Tokens & context windows

Language models read and write tokens — chunks of text roughly ¾ of a word. You’re billed per token for chat and embeddings; understanding them makes cost and limits predictable.

How AIx estimates tokens

Before a call, AIx tokenizes your input with tiktoken o200k_base (the GPT-4o family encoding). It’s a strong cross-model approximation — far more accurate than the naive “characters ÷ 4” rule — and falls back to that heuristic only if the tokenizer fails to load.

max_tokens

max_tokens caps the output length (default 256). It does not affect input cost. Set it deliberately for coding tasks — long completions cost more and can be truncated mid-answer if set too low.

Context windows

The context window is the total tokens (input + output) a model can consider at once. It’s model-specific — small models hold a few thousand tokens, frontier models hundreds of thousands. AIx forwards your messages to the provider; if you exceed the model’s window the provider rejects the call. Keep prompts within the target model’s window (the models page lists ids; consult the model’s own docs for its exact window).

Cost math

For token-priced models, price is quoted per 1M tokens:

cost = (prompt_tokens + completion_tokens) / 1{,}000{,}000 × price_per_1M

Embeddings are billed on input tokens only. Images bill per image (n × price), TTS by input tokens, and transcription by measured audio minutes. Image and transcription billing do not use chat-token accounting.

Request size limits

To protect the service, JSON requests are bounded:

Exceeding these returns 400 invalid_request_error. Large binaries (audio) go through the multipart upload path, not the JSON body.