Costs, budgets and limits

Every Tvijo task has a cap set before the model is called, and every attempt — including a retry or a fallback — gets a row in the ledger. This page explains the three numbers you will actually see: the task cap, the task spend, and the account limit.

📊 The three numbers

A task is reserved before the first request, not priced afterwards. That ordering is the reason a runaway conversation cannot happen: the money is held first, then the call is made.

📊 The free preview, in numbers

New self-service accounts get a sponsored preview:

• 5 tasks
• 50,000 model tokens
• USD 0.05 total

…whichever ends first. The shared preview pool is USD 1 per UTC day. Preview covers the cheap routes; premium routes need paid credit. No card, no automatic debit.

These are preview numbers, not a price list. They can change, and the workspace shows the current ones.

🔹 Unknown is not zero

This is the rule that matters most in practice:

• A provider that refuses before generation (an authentication or balance error) costs 0.
• A timeout or a 5xx may have been billed — nobody knows.
• Tvijo records that as uncertain, keeps the task marked, and blocks further paid turns for that account until the state is resolved.

A tool that printed $0.00 in that moment would be lying, and you would plan your budget on the lie.

🔹 How to spend less without losing quality

1. Ask for a diff, not a rewrite. Output tokens are the expensive half of most tasks.
2. Attach a function, not a repository. Context dominates the input cost.
3. One question, one task. Bundling five questions into one prompt usually produces five shallow answers and a bigger receipt.
4. Judge cheaply. A read-only check of a draft costs a fraction of the draft.
5. Retry deliberately. There is no automatic paid retry, so a retry is a decision — make it after you know why the first attempt failed.

🔹 Working with a daily cap

If you run several lanes in parallel, remember that they share the day:

• Parallel lanes fail together when the cap is reached, while a sequential run degrades gracefully.
• A single lane with a large task can consume the day before the others start.
• Reserve the expensive step (judging, final assembly) for last.

🚧 When this does NOT work

• You need a fixed price per answer. No routed service can promise that; the receipt is the only real number.
• You need per-seat billing. The preview is per account and sponsored, not a subscription plan.
• You expect refunds for failed calls. An uncertain attempt is exactly that — uncertain. It is reconciled, not assumed free.
• You compare prices across providers from marketing pages. Compare receipts from your own tasks instead.

➡️ Next

• Models and routing — why the same question can cost differently twice.
• Recovery after a failed run — what to do when a task ends uncertain.
• Running a pilot — measuring cost per accepted result.