corpus, and it learns to merge the character sequences it sees most often into single tokens. English is over-represented in that corpus, so common English word forms became one token each. Many Portuguese forms did not, so they arrive split into pieces, and each piece is one token. Accented letters push this further, because the tokeniser works on bytes, and a letter such as "ç" or "ã" is two bytes in UTF-8 (unicode transformation format, 8 bit) rather than one.
You can watch this happen instead of believing it. Count the word coração on its own, then count heart on its own, using the command in the next section. Do not take a number from me for the size of that gap. The endpoint will tell you, for your model, today.
Measure the gap instead of reusing a remembered ratio
A ratio like "Portuguese costs 1.5x English" gets quoted second hand and goes stale quietly. It goes stale for a reason you can check: the tokeniser itself changes between model generations.
The data behind this chart
[
{
"label": "Before Claude Opus 4.7",
"words_per_1m_tokens": "750,000"
},
{
"label": "Claude Opus 4.7 and later",
"words_per_1m_tokens": "555,000"
}
]Anthropic's models overview, checked in September 2026, fits about 750,000 English words into a million tokens on the older tokeniser, and about 555,000 on the one introduced with Claude Opus 4.7. Anthropic's token counting documentation states the same thing directly: the same input text produces roughly thirty percent more tokens on the newer tokeniser, and the exact increase depends on the content. A ratio measured two model generations ago is not your ratio.
So measure your own. The token counting endpoint exists for this. It belongs to the developer API (application programming interface), which is a separate account and a separate bill from your Claude Pro subscription, so counting tokens does not spend your Pro usage. You need an API key from the Anthropic Console. Counting itself is free and carries its own rate limit, independent of message creation.
Send the Portuguese paragraph first:
curl https://api.anthropic.com/v1/messages/count_tokens \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "content-type: application/json" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "claude-sonnet-5",
"messages": [{"role": "user", "content": "Preciso de ajuda para configurar um servidor web em um VPS com Ubuntu. Já instalei o Nginx, mas o certificado TLS não está sendo emitido e o navegador mostra um aviso de segurança em todas as páginas."}]
}'Then send the English translation of the same paragraph, changing nothing else:
curl https://api.anthropic.com/v1/messages/count_tokens \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "content-type: application/json" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "claude-sonnet-5",
"messages": [{"role": "user", "content": "I need help setting up a web server on an Ubuntu VPS. Nginx is installed, but the TLS certificate is not being issued and the browser shows a security warning on every page."}]
}'Each call answers with a single line:
{"input_tokens": 14}That 14 is the count for the two-word example in Anthropic's own documentation, shown here so you know the shape of a healthy reply. Your paragraph returns a much larger number. Divide the Portuguese count by the English count and you have your ratio, for that text, on that model, on that day. An HTTP 401 with authentication_error instead means the key in x-api-key is missing or wrong.
Full method, so you or anyone else can repeat it
Record four things next to each result: the exact text you sent, the model value you sent, the date you ran it, and the input_tokens you got back. The model matters because the count comes from that model's tokeniser. The date matters because tokenisers change between generations, as the chart above shows.
Two honest limits on the number. Anthropic's documentation says the count is an estimate, and a real request may differ by a small amount. And the endpoint counts input only. Claude's reply is metered as well, and output is the more expensive half of the arithmetic, which the difference between input and output tokens goes into properly.
Repeat the pair on text that looks like your real work, not on one sample paragraph. A support ticket and a Python traceback do not share a ratio. Code, log output and numbers barely move between languages. Ordinary prose moves the most. This is why publishing one ratio would mislead you: the number is a property of your text, not of the language.
Why long conversations end early
A model has no memory between turns. The whole transcript is sent again as input on every turn, which is why Anthropic lists the length of your current conversation as a factor in your limit. Turn 20 sends turns 1 through 19 again, plus your new message.
The arithmetic is worth doing once. Suppose each turn adds about 1,000 tokens. Turn 1 sends 1,000. Turn 10 sends about 10,000. Ten turns together have sent about 55,000 input tokens to carry 10,000 tokens of new writing. Now raise every one of those numbers by your measured Portuguese ratio. The increase lands on the resent history as well as on the new message, so it grows with the thread rather than with the message.
That is the honest answer to "why did it stop after so few messages". The meter was never counting messages. It was adding up a transcript that you kept making longer.
Four changes that make a Pro subscription last longer
- Start a new conversation when the subject changes. A fresh thread resends nothing, so its first turn is cheap. Continuing yesterday's thread out of habit is the most expensive thing on this list.
- Group related questions into one message. Anthropic's usage limit best practices page recommends grouping multiple related tasks into a single message, and the resend is the reason: four separate turns each carry the history, one turn carries it once.
- Attach a document instead of pasting it, and never paste the same document twice. An attachment is not free, because Anthropic lists the length of files you attach as a factor. The win is that an attachment is one item you can remove, or carry into a fresh chat, while pasted text stays in the transcript for the life of the thread and is resent on every turn after it.
- Switch off connectors and tools you are not using in that chat. Their definitions are sent as input on every turn, and Anthropic's usage guidance describes tools and connectors as token intensive.
Does prompting in English actually help?
It helps, and it costs you something, so decide with your own numbers in front of you.
Write the instruction in English and ask for the answer in Portuguese, and the input side of every turn shrinks. The input side is the part that gets resent, so that saving compounds along the thread. The reply is still Portuguese and still carries the Portuguese token count, and output is metered too. The saving is real and partial.
The cost is that your prompt is now in your second language. A vague prompt gets a vague answer, and one round of re-asking costs more than the tokens you saved. If your English is weaker than your Portuguese, write in Portuguese and shorten your threads instead. That change is free, and it works on the larger term in the arithmetic.
When is the next plan up the real answer?
Measure for a week before you pay more. Keep threads short and group your questions, then watch whether you still hit the limit. If you do, you have a capacity problem, and a larger plan fixes it. If the limit only arrives on turn sixty of one enormous thread, a larger plan buys you a longer version of the same habit.
If you are upgrading, the plan pages carry the numbers rather than repeating them here: what Pro costs and what its limits actually are, the questions to answer before moving to Max and the gap between the two Max tiers. For readers paying in reais, what a Claude Pro subscription really costs in Brazil covers the currency and tax side, which usually moves the decision more than the token ratio does. And if your workload is API shaped rather than chat shaped, what a million tokens costs at list price is the comparison to run, next to whether a Pro subscription earns its price for your kind of work.
FAQ
Does Claude Pro count messages or tokens?
Anthropic's Pro plan page says the number of messages varies with message length, the length of files you attach, the length of your current conversation, and the model or feature you use. As of September 2026, every one of those is an amount of text, so the meter is text volume rather than a message counter. A hundred one-line questions and ten long threads with a PDF attached are not the same load, even though the message counts look similar.
How much more does Portuguese cost than English on Claude?
There is no single number, and anyone quoting one has measured one sample on one model. Send the same paragraph twice to https://api.anthropic.com/v1/messages/count_tokens, once in Portuguese and once in English, and divide the two input_tokens values. Ordinary prose moves the most. Code, log output and numbers barely move at all. Record the model id and the date beside the result, because the tokeniser changed between model generations and your ratio changes with it.
Should I write prompts in English to save my Claude Pro limit?
It reduces the input side of every turn, and the input side is what gets resent, so the saving grows along the thread. The reply is still in Portuguese and is still metered, so the effect is partial. Weigh it against the risk of writing a less precise prompt in a second language, because one round of re-asking usually costs more than the tokens you saved.
Why does one long conversation hit the limit faster than several short ones?
Because the whole transcript is resent as input on every turn. Ten turns in one thread send the earliest turns ten times over. Ten turns spread across five fresh conversations resend far less, since each new thread starts with an empty history. Starting a new chat when the subject changes is the cheapest habit change available to you.
Does the token counting endpoint use up my Claude Pro usage?
No. Token counting belongs to the developer API, which is a separate account and a separate bill from a Claude Pro subscription. Anthropic's documentation states that token counting is free to use, with its own requests-per-minute rate limit that is independent of message creation. You need an API key from the Console to call it, and calling it does not touch your Pro session limit or your weekly limit.