Question 1

How accurate is the LLM token calculator?

Accepted Answer

For OpenAI models (GPT-5, GPT-5 Mini, GPT-4.1 Nano), we use tiktoken — the exact same tokenizer OpenAI uses — so counts are 100% accurate. For Claude and Gemini models, we use Google's recommended 4-characters-per-token approximation, clearly labeled as ~estimated.

Question 2

What is prompt caching and how much does it save?

Accepted Answer

Prompt caching allows you to reuse repeated input tokens across API calls. When enabled, it saves 90% on input token costs (you pay only 10% of the normal input price). This is especially valuable for RAG pipelines and chatbots with large system prompts.

Question 3

How much does the Batch API save?

Accepted Answer

The Batch API processes requests asynchronously and saves 50% on both input and output costs. It's ideal for non-real-time workloads like data processing, content generation, and evaluation pipelines.

Question 4

What are reasoning tokens?

Accepted Answer

Models like Claude Opus 4.6 (in extended thinking mode) use additional 'thinking' tokens before generating their response. These reasoning tokens are billed at the output token rate and can be 2-4x the output token count, significantly increasing costs.