Tech • AI • Robotics • Game
Why Claude Costs You More With Every New Message
5/10TL;DR
Tokens are the units AI models process and bill for, while context is the limited working memory that makes long conversations more expensive and less reliable over time.
KEY POINTS
AI reads tokens, not sentences
AI models do not interpret text as humans do. They break prompts into tokens, or fragments of words, and operate on those pieces as numerical units. A common word may be one token, while a rarer or longer word can be split into several, meaning the same sentence can consume far more tokens than expected.
Tokenization varies by model and language
Each model has its own tokenization system, so the same phrase may be split differently by GPT-4o and another model. Frequent terms tend to get their own tokens, while uncommon words are cut into smaller chunks. English often uses fewer tokens than French, largely because many models were trained more heavily on English text.
Why tokens matter for quality and cost
Tokens are effectively the currency of AI use. Both the input and the output are counted, and the total determines the bill. One example cited for Sonnet put pricing around $2 for input tokens and $10 for generated output, illustrating why long prompts and long answers quickly become expensive.
A rough rule of thumb for size
A token is often estimated at about four characters, though the exact ratio varies. By that rule, 431 characters amount to roughly 104 tokens. That estimate helps explain why large files, lengthy prompts and repeated back-and-forth exchanges can exhaust quotas or trigger usage limits.
Models do not truly remember between turns
A model does not carry memory from one message to the next on its own. To keep a conversation coherent, the system resends the prior exchange into the next request, which means every turn grows heavier. As a result, later messages cost more because the model must process the accumulated conversation again and again.
Context is a finite workspace
The context window defines how much information a model can actively use at once. A limit of 200,000 tokens corresponds to roughly 800,000 characters, but anything outside that window is effectively unavailable. If a file is too large or the running conversation becomes too long, the model may stop reading, truncate replies or lose precision.
Long chats can reduce reliability
As context fills up, models tend to retain the beginning and end better than the middle. Important instructions can fade, especially in prolonged sessions. That is why a model may suddenly reintroduce a discarded element, such as a popup, or forget a style requirement after many exchanges.
Compaction helps, but can lose details
When platforms offer to compact a conversation, they usually replace the full history with a summary. That keeps the session alive, but summaries can drop specifics with each pass. A first summary may omit one design rule, a second may lose another, and after repeated compression the final output can drift far from the original brief.
System prompts, tools and skills also consume context
The visible chat is only part of what fits inside the context window. Hidden instructions, tool definitions, memory files and skill descriptions may all be injected before the user’s message. Keeping those elements short and relevant improves efficiency, and a small set of well-designed skills is generally more effective than loading dozens of loosely relevant ones.
Fresh sessions and sub-agents can cut costs
A practical rule is one task, one session. Starting a new chat often works better than carrying long histories forward. Another strategy is to delegate narrow tasks to sub-agents, such as auditing separate advertising accounts independently and returning short summaries, which keeps the main context lighter and reduces token waste.
CONCLUSION
Understanding tokens and context is now essential to using AI effectively, because they directly shape cost, speed and answer quality. Shorter, cleaner sessions and tighter instructions generally produce better results than endlessly extending a single conversation.
Ask a question