Token Cost Is the New Page Count for Your Notes

A token is the new page. Every paste to an AI is a page billed on the way in; the reply is a page billed on the way out. The math fits in your head, vendor prices are public¹², and the bill keeps asking: what is the smallest piece of my notes that actually needs to be there?

The page-count framing is older than this generation of models. On Hacker News in May 2023, a commenter writing as capableweb left the durable rule of thumb: "A helpful rule of thumb is that one token generally corresponds to ~4 characters of text for common English text," with "100 tokens ~= 75 words."³ Anthropic's own pricing page says the same thing in its FAQ: "1 token is approximately 4 characters or 0.75 words in English. The exact count varies by language and content type."¹ OpenAI's tokens explainer prints the matching pair: "1 token ≈ 4 characters" and "1 token ≈ ¾ of a word."⁴ Three independent sources, one ratio. That ratio is what makes a paste into a page.

The belief: an AI chat is just a paste box

Most people meet a hosted model through a chat window, and a mental model takes shape there: type, paste, read, paste again. There is no meter on the screen, no scale, no per-keystroke counter. The bill, when it arrives, is the surprise of a tool whose unit of work was invisible. That is the belief this essay replaces.

The paste-box belief is harmless until the work scales. One short prompt costs almost nothing. One vault of a thousand notes, summarised on a weekend, can easily produce a four-figure month — and a four-figure month is the kind of surprise that gets a credit card removed from a developer console at speed. The cost is not hidden by the vendor; the vendor publishes it. The cost is hidden by the interface, which is a different problem.

The way out is not a calculator. The way out is a mental model small enough to apply in front of every paste: this note is roughly N pages; the model will charge me on the way in, and again — usually more — on the way out. The numbers below make that model concrete.

The bill, in pages: a worked example

Take a 500-word note. At 1 token ≈ 0.75 words¹, it runs about 670 tokens — call it a page. Feed it to Claude Opus 4.8 at the listed $5 per million input tokens¹ and it costs $0.003 on the way in; a 300-word reply (~400 tokens) at $25-per-million output¹ costs $0.010. Round-trip: $0.013 per paste.

A 10,000-word wing of a vault is ~13,400 tokens, or about $0.067 to send into Opus 4.8 once. None of these numbers is large in isolation. That is precisely the trap: a small number per page, repeated thousands of times across a vault and a tool that auto-includes "context" from every adjacent note, scales sneakily. The reference matrix lists the rest of the menu (Sonnet 4.6 at $3 / $15 per million tokens, Haiku 4.5 at $1 / $5)¹, and choosing between Opus and Haiku is a five-times-cheaper input lever for the same paste.

What raises the bill in practice, beyond the napkin math, is what the chat window does for you without showing it: a hidden system prompt, retrieved context from neighbouring notes, prior turns of the conversation, and tool-use scaffolding. Anthropic's tool-use documentation lists tool-use system prompts of 290 to 804 tokens per request across the current Claude 4.x lineup⁵. Every one of those is a page silently added to your bill before your text arrives. The honest version of "$0.013 per paste" is "$0.013 plus whatever the wrapper added." Count what you can; assume the wrapper added more than nothing.

The structural discounts (and the discipline they imply)

Three pricing levers are public and consistent across vendors: model choice, prompt caching, and the Batch API. Each can change the bill by an order of magnitude, and each only pays off when you have already decided, deliberately, what to send. The discipline they reward is the one they depend on: choose the work, then choose the lever.

The first lever is model choice, and it is the bluntest. Anthropic's current list prices put Haiku 4.5 at $1 input / $5 output per million tokens, Sonnet 4.6 at $3 / $15, and Opus 4.8 at $5 / $25¹. That is a five-times spread on input from the cheapest tier to the flagship. The rule that follows is unromantic: match the model to the task. Anthropic's own pricing-page guidance lists it first under cost optimisation: "Choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning."¹ Most note-handling jobs — tagging, summarising, drafting a back-of-envelope outline — are simple tasks in that taxonomy.

The second lever is prompt caching, and it is the one a careful workflow can lean on hardest. Anthropic's pricing page publishes the multipliers directly: a 5-minute cache write costs 1.25× the base input price, a 1-hour cache write costs 2× base, and a cache read costs 0.1× base — ten percent of standard input¹. The page states the break-even out loud: "A cache hit costs 10% of the standard input price, which means caching pays off after just one cache read for the 5-minute duration (1.25x write), or after two cache reads for the 1-hour duration (2x write)."¹ OpenAI's prompt-caching guide reports the same shape from the other vendor: "Prompt Caching can reduce latency by up to 80% and input token costs by up to 90%."² Two vendors, two independent docs, the same ~90-percent input discount for content the model has already seen. That is not a coupon; it is a structural feature of how transformer inference works.

The honesty caveat: "pays off after one read" assumes the cached prefix actually is reused. A cache that nothing hits is a 1.25× write that pays for itself never. The discipline the discount rewards is workflow design — a stable system prompt, a stable glossary, a stable reference document — not a clever trick at paste time.

The third lever is the Batch API. Anthropic states it cleanly: "The Batch API allows asynchronous processing of large volumes of requests with a 50% discount on both input and output tokens."¹ OpenAI publishes the same headline number — a "50% cost discount compared to synchronous APIs"⁶ — for jobs you can let run in the background. Half-price on both sides for any work that does not have to come back in thirty seconds. Bulk vault summarising, mass tagging, archival re-rendering — the natural fit for batch is most of the work a note-app user ever does on more than one note at once. The lever rewards a single question being answered honestly: which of these jobs can wait?

There is one more economic surface worth knowing. Anthropic now offers the full 1M-token context window at standard per-token pricing on its current Opus and Sonnet 4.x lineup: "A 900k-token request is billed at the same per-token rate as a 9k-token request."¹ The shape that follows is dangerous. Long context is no longer rationed by a premium tier; it is rationed by you. A million-token paste is technically affordable and almost always the wrong move — which is the bridge to the durable lesson.

The durable lesson: prices fall; deciding what to send doesn't

Prices have been falling and may keep falling. Simon Willison's 2024 year-end review ran a section headed "LLM prices crashed, thanks to competition and increased efficiency," anchored on one line: "GPT-4o mini is $0.15/mTok—200x cheaper than GPT-4, nearly 7x cheaper than GPT-3.5."⁷ Two hundred times cheaper in eighteen months. A fact about the past, not a forecast.

The skeptic is right to refuse to draw the trajectory forward. The only thing the line legitimately says is that the number on the bill is not the durable problem. The durable problem is on the other side of the equation, and the academic floor for it has been in print since 2023. In Lost in the Middle: How Language Models Use Long Contexts (TACL, 2024), Nelson F. Liu and six coauthors reported that "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models."⁸ Bury the fact that matters in the middle of a giant paste, and a model with a million-token window will still miss it.

Industry replication has caught up. A 2025 technical report from Chroma Research, by Kelly Hong, Anton Troynikov, and Jeff Huber, tested eighteen frontier models across input lengths and named the failure mode context rot: "models do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows."⁹ (Chroma is industry research, not peer-reviewed; cite it as the eighteen-model replication of Liu et al., not as the industry consensus.) Two independent shops, one peer-reviewed paper and one open methodology report, agree on a result the marketing page does not advertise: more tokens does not buy more answer.

Which puts the durable question in a single sentence the bill keeps asking and the research keeps answering: what is the minimum context that actually answers the question? That question is invariant to next quarter's price drop. The price falls; the discipline does not.

What to do tomorrow

Five small habits, none of them clever, each of them paid back per paste.

1. Count before you send. A typical 500-word note is about 670 tokens; call it a page. For a vault wing of 10,000 words, expect about 13,400 tokens going in. wc -w is a usable estimate; OpenAI's tiktoken library¹⁰ is the exact count for OpenAI models. Anthropic's pricing page warns that "Opus 4.7 and later use a new tokenizer compared to previous models... This new tokenizer may use up to 35% more tokens for the same fixed text,"¹ so use tiktoken for order-of-magnitude, not for an Anthropic bill; for that, use Anthropic's own token-counting API.
2. Pick the model to the job, not the other way around. Tagging and summarising are Haiku-class problems. Use the flagship when the reasoning needs it; use the cheap rung when it does not.
3. Cache the stable parts. A system prompt, a glossary, a long reference document, anything that does not change between calls, should be sent through prompt caching once and read back at 0.1× thereafter¹². Design the workflow that reuses; the discount follows.
4. Batch what can wait. If a job does not need to come back in thirty seconds, run it through the Batch API at half price¹⁶. Vault-wide tagging on a Tuesday morning is not a real-time request.
5. Decide what not to send. The cheapest token is the one you never paid for, because you decided in writing that some folders are not eligible to be pasted at all, a sibling discipline covered in Some Notes Should Never Reach the AI. The companion question, what to keep on your own side after a summary, lives in Let the AI Summarise, but Keep the Original.

The room your model has to think in is finite (the context window) and metered (the token bill). The context window is working memory is the ceiling; this post is the meter. The structure of the note you feed it changes both sides of that math; how you structure a note is its AI retrieval. Whether what you send is then trained on is a separate axis covered in Does Your AI Assistant Train on Your Notes?. And the format you write in already saves you tokens before you do anything else: Markdown is denser and cleaner than HTML for a model to read, as Why an LLM Reads Your Markdown Better Than an Export walks through. The cluster anchor is What You Type Into AI Leaves Your Walls; this post is the missing economic post in the cluster.

A short closing honesty: the dollar figures in this essay are vendor list prices as of 28 June 2026. They will move. The Worked Example uses a clean round-trip on Opus 4.8 base pricing; a real workload includes system prompts, retrieved context, and tool-use overhead that the napkin does not show. Treat the math as the right shape, not the precise number. Re-pull the live pricing pages¹²⁶ before you bank a budget on them.

Frequently asked questions

How much does it cost to feed my notes to an AI? At Claude Opus 4.8's listed base pricing ($5 / $25 per million tokens input and output¹), a 500-word note is roughly $0.003 on the way in and $0.010 on the way back, about $0.013 per round-trip. A 10,000-word wing of a vault costs about $0.067 to send once. Numbers that small are exactly why they scale sneakily across a year of small pastes.

What is the difference between input and output token pricing? Output tokens cost three to six times more than input on every major model. Anthropic's Sonnet 4.6 sits at $3 input / $15 output per million tokens (5×); Opus 4.8 at $5 / $25 (5×)¹. The implication: short, well-shaped prompts that ask for short, well-shaped answers are the cheap path; rambling prompts that ask for long replies are where bills surprise people.

How many tokens is 1,000 words? How many pages is 32K tokens? About 1,333 tokens for 1,000 words, and roughly 24,000 words (call it forty to fifty paperback pages) for a 32K-token window. The conversion is the durable rule of thumb from capableweb's 2023 HN comment and both vendors' docs: "1 token ≈ ¾ of a word" / "100 tokens ≈ 75 words."³¹⁴

Does prompt caching actually save money? Yes, but only for content you actually reuse. Anthropic publishes the multipliers verbatim: 5-minute cache write at 1.25× base, 1-hour write at 2× base, and cache read at 0.1× base¹. OpenAI reports up to 90% off input tokens for cached content². The discount is real and structural; it only triggers if the cached prefix is read back, which is a workflow question, not a settings toggle.

Is the Batch API worth it for my notes? For non-time-sensitive work, almost always. Both Anthropic and OpenAI offer 50% off input and output tokens for asynchronous batch jobs¹⁶. Bulk vault tagging, summarising a year of journal entries, re-rendering an archive: anything that does not need to come back inside the minute is a natural batch job.

Why does sending more context sometimes give worse answers? Because models do not read uniformly. Liu et al.'s peer-reviewed Lost in the Middle (TACL, 2024)⁸ and Chroma Research's Context Rot (2025)⁹ independently find that performance degrades as input length grows, and that information buried in the middle of a long prompt is the most likely to be missed. A 1M-token window is technically affordable; it is almost never the cheapest way to get the right answer.

Will AI prices keep falling? They fell hard between 2023 and 2024. Simon Willison's year-end review for 2024 was headed "LLM prices crashed, thanks to competition and increased efficiency," and noted "GPT-4o mini is $0.15/mTok—200x cheaper than GPT-4, nearly 7x cheaper than GPT-3.5."⁷ That is a dated fact about the past, not a forecast. Plan on the current vendor list prices¹²; treat any extrapolation as your bet, not the vendor's promise.

Can I trust tiktoken to count tokens for Claude? No. tiktoken is OpenAI's BPE tokeniser¹⁰, and Anthropic's pricing page notes that "Opus 4.7 and later use a new tokenizer compared to previous models... This new tokenizer may use up to 35% more tokens for the same fixed text."¹ Use tiktoken as an order-of-magnitude estimator across vendors; for an exact Anthropic invoice, call Anthropic's own token-counting endpoint.

A token is the new page; every paste you make to the AI is a page billed both ways. The prices will move, the models will swap names, the windows will grow — and the one question that survives every quarter is still small enough to write on a sticky note: what is the smallest piece of this that needs to be there?

This essay builds on Simon Willison's running chronicle of LLM pricing⁷ and on the Lost in the Middle and Context Rot findings⁸⁹ for why bigger pastes do not buy better answers; the page-count framing was first written down on Hacker News in 2023 by capableweb³. Your notes are already plain Markdown, the kind of file you can count tokens on with one shell command and feed on purpose, one page at a time, and that file lives at mnmnote.com.