Can GPT-6.1 Terra help match your product photos?
A folder of product photos arrives with names such as IMG_4021. Someone has to work out which item belongs to which listing. GPT-6.1 Terra is worth a small trial if that search takes time every week. Ask it to suggest matches for a person to check, with a clear way to flag a size or variant it cannot identify.
What the new model changes
OpenAI released GPT-6.1 Terra on 2026-10-26. Its model documentation explicitly lists matching product photos to catalogue entries as a use case. It accepts images and text, so an application can provide a photo alongside candidate products and their descriptions.
TLDR AI and The Rundown AI highlighted the release in their 27 October editions. The lower API price makes it worth checking whether Terra can reduce the work of finding a match. Its accuracy on your catalogue still needs testing.
The useful first result is a short list of candidates and the evidence needed to choose. Forcing one answer every time can turn a missing label into a wrong listing.
OpenAI release changelogTerra capabilities and API accessNewsletter discovery: TLDR AI, 27 OctoberThe Rundown AI, 27 October
An illustrative example: three almost identical lamps
Illustrative example: a homeware shop has three catalogue entries — a small brass lamp, a large brass lamp and the small lamp in black. These are fictional products for explaining the decision.
A new photo shows the brass finish and the familiar lampshade, but contains no label or reliable size reference. The sensible answer is to identify the two brass candidates and request the product code or a measurement. Choosing the small lamp because it looks smaller in the frame would go beyond the evidence.
Now add a clear photograph of the product label. Ask the system to read the code, compare it with the catalogue and check whether the visible item agrees. OpenAI warns that vision models can misread small text and describe images incorrectly. Keep the label legible and check the proposed match.
Three outcomes that help someone check the result
For a pilot, ask the model for a possible product code, the details it can see and the information it still needs. We suggest these three outcomes. Your developer would need to add them to the application.
- Match to review: a candidate agrees with the visible attributes and supplied identifier. Show the original photo beside the catalogue entry so a person can confirm it.
- More information needed: two variants remain plausible, or a key detail cannot be read. Name the missing detail, such as size or label text, so the person checking the result knows what to find.
- No catalogue match found: none of the supplied candidates fits. Set the item aside for a person to identify.
Budget for the whole matching job
On 2026-10-28, OpenAI lists standard API rates for GPT-6.1 Terra at US$1 per million input tokens and US$5 per million output tokens. GPT-6.1 Sol costs US$2 and US$10 respectively. These rates apply to requests with up to 272 thousand input tokens. Tokens measure the information processed and generated. A million tokens does not mean a million photographs.
Images also consume billable tokens. Image size and detail, candidate images, repeated attempts and the generated answer affect the bill. There are different rates for longer requests, reuse of stored input, faster processing and processing in a specified region. Ask whoever builds the pilot to record actual usage.
Terra requires paid API access; the free API option does not support it. OpenAI lists the paid tiers as Build, Launch and Grow, each with usage limits. Your developer would still need to connect the model to the catalogue and build a screen for checking results. Confirm API access and that implementation cost before commissioning a pilot.
Include the time needed to prepare catalogue records and review results. Our automation-cost guide explains how to include setup and ongoing work without treating every minute released as cash saved.
Current OpenAI API pricingTerra API tiers and limitsEstimate the full workflow cost
A small trial to help you decide
Prepare a sample with known answers. Include clear matches, near-identical variants, unreadable labels and products missing from the supplied catalogue. Use public product material you are permitted to process or synthetic examples.
Keep the answer sheet separate from the model’s inputs. Give Terra and any comparison model the same photos, candidate records and instructions. Record matches presented as certain that turn out to be wrong, useful requests for more information, missed matches, API usage and the time a person spends checking each result.
Compare the total time to a correct decision with the manual process. Include checking and correction in both. This proposed evaluation has not been run: use a small sample to decide whether a larger trial is worthwhile.
When to try it, and when to wait
Run a small trial if the catalogue has dependable product codes and reference photos, the matching work recurs, and someone can verify the result. Start with one product family so that mistakes are easy to understand.
Wait if the catalogue itself cannot distinguish variants, or if identifying an item requires information absent from every photo. Improving labels or capturing a product code at the source may remove the search altogether.
For a useful first brief, choose one product family. Define what counts as a correct match, what to do when the variant is unclear, and how to handle an item absent from the catalogue.
Sources
Scope one catalogue task.
Tell forvertz which product family needs matching and what makes a match acceptable. A synthetic example is enough to discuss the first step; keep confidential client material private.
Discuss an AI workflowUpdated