# GPT-6.1 Terra's 48 won't tell me where it should click **Summary:** GPT-6.1 Terra scores 48 on Artificial Analysis's primarily text-based index. I would judge its visual reasoning and computer-use pitch on actual images, accepted outcomes, and cost across reasoning settings before assigning it production tasks. - Canonical: https://markhuang.ai/news/gpt-6-1-terra-visual-score-gap - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-10-26 - Section: News - Tags: GPT-6.1 Terra, OpenAI, Visual Reasoning, Computer Use, Model Evaluation, AI Pricing - Source: [Artificial Analysis](https://artificialanalysis.ai/models/gpt-6-1-terra) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![A glass optical cube faces a blue vessel obscured by ribbed glass, beside a stack of blank test cards](https://cdn.markhuang.ai/news/gpt-6-1-terra-visual-score-gap/hero.webp) *Conceptual illustration: a general test result leaves a separate question about what the model can see.* [Artificial Analysis lists GPT-6.1 Terra as released on October 26, 2026, with an Intelligence Index score of 48 at maximum reasoning effort](https://artificialanalysis.ai/models/gpt-6-1-terra) and a weighted benchmark cost of $0.40 per task. Its listed base prices are $1 per million input tokens and $5 per million output tokens. That is enough to get my attention as an option for a product that pays for inference. I would try Terra on a specific visual task before considering it as a general replacement. OpenAI's [model documentation](https://developers.openai.com/api/docs/models/gpt-6.1-terra) pitches it for visual reasoning, computer use, and writing. The Intelligence Index is primarily a text-based evaluation. I would test Terra on the screens it is supposed to interpret before deciding whether 48 makes it a bargain or a compromise. I also would not dismiss Terra from that score. For someone building catalog matching or screen automation, the buying decision depends on whether it handles the particular images and decisions the application needs. ## The score answers a different question Artificial Analysis's [methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking) describes Intelligence Index v4.3.2 as a combination of 10 evaluations. It explicitly calls the suite primarily text-based and English-language, with image-input performance evaluated separately. The index includes agentic work and coding, so it is relevant to broad capability. It is not a direct verdict on visual interaction. OpenAI lists matching product photos to catalog entries, interpreting technical drawings, and acting on application screens among Terra's intended uses. Those are vendor descriptions of where to try the model. The index number alone does not tell me whether it can distinguish two similar products or notice that a screen changed after an action. I find that distinction more useful than arguing over whether 48 is impressive. A model can be worth buying for a particular job without winning the aggregate, provided it can demonstrate that narrower ability. The source page also lists output speed as unavailable, so I cannot turn OpenAI's description of Terra as fast into an independently measured latency claim. ## A middle tier has to justify its place The skeptical question I find persuasive appears in the [Hacker News launch discussion](https://news.ycombinator.com/item?id=50248646): why choose Terra instead of a smaller model working harder or a larger model working less? That is a fair objection to paying for the middle tier. It leaves Terra with a specific job to prove it can do economically. Reasoning effort makes the comparison less straightforward. OpenAI supports settings from `none` through `max`, with `medium` as the default. The Artificial Analysis page reports the maximum-effort configuration. Its $0.40 benchmark task cost should not become a forecast for a product running medium, or a promise about the cost of a successfully completed workflow. [Simon Willison's independent pelican SVG comparison](https://static.simonwillison.net/static/2026/openai-pelicans-october-2026.html) shows how much the setting can change one deliberately narrow exercise. Terra used 22,629 output tokens at max, with an estimated cost of 11.32 cents, versus 1,026 tokens and 0.51 cents at low. Generating SVG code is not a test of understanding an input image, and one prompt cannot rank the models. I would compare effort settings before assuming Terra has one characteristic task cost. ## I want to catch the wrong click before it happens For a screen workflow, I would separate reading the screen from choosing an action and confirming the result. A model might identify the correct control while choosing an inappropriate action. It might choose the intended action but fail to notice that nothing happened. A single completion grade would make those failures harder to diagnose. My evaluation would include screens where the requested control is missing, similar objects compete for attention, or an earlier screenshot no longer matches the current state. These are proposed test cases, not failures I have observed in Terra. The application should let it stop when the visual evidence does not support proceeding. That is the same distinction behind [my Jev bounding-box analysis](https://markhuang.ai/news/jev-1-14-bounding-box-no-not-found): returning a usable location does not establish that the requested object exists. Keeping the input image and expected outcome would help a reviewer distinguish a visual mistake from a bad decision. Catalog matching would need a different acceptance check. I would include items with no valid match and judge the selected catalog entry against a known answer. A fluent explanation should not rescue a wrong match. I would want cheaper accepted matches, with review time included in the comparison. ## The saving has to survive the whole task The Artificial Analysis page reports 130 million output tokens across its Intelligence Index run. That is useful context for the maximum-effort result, not a prediction about the length of an ordinary answer. I would measure tokens and elapsed time on the same visual cases across the effort settings, then include retries and review in the cost comparison. I would give Terra the tasks where it meets the quality bar at an acceptable cost and delay. Success at catalog matching would still leave screen automation to prove separately. I like having another candidate at these prices. Before letting it act, I want evidence from the images it will actually receive and the outcomes I will have to own.