OpenAI claims that GPT-6.1 Terra improves on GPT-5.6 Terra in desktop operation, visual reasoning, factual accuracy, and some coding benchmarks, while emphasizing cost and speed advantages. The published article does not disclose sample sizes, error ranges, or full test configurations, and the specific scores in several charts are not visible. These results are therefore best treated as directional evidence reported by the vendor, not as conclusive evidence of general task performance or independent replication.
Tasks the results can help assess: Visual understanding and multi-step desktop operation; general command-line software engineering tasks; prompts prone to factual errors; vendor-reported comparisons of cost and speed between models.
Tasks the results should not be generalized to: Success rates across all real-world workflows, untested specialist tasks, complex software engineering tasks, and undisclosed chart values or statistical significance.
Model versions covered: GPT-6.1 Terra; comparisons in the article include GPT-5.6 Terra, GPT-6.1 Sol, GPT-6 Astra, and Claude Opus 5.5.
Test environment or client: The article mentions Agents’ Last Exam, MMMU-Pro, OSWorld 2.1, Terminal-Bench 4.0, Terminal-Bench-Science 0.1, and an internal OpenAI factual-accuracy benchmark. It does not disclose the runtime environment, tool versions, or full evaluation configuration.
Reasoning levels and parameters: Some results are explicitly reported at low, medium, and highest reasoning effort. Temperature, sampling parameters, tool configuration, and number of runs are not provided.
The following are OpenAI’s descriptions of its evaluations, not independent retests:
Agents’ Last Exam: A professional-task benchmark covering tasks such as converting architectural drawings into 3D models and inspecting an animation shot by shot to generate timestamped notes for a repair team. The article reports some comparisons at medium and highest reasoning effort.
MMMU-Pro: Used for visual-reasoning comparisons. The article reports only relative performance and gives no scores that can be read in the body text.
OSWorld 2.1: Tests an agent’s ability to complete multi-step tasks through desktop applications. One example in the article is building a 3D model from a technical drawing.
Internal factual-accuracy benchmark: Consists of what OpenAI calls “prompts prone to errors” and compares error rates at low and highest reasoning effort. The article does not disclose the questions, sample size, or scoring rules.
Terminal-Bench 4.0: Evaluates the ability of coding agents to solve problems through a command line. The article gives fixing layout shifts during webpage loading as an example.
Terminal-Bench-Science 0.1: Evaluates agents’ ability to use code to conduct scientific tasks. The article gives tracking cell division in delayed video as an example.
| Benchmark or metric | Results reported by OpenAI in the article | Comparison or conditions |
|---|---|---|
| Agents’ Last Exam | At medium reasoning effort, GPT-6.1 Terra scores 15 percentage points higher, at about one-third the task cost of GPT-5.6 Terra. At the highest reasoning effort, OpenAI says it is more than three times as token-efficient as Claude Opus 5.5, with a per-token price one-quarter as high, bringing the cost per task to about one-ninth; it completes tasks more than three times as fast as GPT-6 Astra and GPT-6.1 Sol. | The page does not provide a complete, readable score table or details of the cost calculation. |
| MMMU-Pro | GPT-6.1 Terra outperforms GPT-5.6 Terra at every reasoning setting; OpenAI says its scores are close to those of GPT-6 Astra and GPT-6.1 Sol at much lower cost. | No specific scores readable in the body text, breakdown by reasoning level, or cost basis are disclosed. |
| OSWorld 2.1 | At medium reasoning effort, GPT-6.1 Terra surpasses GPT-5.6 Terra’s previously published best score; the cost per task is only a small fraction of the latter’s. | No specific score or cost figures are provided. |
| Internal factual-accuracy benchmark | At low reasoning effort, the error rate falls from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra. At the highest reasoning effort, the error rate drops by 18 percentage points to 4.3%. | OpenAI says the benchmark consists of prompts prone to errors; the question set, sample size, and error range are not disclosed. |
| Terminal-Bench 4.0 | At the highest reasoning effort, it solves 50.3% of tasks, more than twice GPT-5.6 Terra’s previously published 21.5%. | OpenAI also notes that GPT-6.1 Sol or GPT-6 Astra is better suited to the most complex coding tasks. |
| Terminal-Bench-Science 0.1 | At the highest reasoning effort, GPT-6.1 Terra surpasses GPT-6.1 Sol; its score is less than five points behind Opus 5.5, at an estimated cost of less than one-fifth as much. | The score unit, estimation method, and full scores are not specified. |
Values published in the article body: On Agents’ Last Exam, the difference at medium reasoning effort is 15 percentage points; the task cost on this benchmark is about 1/3 that of GPT-5.6 Terra. At the highest reasoning effort, OpenAI reports more than 3× the token efficiency of Opus 5.5, a per-token price of 1/4, and a task cost of about 1/9; speed is more than 3× that of GPT-6 Astra and GPT-6.1 Sol.
Internal factual-accuracy benchmark: At low reasoning effort, the error rate is 53.6% → 16.7%; at the highest reasoning effort, it decreases by 18 percentage points to 4.3%.
Terminal-Bench 4.0: 50.3% at the highest reasoning effort; GPT-5.6 Terra’s previously published score is 21.5%.
Terminal-Bench-Science 0.1: At the highest reasoning effort, it surpasses GPT-6.1 Sol, trails Opus 5.5 by less than 5 points, and has an estimated cost below 1/5 as much.
The body text gives only relative results for MMMU-Pro and OSWorld 2.1, with no specific scores. The embedded chart area on the webpage did not show readable data during this collection, so no chart values are transcribed or estimated.
These are paraphrased public claims by the vendor, not independently verified results. The article provides no sample sizes, confidence intervals, error bars, full prompts, number of runs, or cost-calculation formulas.
The conclusions supported by the article are limited to OpenAI’s reported relative improvements on several internal and public benchmarks, especially factual accuracy, Terminal-Bench 4.0, and some desktop-operation results.
OpenAI positions Terra as a model with speed and cost advantages, while explicitly advising that GPT-6.1 Sol or GPT-6 Astra may be preferable for the most complex software engineering tasks. Both this positioning and the comparisons come from OpenAI and cannot replace tests on the target workflow.
The benchmarks test different tasks, so comparisons of scores, costs, and speed cannot be treated as the same metric. Undisclosed evaluation configurations also limit external replication.
The page does not state a publication date. The collection date should not be treated as the publication date; if the article is updated later, this record should be checked against the new page version.
On 2026-10-27, the original article was opened through Tabbit, and its body text, article metadata, and chart area were read. The page displayed no publication-date field; the benchmark chart area in the article did not show readable values, so only numbers and qualitative comparisons explicitly stated in the body text were recorded. No sample sizes were inferred, no chart data was reconstructed, and no independent benchmarking was performed.
GPT-6.1 Terra
