GPT-6.1 Sol errs less, but still pass prices as Format input
OpenAI says GPT-6.1 Sol cut factual errors from 11.4% to 7.7%. For ad video, give the run your prices and claims as input data instead of trusting recall.
A 7.7% factual-error rate is still a rate, and for an ad it means roughly one claim in thirteen could be wrong. OpenAI's GPT-6.1 Sol announcement says the share of responses with factual errors fell from 11.4% to 7.7% at low reasoning effort (read 2026-10-04). That is a real improvement, and it is not a reason to let a model recall your price, your dates or your legal line.
On a Sume Format run the fix is built in: put the facts in input and tell the run to use them.
Why does `input` beat a prompt?
Sume's Create a run page says input is an object of up to 64 top-level keys and 2 MiB, written whole to /workspace/inputs/sume-action-input.json in the run's workspace, and treated as data. The instruction field is capped near 4,000 characters of carried text, so a long product sheet does not fit there but does fit in input.
| Content | Field | Why |
|---|---|---|
| What to make and the tone | instruction | Carried to the agent as the task |
| Price, SKU, dates, claims | input | Written whole; the model reads it rather than recalling it |
| Product photos | attachments | Image files or public HTTPS URLs |
What does the instruction say?
Name the file and the rule in one sentence: use only the figures in /workspace/inputs/sume-action-input.json, and leave any claim out if it is not there. That makes a missing fact an omission instead of an invention, and an omission is easy to spot in review.
Does a better model remove the review step?
No. Check prices and claims in the finished video against your source sheet before you publish, and do it every time while the error rate is above zero.
Sources
Related posts
More in Formats
- Hashtag and caption limits in a Sume output schema, checked post-run
Use pattern, maxItems and maxLength in a Sume output_schema so captions and hashtags are checked after the run. A violation returns null output, not bad data.
- Know a Sume bulk batch is done: count terminal webhooks vs item count
A Sume bulk queue has no webhook. Put a webhook_url on each item, verify the signature, dedupe by request_id and count terminal events to the item count.
- Localized video batch: one Sume bulk queue per language
Split a multilingual batch into one Sume bulk queue per language so each language gets its own concurrency, spend cap, idempotency key and retry.
- Make Iterator and Array aggregator: build a Sume bulk items array
In Make, an Iterator splits an array into bundles and an Array aggregator joins them. Use the pair to turn sheet rows into a Sume bulk items array.
Written by Sume