AI & LLM tools

Utilities for anyone building on top of hosted language-model APIs — counting tokens, estimating spend, and comparing costs across major vendors. Pricing is refreshed periodically against each provider’s public list price.

Token counts are heuristic, not exact. Each major LLM family uses its own tokenizer — OpenAI’s tiktoken, Anthropic’s internal BPE variant, Google’s SentencePiece — and they disagree on the same input by 5–15% depending on language and formatting. The counter here uses a tiktoken-compatible cl100k_baseencoding, which is correct for the GPT-4 / GPT-3.5 family and a close approximation for the rest. Treat the cost estimate as a budgeting floor: add a 10–20% margin for system prompts, retries, and the conversation turns you don’t see until you ship. Output tokens are billed at a higher rate than input tokens for most providers, so a chat workload with a 1:3 input-to-output ratio costs more per turn than a summarization workload with a 10:1 ratio at the same total token volume — model and shape of the call matter more than raw length.