1Tokens The units a language model reads and writes
Anthropic defines tokens as “the smallest individual units of a language model,” which “can correspond to words, subwords, characters, or even bytes.” The model never sees your sentence. It sees a row of tokens.
A bag of Scrabble tiles where a common word comes as one tile and a rare word has to be spelled out from several. The model reads tiles and writes tiles back, one at a time. Anthropic's version: larger tokens make the model more efficient, smaller ones “allow a model to handle uncommon or never-before-seen words.”
About 3.5 English characters per token on earlier Claude models. Anthropic says its newer tokenizer, used from Claude 4.7 on, “produces approximately 30 percent more tokens for the same text.” Same paragraph, about 30% more tiles. That matters because tokens are how the work is metered: NVIDIA writes that “the cost of AI inference is often measured in terms of the cost per token.”
The October essay on running the tkal machine: “My token spend has climbed every month since that first article, and the tool did not get more expensive.” The essay does not break the climb into more questions versus longer ones. Both show up on the same meter.
A token is not a word. It is a slice of text, and how the text gets sliced depends on the model and the language: Anthropic says the characters per token “depends on the model and the language.” Two models can count the same email differently, which is why Anthropic tells developers to count with the model they plan to use.
2Context window The model's working memory for one request
Anthropic's definition: “all the text a language model can reference when generating a response, including the response itself.” It is “different from the large corpus of data the language model was trained on,” and works as a “working memory” for the model.
A desk. The training data is the library the model read long ago. The context window is the desk in front of it right now. Your question, the whole conversation so far, the files you pasted in, the instructions the app adds that you never see, and the answer being written all have to fit on that desk at once.
1M tokens on current Claude models, and a single request can generate up to 128,000 tokens (128k) of output. Other Claude models, including Sonnet 4.5, have a 200,000-token (200k) window. The desk got five times bigger. One request can also carry up to 600 images or PDF pages.
Not on a tkal page yet, which is part of why it is here. You have seen it in every model launch since the chatbots arrived, usually as the first spec on the slide. Anthropic's context window guide lists it model by model.
A bigger desk is not automatically better. Anthropic warns that “as token count grows, accuracy and recall degrade,” and it has a name for it: “context rot.” Chat apps such as claude.ai “can also manage the context window on a rolling 'first in, first out' basis,” so the oldest messages fall off the desk first. That is one reason a very long chat can lose track of how it started. What goes on the desk matters as much as how big the desk is.
3Inference Running a trained model to answer (vs training it)
NVIDIA defines inference as “the process where a trained AI model generates new outputs by reasoning and making predictions on new data.” Training is the other half: the model “learns to perform a specific task by adjusting its weights based on a set of training data.”
A restaurant. Training is culinary school: long, expensive, and paid for once per chef. Inference is every plate that leaves the kitchen during service, and every plate costs ingredients again. Your question is one order.
251% to 48%. In the first quarter of 2025, Anthropic's compute and infrastructure bill was 251% of its revenue, or $2.51 for every dollar it took in. Quarter by quarter the line ran 251%, 177%, 148%, 144%, 72% and 48%, from Q1 2025 to Q2 2026. In Q2 2026 that was $5,558 million (about $5.56 billion) of compute against $11,555 million (about $11.56 billion) of revenue. tkal's read of the S-1: the compute base is “committed 12 to 18 months ahead and does not move with demand,” so when revenue grew 144% in one quarter and compute grew 64%, the ratio fell by half.
Our read of Anthropic's S-1, filed October 26. The filing names its own risk in one sentence: “Our compute commitments are substantial and generally must be satisfied regardless of our revenue or utilization.”
Training and inference are not the same bill, and in that filing they arrive as one line. The S-1 table we read shows compute and infrastructure as a single number, with no split between the chips that trained the next model and the chips that answered this quarter's questions. That split is the number that would tell you how much of the 48% recurs with every question. It is not disclosed.
Should you care?
If you follow AI stocks: training happens in bursts, but inference runs every time anyone asks anything, and it is metered in tokens. That is the bill that keeps growing after a model is finished, and it sits under the GPUs and memory every other tkal page tracks. The thing to watch at Anthropic is the compute line as a share of revenue, against 48%, when the next filing lands.
If you just use chatbots: a long conversation costs more tokens and can lose accuracy. Starting a fresh chat with only what matters on the desk is often the better move.
This is education, not advice. Nothing here says to buy or sell anything.