Skip to content
Browse documentation
Intelligence / ATH documentation

ath predict

Predict a benchmark with a model, or prepare evidence for your agent.

Usage

$ath predict fran

Arguments

<benchmark>
Required benchmark ID, such as fran. Find IDs with ath stats.

What it does

In a terminal, uses a configured gateway and a model to produce a prediction, then asks before saving. With --json, it only returns evidence for your agent: no model call, key or prediction write.

Examples

Set a gateway key and select a model with ath models first for terminal predictions. --dry-run generates a prediction without saving; --as-of limits the evidence to a historical date and prevents a write. A saved prediction can later be evaluated with ath grade.

$ath predict fran --json
$ath predict fran --as-of 2026-06-01 --json

Options

--model <name>
the model for this run. See them with ath models
--gateway <name>
vercel or openrouter, when both keys are available
--as-of <date>
pretend it is this day, hiding everything after it. Does not write. Replay the past with ath backtest (YYYY-MM-DD)
--file <path>
read a file other than the one in this folder
-y, --yes
write the prediction without asking
--dry-run
show the prediction, and write nothing
--json
evidence for a harness, not a prediction

After the attempt

Use ath grade to record your result and check the prediction. The ath grade page covers scoring, dates and analysing a miss.

Related commands

Up nextath grade →