---
name: bulk-work-on-free-models-verified
description: Run repetitive bulk jobs (trawls, categorization, extraction, first-pass summaries) on free gateway models without burning your main budget — with a hard, live verification that the model costs $0 before every run.
---

# Bulk Work on Free Models, Verified

Your main model is expensive and its context is precious. Bulk work — sorting a thousand messages, extracting fields from documents, categorizing a backlog — belongs on a cheap worker. Free gateway models are ideal, with one catch: "free" is a pricing state, not a property. Models rotate out of free tiers without announcement.

## The rule that makes it safe

**Before every run, fetch the provider's live model catalog and refuse to proceed unless every pricing component of the chosen model is exactly $0** — input, output, cache reads, cache writes, everything the catalog lists. If any component is nonzero or missing, abort and report. Never cache yesterday's "it was free" and never trust a model name containing the word "free" as proof.

## The pattern

1. **File in, file out.** The job reads input from files and writes results to files. Bulk content never passes through the main agent's context — that's the entire point. The main agent writes the job spec; the worker chews; the main agent spot-checks the output file.
2. **Chunk to fit.** Split large inputs into chunks under the model's comfortable context, process each, then run one synthesis pass if the job needs a single answer.
3. **Cap the calls.** Set a per-run call ceiling in advance so a malformed job can't loop forever.
4. **Denylist bad actors.** Some "free" models are anonymous, retain prompts, or are otherwise not worth the risk at any price. A $0 price tag doesn't waive judgment — exclude them permanently.
5. **Reasoning models need handling.** Many free models are heavy reasoners that will burn their whole output budget thinking. Pin reasoning effort low and cap output tokens to something sane.

## Verification duty

The worker's output is **always an unverified draft**. The main agent samples it, checks it against the source material, and only then reports results to the human. Bulk workers hallucinate counts, misclassify edge cases, and state errors with total confidence. The savings are real only if the checking is.

## Hard limits

- Nothing confidential ever goes to a free third-party model — client material, credentials, private documents. If the content is sensitive, the job runs on the main model or not at all.
- The worker never sends anything, never talks to the human, never acts on the world. It reads files and writes files.