The Claude AI token limit question has two answers: current Opus, Sonnet, and Fable models accept up to 1 million tokens of context and can generate up to 128K tokens of output per request, while Haiku 4.5 works with 200K context and 64K output. But those headline numbers only make sense once you know what a token actually is — roughly 4 characters of English text — and how input and output tokens are counted differently. This article explains the whole system, from the unit itself to the per-model limits table to what tokens cost on the API.
TL;DR:
• A token is roughly 4 characters or about three-quarters of an English word; 1,000 tokens ≈ 750 words.
• Input tokens (what you send, including conversation history) and output tokens (what Claude writes) are counted and priced separately.
• Current Opus/Sonnet/Fable models: 1M-token context, 128K max output. Haiku 4.5: 200K context, 64K output.
• On the API, output tokens cost about 5× more than input tokens on every model.
• On claude.ai, token limits show up indirectly — as usage budgets and the conversation-length ceiling.
What Is a Token, Exactly?
Claude doesn’t read letters or words — it reads tokens, chunks of text produced by a tokenizer. A token is usually a word fragment, a short whole word, or a piece of punctuation. The working rule of thumb: one token ≈ 4 characters of English, which works out to about ¾ of a word. So:
- 100 tokens ≈ 75 words (a short paragraph)
- 1,000 tokens ≈ 750 words (a blog post section)
- 100,000 tokens ≈ 75,000 words (a short novel)
- 1,000,000 tokens ≈ 750,000 words (a large codebase or several books)
Common words like “the” are a single token; rarer words split into pieces (“tokenization” might be 3–4 tokens). Code, non-English languages, and dense punctuation typically consume more tokens per visible character than plain English prose — worth remembering when you paste in a source file and it “weighs” more than its word count suggests.
Input Tokens vs Output Tokens
Every Claude request has two token meters running:
- Input tokens — everything you send: your message, the system prompt, attached documents, and crucially the entire conversation history, which is re-sent on every turn.
- Output tokens — everything Claude generates in response, including Extended Thinking’s reasoning tokens where enabled.
The distinction matters for two reasons. First, cost: on the API, output tokens are priced about five times higher than input tokens across the lineup. Second, limits: the context window caps the total (input plus output so far), while the max-output figure caps how much Claude can write in a single response. A model can read a million tokens but will still stop writing at its per-response output ceiling.
That re-sent history is also why long threads on claude.ai burn through your usage budget — a dynamic we cover in detail in our guide to Claude AI usage limits across Free, Pro, and Max.
Claude AI Token Limits by Model (2026)
Here are the current token limits for every active Claude model, as of July 2026:
| Model | Context window | Max output | Positioning |
|---|---|---|---|
| Claude Fable 5 | 1M tokens | 128K tokens | Newest flagship, thinking always on |
| Claude Opus 4.8 | 1M tokens | 128K tokens | Most capable Opus-tier model |
| Claude Opus 4.7 | 1M tokens | 128K tokens | Previous-gen Opus, still active |
| Claude Sonnet 5 | 1M tokens | 128K tokens | Best speed/intelligence balance |
| Claude Sonnet 4.6 | 1M tokens | 128K tokens | Previous-gen Sonnet, still active |
| Claude Haiku 4.5 | 200K tokens | 64K tokens | Fastest, cheapest tier |
Two takeaways from the table. The 1M-token context — once exclusive territory — is now standard on everything except Haiku, which stays at the long-time 200K standard. And max output has grown to 128K tokens (~96,000 words), which means a single request can genuinely produce a full-length report or a large multi-file code change. The history of that context growth, and what a million tokens actually holds, is the subject of our companion piece on the Claude context window.
Token Limits on claude.ai vs the API
On claude.ai, you never see a raw token counter. Token limits surface in two indirect ways: your plan’s usage budget (tokens consumed per 5-hour window and per week) and the conversation-length ceiling — when a chat’s accumulated history approaches the context window, Claude warns you and eventually asks you to start a new conversation.
On the API, tokens are explicit and everywhere: you set a max_tokens value per request, every response reports exact input and output token counts in its usage field, and your bill is a straight function of both. Rate limits are token-denominated too — each tier (1–4, plus custom enterprise arrangements) sets requests-per-minute and input/output tokens-per-minute per model, with 429 responses carrying a retry-after header. Our Claude API rate limits guide breaks down the tier system.
Text isn’t the only thing that consumes tokens, either. Images you upload for Claude’s vision analysis are converted into tokens based on their dimensions, so a screenshot-heavy conversation carries real token weight even if you type very little. Tool calls, code execution results, and web search results in agentic workflows all land in the context as tokens too — which is why a Claude Code session can consume the equivalent of dozens of ordinary chat messages.
How Tokens Map to Cost on the API
API pricing is quoted per million tokens (MTok), input and output separately:
| Model | Input / MTok | Output / MTok |
|---|---|---|
| Claude Fable 5 | $10 | $50 |
| Claude Opus 4.8 / 4.7 | $5 | $25 |
| Claude Sonnet 5 | $3 (intro $2 through Aug 31, 2026) | $15 (intro $10) |
| Claude Sonnet 4.6 | $3 | $15 |
| Claude Haiku 4.5 | $1 | $5 |
A worked example: send Sonnet 5 a 10,000-token prompt (about 13 pages of text) and get a 1,000-token answer back. At standard rates that’s 10,000 × $3/1M = $0.03 of input plus 1,000 × $15/1M = $0.015 of output — 4.5 cents total. The same request on Fable 5 costs $0.15; on Haiku 4.5, about a cent and a half. Prompt caching changes the math dramatically for repeated context: cache reads cost roughly 0.1× the input price, with cache writes at 1.25× (5-minute TTL) or 2× (1-hour TTL). For full tier-by-tier detail see our Claude API pricing explainer, or plug your own numbers into the Claude API cost calculator.
How to Count Tokens Before You Send
For quick estimates, divide characters by 4 or multiply word count by 1.33. For exact numbers, the API gives you two tools: a token-counting endpoint that returns the precise input token count for a request before you run it, and the usage object on every Messages API response, which reports actual input and output consumption. In practice I estimate with the 4-characters rule when drafting and rely on the usage field for billing math — the estimate is usually within about 10% for English prose, further off for code. Anthropic documents both in the Messages API reference.
What People Get Wrong About Token Limits
- Confusing tokens with words. A “1M-token context” is about 750,000 words, not a million words — and code or non-English text shrinks that further.
- Confusing the token limit with the usage limit. The context window is a per-conversation capacity; your plan’s usage cap is a spend budget over time. You can hit either one independently.
- Forgetting history counts as input. Turn 50 of a conversation re-sends turns 1–49. Long threads are quadratically expensive, which is why fresh chats are cheaper.
- Assuming input and output cost the same. Output is ~5× input on every model. Verbose responses, not big prompts, dominate many API bills.
- Expecting a 1M-token output. The context window is read capacity. No current model writes more than 128K tokens in one response.
FAQ
What is the Claude AI token limit?
Current Claude Opus, Sonnet, and Fable models accept up to 1 million tokens of context per conversation and generate up to 128K tokens per response. Claude Haiku 4.5 has a 200K context window and 64K max output.
How many words is a token?
One token is roughly 4 characters of English, or about three-quarters of a word. That means 1,000 tokens is approximately 750 words, and a 1M-token context holds roughly 750,000 words of plain English text.
What is the difference between input and output tokens?
Input tokens are everything you send to Claude, including the system prompt, attachments, and the full conversation history re-sent on each turn. Output tokens are everything Claude generates. They are metered separately, and on the API output tokens cost about five times more than input tokens.
How much do Claude tokens cost?
API pricing per million tokens runs from $1 input / $5 output on Haiku 4.5, through $3 / $15 on Sonnet and $5 / $25 on Opus 4.8, up to $10 / $50 on Fable 5. Sonnet 5 has introductory pricing of $2 / $10 through August 31, 2026.
How do I count tokens before sending a request?
For estimates, divide your character count by 4. For exact figures, the Claude API provides a token-counting endpoint that measures a request before you run it, and every response includes a usage field reporting actual input and output token consumption.
Does Claude on claude.ai have a token limit too?
Yes, but it surfaces indirectly. Your plan’s usage budget is consumed in tokens, and each conversation is bounded by the model’s context window – when a long chat approaches it, Claude prompts you to start a new conversation.
Next Steps
Once tokens click, everything about Claude’s limits and pricing becomes predictable arithmetic. The natural next read is our deep dive on the Claude context window — how the limit grew from 200K to 1M and how to actually use that space well. And for more quick answers, browse the full Claude AI FAQ section.