Run-time AI steps and credit metering

Run-time AI steps bring AI into the runs themselves: a step in your workflow that reads a document or a scraped page and produces structured output. They are the one deliberate exception to Tyrbo's determinism — browser-step replay stays fully deterministic, while AI steps trade determinism for judgment.

What AI steps can do

AI steps are ordinary steps you add to a workflow, with four actions:

• Extract structured data — turn a document or page into clean JSON in exactly the shape you define (fields, types), ready for the next step.
• Classify — assign an input to one of your categories, like routing support emails by topic.
• Summarize — condense long content into a short summary.
• Generate — draft text from instructions and input data.

They compose with everything else: scrape a page, extract structured fields from it with an AI step, then loop over the results into your CRM.

Models and keys

Each AI step picks a model from a curated list of Anthropic and OpenAI models. You don't bring your own API key — there is nothing to configure, no provider account to create, and no key to paste. Tyrbo supplies and manages the provider access; you pay in credits like everything else.

How metering works

AI step costs are proportional to what the models actually process:

1. Every AI step execution records the model used and the tokens in and out (tokens are the units AI providers charge by — roughly, chunks of text).
2. When the run completes, that usage is converted to credits using the posted per-model rates below. Bigger, more capable models cost more per token than smaller ones; each step's cost is rounded up to a whole number of credits.
3. The result appears in your credit ledger as an AI usage debit alongside the run's base debit — two line items, same run — and run history shows the per-step breakdown.

So a run with AI steps costs: the run's flat base (5 on your device, 10 in the cloud), plus AI usage credits that scale with how much text the AI steps processed.

The rate card

Credits per million tokens, by model:

| Model | Input tokens | Output tokens | | --- | --- | --- | | claude-haiku-4-5 | 200 | 1,000 | | gpt-5-mini | 50 | 400 | | gpt-5.1 | 250 | 2,000 | | claude-sonnet-5 | 600 | 3,000 |

For scale: extracting the fields from a typical one-page invoice with claude-haiku-4-5 processes roughly 2,500 tokens in and 300 out — about 1 credit. A long multi-page summary with claude-sonnet-5 might process 20,000 in / 1,500 out — about 17 credits. Rates are reviewed when providers reprice; the current table always applies.

Cost controls

Three guardrails keep AI spend predictable:

• Per-step token cap. Every AI step has a maximum-tokens option that hard-limits how much a single execution can produce. Set it to match the output you expect — a classification needs far fewer tokens than a long summary.
• Organization AI spend cap. Set an org-level cap on AI credit spend; it works like the overage hard cap, stopping AI usage before it exceeds the limit you chose.
• Pre-run balance check. Before a run with AI steps is queued, Tyrbo checks your balance against the base run cost plus an estimate of the AI usage. The actual amount debited is always the measured usage on completion, not the estimate.

Keeping costs down

• Choose the smallest model that does the job — classification rarely needs a top-tier model.
• Feed AI steps only what they need. Extracting from one scraped section costs less than passing an entire page.
• Set tight maximum-token caps and raise them only if outputs get truncated.

Where to see AI usage

Run history shows each run's total cost, and the credit ledger separates base run debits from AI usage debits, so you can see exactly which automations drive AI spend. Billing details: Billing, credits, and overage.