AI vendors charge for every word the AI reads and every word it writes back. The path you choose to do the work changes the bill more than the model you pick. This page lets you move each lever and watch cost respond — same job, twenty times the cost depending on the design.
Small to premium is a 10× rate jump within Claude (Haiku 4.5 to Fable 5.1) and 100× within GPT-6 (Luna to Astra). Most production work doesn't need the top tier; a triage-then-validate pattern (a fast cheap model filters, a stronger model adjudicates the shortlist) captures most of the quality at a fraction of the cost. Read rates with the tokenizer in mind: Claude models from 4.7 onward count about 30% more tokens for the same text, so Sonnet 5.5's 33% lower rate works out about 13% cheaper per page than Sonnet 4.6.
Match the model to the decision, not the workflow.
Every page you send is re-read on every round. A 30-page protocol re-read four times in a back-and-forth workflow is 120 pages of reading for one job — before the AI has written a single word back. Documents with images and tables cost 2–3× more than plain text.
Send only what's needed. Prompt caching (under rounds) cuts the cost of re-reading unchanged documents.
What the AI writes back is the most expensive part of the bill. Verbose responses, long reasoning explanations, and structured-output overhead all bill at the premium rate. "Reasoning" models also produce hidden thinking-text you still pay for.
Constrain answer length. Ask for what you actually need.
A 4-round workflow doesn't cost 4× a single call. It can cost 10× or more, because round 4 re-reads the entire conversation history from rounds 1 through 3. This is where naive forecasts break, and where most of the hidden cost lives. Prompt caching flattens most of it: after the first round, re-read history bills at about a tenth of the input rate.
Cap rounds in code, set ceilings on agent loops, and cache what repeats.
When the AI looks something up and the database returns 2 pages of results, those pages enter the conversation and get re-read on every round after that. Five lookups in an 8-round workflow can quietly add 40 pages of reading to the bill.
Summarize lookup results before re-reading. Drop stale findings.
Multi-specialist designs look elegant in a diagram. In billing terms, each specialist runs its own full conversation. Three specialists on a 4-round workflow cost roughly 3× a single specialist's full run — even if they're collaborating on the same job.
Parallelism multiplies. Sequential delegation can be cheaper.
Different models produce different output volumes for the same task, independent of the rate card. A more expensive model can be more concise and still cost more, because the rate jump exceeds the token savings. A "reasoning" model can produce 3–5× the visible output because the hidden thinking is also billed. The toggle above lets you explore the directional effect. Tokenizers differ too: the tool applies Anthropic's published figure of about 30% more tokens for Claude 4.7+ models, but not other vendors' differences.
Cross-model comparisons here are directional, not like-for-like.