OpenAI has launched a new series of GPT-6 models, focusing on completing everyday office and computer tasks at a lower cost.
On October 26 Eastern Time, OpenAI released GPT-6.1 Terra, highlighting advancements in visual understanding, cross-application operations, and natural language generation. Alongside improved model performance, OpenAI reduced the standard API pricing by 50% compared to the previous generation, GPT-5.6 Terra. Independent evaluation firm Artificial Analysis also found that the new model shows improvements in intelligence level, task cost efficiency, and hallucination performance on specific benchmarks.
Enhance visual comprehension and computer operations throughout the workday.
OpenAI positions GPT-6.1 Terra as a fast and efficient model designed to help users complete computer tasks spanning an entire workday at a lower cost. The model can understand screen content, perform actions across applications, and generate writing that is clearer and more natural.
From drafting business proposals and interpreting technical drawings to filling out forms and editing videos, OpenAI aims to position Terra as a general-purpose tool for high-frequency office tasks. For businesses, these capabilities mean the model can take on more repetitive tasks, such as reading incoming invoices and entering related information into accounting software.
This upgrade does not merely aim to enhance a single capability, but seeks to integrate visual understanding, computer operations, and text generation into a unified workflow. OpenAI states that GPT-6.1 Terra shows significant improvements over GPT-5.6 Terra in computer operation and communication tasks, while emphasizing that the new model enables existing ChatGPT Work and Codex usage allowances to accomplish more tasks.
Computer operation benchmarks improved by 15 percentage points, with speeds more than three times faster than Astra.
In the Agents’ Last Exam benchmark, which measures agents' ability to use software to complete professional tasks, OpenAI reported that GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra under medium reasoning intensity, while its task cost was approximately one-third that of the latter. Its performance is even comparable to the flagship model GPT-6 Astra, but at one-tenth the price per token.
At maximum reasoning intensity, OpenAI states that GPT-6.1 Terra achieves more than three times the token efficiency of Claude Opus 5.5, with a per-token cost only one-fourth that of the latter. Combining its superior token efficiency and lower unit price, Terra’s cost to complete tasks is approximately one-ninth that of Claude Opus 5.5. In terms of speed, OpenAI reports that Terra completes tasks at an average rate more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
In terms of visual understanding, GPT-6.1 Terra outperforms GPT-5.6 Terra across all inference settings on the MMMU-Pro benchmark and approaches the performance of GPT-6 Astra and GPT-6.1 Sol at a lower cost. On the OSWorld 2.1 benchmark, which tests agents' ability to complete multi-step tasks using desktop applications, Terra also surpasses GPT-5.6 Terra’s previous best result under medium inference intensity.
These results reflect OpenAI's positioning of Terra: not necessarily striving for flagship-level performance across all capabilities, but rather aiming to accomplish a large volume of practical work at lower cost and in less time.
Independent evaluation: Intelligence level improved, cost to complete tasks significantly reduced
Artificial Analysis's independent evaluation further compared the performance and cost of GPT-6.1 Terra with its predecessor. The firm reported that Terra achieved a 6-point increase in peak reasoning strength on the Artificial Analysis Intelligence Index compared to GPT-5.6 Terra, while the cost per task dropped from $1.40 to $0.40—less than one-third of the previous amount.
At lower and medium reasoning intensities, Terra achieves a greater improvement in index scores compared to its predecessor, ranging from 7 to 13 points. Artificial Analysis believes these two reasoning tiers lie on the efficiency frontier of the trade-off between intelligence and cost, meaning users can achieve better model performance at relatively lower expenses.
In the Coding Agent Index v2.0 for coding agents, GPT-6.1 Terra scored 48, below GPT-6.1 Sol’s 54 but above GPT-6 Luna’s 31. The agency noted that Terra also sits on the cost-efficiency frontier for certain reasoning configurations in this index, indicating it is not only suitable for everyday office tasks but also possesses strong coding agent capabilities.
However, the strength of a model’s reasoning and the complexity of the task remain important. Artificial Analysis’s evaluation shows that Terra’s Intelligence Index score under maximum reasoning intensity still falls below that of GPT-6.1 Sol; OpenAI also explicitly recommends that users prioritize GPT-6.1 Sol or GPT-6 Astra for in-depth research or complex coding tasks.
Hallucination rate decreased, encoding task success rate improved
Reliability was another key focus of this upgrade. Artificial Analysis found on its private knowledge and hallucination benchmark, AA-Omniscience, that GPT-6.1 Terra provided guesses 38% of the time when it did not know the answer, compared to 88% for GPT-5.6 Terra—a 50-percentage-point reduction. Meanwhile, Terra’s accuracy on this benchmark improved by 2 percentage points.
It is important to note that this result measures the model’s tendency to guess answers under uncertainty within a specific benchmark and does not indicate a 50-percentage-point reduction in the model’s overall error rate across all real-world applications.
OpenAI’s internal testing also showed similar improvement trends. In prompt tests designed to easily induce errors, GPT-6.1 Terra reduced its error rate from 53.6% for GPT-5.6 Terra to 16.7% under low reasoning intensity—a decrease of 36.9 percentage points. Under maximum reasoning intensity, the error rate dropped by 18 percentage points to 4.3%, approaching the levels of GPT-6 Astra and GPT-6.1 Sol.
Coding ability has also shown significant improvement. On the OpenAI-released Terminal-Bench 4.0 benchmark, GPT-6.1 Terra achieved a task success rate of 50.3% at maximum reasoning intensity, more than doubling the previous 21.5% recorded by GPT-5.6 Terra. Independent evaluations by Artificial Analysis similarly show that Terra’s performance on this benchmark improved by approximately 20 percentage points, nearing that of GPT-6.1 Sol at maximum reasoning intensity.
This means Terra’s upgrade not only makes the model faster and cheaper but also reduces certain errors and improves performance on complex tasks. However, for the most complex software engineering tasks, OpenAI still recommends using Sol or Astra.
API prices halved, targeting automated tasks for high-frequency enterprises
Pricing is another key highlight of this GPT-6.1 Terra release. OpenAI has announced standard API pricing at $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. Compared to GPT-5.6 Terra, standard API pricing has been reduced by 50% overall; compared to GPT-6.1 Sol, GPT-6.1 Terra’s standard API pricing is also 50% lower.
The artificial analysis shows that for GPT-5.6 Terra, the input and output token prices were $2 and $12 per million, respectively, while for the new model, these have decreased to $1 and $5. The cached input price has dropped from $0.20 per million to $0.05 per million, a 75% reduction. The institution also reported that the cached write price for the new model is $1.25 per million.
In ChatGPT Work and Codex, OpenAI states that the same usage allowance enables GPT-6.1 Terra to accomplish more tasks than GPT-5.6 Terra, and even more than GPT-6.1 Sol and GPT-6 Astra. For enterprise users requiring repeated file access, office software operations, or handling large volumes of structured data, this cost difference may directly impact whether agents can be deployed at scale.
GPT-6.1 Terra is available to Plus, Pro, Enterprise, and Edu users of ChatGPT Work and Codex starting October 26, and developers can access it via API using the model name gpt-6.1-terra.
From a product segmentation perspective, OpenAI is further refining the capability and cost tiers of the GPT-6 series: Terra is designed for high-frequency, repetitive tasks requiring computer operations, while Sol and Astra continue to target in-depth research and complex coding. As model efficiency on specific tasks improves, the competitive focus is shifting from merely comparing maximum model capabilities to measuring the time, tokens, and total cost required to complete real-world tasks.