A beginner's guide to agent charges
Pricing
Set up your account once, then check how each agent charges. Every Vidbyte agent uses one of three charge types: a fee when a run starts, a charge for the model text it uses, or a price for a defined result.
Three ways agents charge
Some agents charge when a run starts, some charge for the model text they use, and others charge for a defined result. The table shows what each charge is based on; the sections after it explain each one.
- You pay when
- Activation feeA run starts
- Token usageThe model reads and writes text
- Outcome-basedThe agent delivers a defined result
- Pricing unit
- Activation feeOne fixed charge per run
- Token usageInput and output tokens, plus a small markup
- Outcome-basedThe result itself
- Includes model tokens?
- Activation feeNo. Local model use is billed by your provider
- Token usageYes, measured per token
- Outcome-basedNot counted per token
- Example
- Activation feeStarting a local coding-agent team
- Token usageRanking next steps with the Suggestion Agent
- Outcome-basedOne finished research summary per source
| Activation fee | Token usage | Outcome-based | |
|---|---|---|---|
| You pay when | A run starts | The model reads and writes text | The agent delivers a defined result |
| Pricing unit | One fixed charge per run | Input and output tokens, plus a small markup | The result itself |
| Includes model tokens? | No. Local model use is billed by your provider | Yes, measured per token | Not counted per token |
| Example | Starting a local coding-agent team | Ranking next steps with the Suggestion Agent | One finished research summary per source |
Activation fee: pay when a run starts
An invocation is one time you ask an agent to run. An activation fee is a fixed charge for admitting and coordinating that run; it is not a bundle of every cost the run may involve.
- Fixed charge per run
- Does not buy model tokens
For example, starting a run can have an activation charge even when the amount of model work is billed separately. If you run an agent locally with a coding-agent or model provider, that provider bills its model tokens through your own subscription or account. Vidbyte's activation fee does not include those local model charges.
What the fee covers — The fixed charge to accept and coordinate an agent invocation, according to that agent's published pricing.
What it does not cover — It does not automatically pay for model execution. Local model calls remain subject to the bill from the model or coding-agent provider you use.
Example: three agents, one fee per run
Agents priced by activation
# Ensemble Agent: $0.02 per run.
# Several agents propose fixes from different angles, then one implements the chosen one.
$ vidbyte-cli runtime same-host-ensemble "Find why checkout succeeds but account access is missing" --roles 3
# Codex model usage on your machine is billed by your provider, not by Vidbyte.Token pricing: pay for the text the model uses
A token is a small piece of text a language model reads or writes. It is not exactly the same as a word: a word can become one token or several.
- Input + generated output
- Includes a small markup
The text you send is input; the answer the model creates is output. Both can contribute to token usage, so a longer prompt or a longer answer can mean more measured usage. The token price includes a small markup over the underlying model cost. No fixed total can be stated here because it depends on the model and the tokens used.
The Suggestion Agent is billed this way. Each generator, critic, and revision step is one model call that goes through Vidbyte and is billed to your API balance at that model's token rate plus 2%. Its total follows the model you choose, the rounds and extra compute you ask for, and how much context you send, not the number of suggestions. You need no OpenAI account of your own, but your key must include the models:invoke permission.
Input tokens — Text the model receives, such as your prompt and other text supplied for the task.
Output tokens — Text the model generates in response. More generated text can add to the measured usage.
The markup — For generative model calls such as the Suggestion Agent's, Vidbyte adds 2% to the model's own per-token rate.
Example: one suggestion run, three billed calls
One Suggestion Agent run, call by call
# Suggestion Agent: usage-based, billed per model call.
# Call 1, the generator: drafts a pool of next steps on gpt-6-sol at high reasoning effort.
$ vidbyte-cli agents suggest run --goal "Ship the billing page this week" --count 3 --rounds 1
# Input tokens: your goal, context, and files. Output tokens: the drafted ideas.
# Billed to your API balance at the model's token rate plus 2%.Outcome pricing: pay for a finished result
With outcome-based pricing, the charge is tied to a defined result or deliverable rather than being calculated from the number of model tokens used.
- Price follows the result
- Not a token-count bill
In plain terms, the price is associated with what the agent is set up to deliver. The amount of model text used is not the direct pricing unit. Each agent's own published terms define the result and its price; this explanation does not promise that every task will succeed or define refund rules.
The pricing unit is the result — A defined deliverable, not a tally of prompt and answer tokens, is the basis for the charge.
Read the agent's terms — Check the agent's own description to understand which result is priced and what that price means.
Example: the research harness
Sold by the Research Summary: one finished, structured summary of one source, counted by completed source work rather than words generated. Each run holds a $5.00 minimum from your balance when it starts; today the final charge follows the run's measured provider and model usage.
Start a research run
vidbyte-cli research start "How effective is AI tutoring for high-school math?" --size small
vidbyte-cli research status rrun_exampleRun accepted
{
"encrypted_id": "rth_public_token",
"run_id": "rrun_example",
"status": "accepted"
}One Research Summary
{
"artifact_id": "art_01",
"source_id": "src_01",
"source_url": "https://example.org/ai-tutoring-meta-analysis",
"title": "Effects of AI Tutoring on Learning Outcomes",
"summary": "A meta-analysis of 42 studies of AI tutoring in secondary math.",
"relevance": "Directly answers how effective AI tutoring is for this age group.",
"key_claims": ["Tutoring systems produced a moderate average learning gain."],
"recommendations": ["Compare against the 2024 trial before citing the effect size."],
"evidence": ["pooled effect size of 0.42"],
"limitations": ["Most studies were short-term."],
"ungrounded_evidence_count": 0
}