Pricing only makes sense in the context of a concrete workload. Each architectural blueprint below models a workflow's realistic token profile (system prompt, retrieved context, generated completion, cache hit share) and prices it across every model in the catalog.
A support bot answers customer questions using product documentation. Each turn sends a system prompt plus retrieved docs as input and generates a short helpful reply.
Retrieval-augmented generation embeds a query, retrieves top-k chunks, and asks the model to answer grounded in those chunks. Input is dominated by retrieved context.
An IDE or code-review assistant sends the open file plus relevant repository context and expects long, precise code output — output tokens dominate the bill.
Summarizing contracts, reports, tickets or transcripts: the entire document is input, and the summary is a small fraction of its length.
An agent plans, calls tools and iterates: every step re-sends the growing conversation, so input tokens compound across steps before the final answer.
Marketing copy, product descriptions, emails and SEO drafts: short instruction input, long creative output.
Turning unstructured text into JSON: invoices, tickets, leads, moderation labels. Structured input in, compact structured output.
Translating content across languages. Token counts vary by language — CJK scripts tokenize differently from Latin scripts, which changes cost per word.