Token (LLM)
A token is the unit of input and output a language model processes. Tokens are roughly word-sized but vary; common short words are one token, uncommon or long words may be three or four. Token count drives cost, latency, and the context-window limit.
Why it matters
Tokens are the model's view of text. When a user types "hello there friend," the model sees three tokens (or thereabouts; tokenizers vary). When the model responds, it generates one token at a time, predicting the next based on what came before. Every input and every output is measured in tokens.
Why this matters for users:
- Cost. Most LLM APIs price per token (input and output rates differ). A long conversation costs more than a short one. A model with a larger context window does not change the per-token cost; it changes how many tokens fit in one call.
- Latency. Generation latency is proportional to output token count. A 50-token response feels instant; a 5,000-token response takes seconds.
- Context window. The maximum input + output token count for a single call. Modern frontier models support 100K-200K+ tokens. Beyond that, the call fails or older content is dropped.
For agent products, token economics shape what is feasible. A multi-turn conversation with tool calls accumulates tokens fast; bounding context budget per turn is real engineering work.
How Ajeris handles this
Ajeris's agent loop bounds each turn with explicit token budgets sized to the channel. A 90-second voice call uses a tighter budget than an asynchronous WhatsApp thread. When context approaches the model's limit, older turns are archived into memory and the agent continues fresh.
The token cost flows through to the usage-based pricing the user sees. Most personal-use customers do not need to think about tokens; the per-message cost is small enough that the budget is invisible.
