智东西（公众号：zhidxcom）
编译 | 杨京丽
编辑 | 李水青
智东西10月27日消息，今天，OpenAI推出新模型GPT-6.1 Terra，主打视觉推理、精准电脑操作以及更自然的写作能力。相比上一代模型GPT-5.6 Terra，新模型在日常沟通和电脑任务方面进一步升级，可用于撰写提案、解读技术图纸、填写表单和编辑视频等工作。
GPT-6.1 Terra继续主打速度、能力与成本之间的平衡。其标准API价格为每百万输入token收费1美元（约合人民币6.67元），每百万缓存输入token收费0.05美元（约合人民币0.33元），每百万输出token收费5美元（约合人民币33.35元），较GPT-5.6 Terra和GPT-6.1 Sol均低50%，方便开发者将其用于读取发票、向会计软件录入信息等大批量自动化工作。
从模型定位来看，OpenAI的GPT-6和GPT-6.1系列包括Astra、Sol、Terra、Luna等不同型号。GPT-6.1 Terra侧重速度、成本和日常任务执行，适合大规模调用；GPT-6 Astra是OpenAI能力最强的旗舰模型，面向高难度推理和复杂任务；GPT-6.1 Sol在相对更低价格下提供接近Astra的能力，适合复杂编程和推理任务；GPT-6 Luna则主打快速、高效的规模化日常工作。
在ChatGPT Work和Codex中，使用同样的额度，GPT-6.1 Terra能够完成比GPT-5.6 Terra、GPT-6.1 Sol和GPT-6 Astra更多的任务。
GPT-6.1 Terra现已面向ChatGPT Work和Codex中的所有Plus、Pro、Enterprise及Edu用户开放。
目前，在Artificial Analysis智力指数榜单上，GPT-6.1 Terra的得分为48分，较GPT-5.6 Terra提升6分，领先Grok 4.7（xhigh）和Mimo-V2.6-Pro的47分，与GPT-6.1 Sol的52分成绩接近。
一、电脑操作得分提升15个百分点，部分任务表现追平GPT-6 Astra
OpenAI使用Agents’ Last Exam（Agent终极考试基准）评估GPT-6.1 Terra的电脑操作能力。这项基准包含多种专业任务，例如将建筑图纸转换为3D模型，以及逐镜头审阅动画电影，为修复团队生成带时间码的记录。
在中等推理强度下，GPT-6.1 Terra的得分较GPT-5.6 Terra提高15个百分点，成本约为后者的三分之一。它的表现与GPT-6 Astra相当，每token价格仅为后者的十分之一。
在最高推理强度下，GPT-6.1 Terra的token使用效率超过Claude Opus 5.5的3倍，每token价格为后者的四分之一，单项任务成本约为后者的九分之一。GPT-6.1 Terra还具备更快的速度，其任务完成速度比GPT-6 Astra和GPT-6.1 Sol快3倍以上。
在视觉推理基准MMMU-Pro测试中，GPT-6.1 Terra在各档推理强度下表现均优于GPT-5.6 Terra。新模型以低得多的成本，取得了接近GPT-6 Astra和GPT-6.1 Sol的成绩。
OSWorld 2.1用于测试AI Agent操作桌面应用、完成多步骤任务的能力，例如根据技术图纸制作3D模型。在该基准的离线测试集中，GPT-6.1 Terra在中等推理强度下，便超过了GPT-5.6 Terra的最高得分，单项任务成本也明显更低。
二、高推理强度下事实错误率降至4.3%，编程任务解决率翻倍
OpenAI称，他们今年在减少模型幻觉和事实错误方面投入了大量工作。在其内部一项使用易诱发错误提示词的基准测试中，低推理强度下，GPT-6.1 Terra的错误率从GPT-5.6 Terra的53.6%降至16.7%，下降近37个百分点。
在最高推理强度下，GPT-6.1 Terra的错误率较上一代下降18个百分点，降至4.3%，接近GPT-6 Astra和GPT-6.1 Sol的水平。
编程方面，在Terminal-Bench 4.0中，GPT-6.1 Terra的任务解决率达到50.3%，较GPT-5.6 Terra 21.5%的成绩翻倍。该测试考察编程Agent通过命令行解决问题的能力，例如修复网页加载时出现的布局偏移。
不过，OpenAI称，对于最复杂的编程任务，GPT-6.1 Sol和GPT-6 Astra仍是更好的选择。
Terminal-Bench-Science 0.1则考察Agent使用代码完成科学研究任务的能力，例如追踪延时摄影视频中的细胞分裂。在最高推理强度下，GPT-6.1 Terra的得分超过GPT-6.1 Sol，且成本更低。GPT-6.1 Terra与Claude Opus 5.5的得分差距只有5个百分点，估算成本则不到后者的五分之一。
三、延续GPT-6系列对齐改进，加强任务边界与提示词注入防护
OpenAI称，在提升AI的效率和普及性的同时，他们也在持续训练、测试并完善防护措施。GPT-6.1 Terra在其“前沿风险防范框架”下，达到网络安全能力的Critical阈值，并保留了为GPT-6 Astra开发的防护措施。
GPT-6.1 Terra还继承了GPT-6系列的对齐改进。根据OpenAI的评估，相比GPT-5.6 Terra，新模型更能在用户授权的任务范围内行动，并如实向用户说明自身限制和可能阻碍任务完成的因素。
在提示词注入测试中，GPT-6.1 Terra对试图使其偏离用户任务的恶意指令，也展现出较强的抵抗能力。
结语：以更低成本推进电脑任务自动化
GPT-6.1 Terra重点提升视觉理解、电脑操作和事实准确性的同时，降低了完成任务所需的时间与成本。与GPT-6.1 Sol和GPT-6 Astra相比，该模型具备更高的性价比优势，但在复杂编程任务上仍存在一定差距。
对于需要反复读取信息、操作软件和录入数据的工作，更低的单项任务成本，有助于扩大电脑操作等Agent任务的应用范围。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Compiled by Yang Kyung-Li
Editor, Lee Shui-qing
In the news of October 27th, Open AI today launched a new model GPT-61 Terra, which focuses on visual reasoning, precision computer operations and more natural writing skills. Compared to the previous generation model GPT-5.6 Terra, the new model is further upgraded in terms of day-to-day communication and computer assignments, which can be used to write proposals, interpret technical drawings, fill out forms and edit videos.
GPT-6.1 Terra continues to balance the speed, capacity and cost of the main hit. Its standard API price is US$ 1 per million input token (approximately RMB 6.67), US$ 0.05 per million cache input token (approximately RMB 0.33), US$ 5 per million output token (approximately RMB 33.35), 50 per cent lower than GPT-5.6 Terra and GPT-6.1 Sol, facilitating their use by developers for substantial automation, such as reading invoices and entering information into accounting software.
Based on model positioning, the series GPT-6 and GPT-6.1 in Open AI include different models such as Astra, Sol, Terra and Luna. GPT-6.1 Terra focuses on speed, cost and day-to-day task execution and is suitable for large-scale calls; GPT-6 Astra is the most capable flagship model of Open AI, oriented towards difficult reasoning and complex tasks; GPT-6.1 Sol provides the ability to approach Astra at relatively lower prices, suitable for complex programming and reasoning tasks; and GPT-6 Luna is primarily responsible for fast, efficient scale-up of day-to-day work.
In ChatGPT Work and Codex, using the same quota, GPT-6.1 Terra was able to perform more tasks than GPT-5.6 Terra, GPT-6.1 Sol and GPT-6 Astra.
GT-6.1 Terra is now open to all Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex.
Currently, on the Artificial Analysis Intellectual Index List, the GPT-61 Terra score is 48, an increase of 6 points over GPT-5.6 Terra, leading by 47 points in Grok 4.7 (xhig) and Mimo-V2.6-Pro, close to 52 points in GPT-6.1 Sol.
I. Computer operations score increased by 15 percentage points, with some missions scoring GPT-6 Astra
Open AI uses Ages' Last Exam (Agent Final Examination Benchmark) to assess GPT-61 Terra's computer performance. The benchmark contains a wide range of professional tasks, such as conversion of building drawings to a 3D model and an animated film-by-shot to generate time-coded records for the repair team.
Under the medium reasoning intensity, the GPT-6.1 Terra score was 15 percentage points higher than the GPT-5.6 Terra, at a cost of about one third of the latter. Its performance is comparable to that of GPT-6 Astra, whose price is only one tenth per token.
At the highest level of reasoning, GPT-6.1 Terra token was three times more efficient than Claude Opus 5.5, with a price of one quarter per token and a single mission cost of about one ninth of the latter. GPT-6.1 Terra is also faster, with tasks completed more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
In the MMU-Pro test, the visual reasoning benchmark GPT-6.1 Terra performed better than GPT-5.6 Terra under the various levels of reasoning. The new model achieved near GPT-6 Astra and GPT-6.1 Sol at much lower cost.
OSWorld 2.1 is used to test AI Agent ' s ability to operate desktop applications and perform multi-step tasks, such as producing 3D models based on technical drawings. With the concentration of offline tests on the benchmark, GPT-6.1 Terra exceeded the highest GPT-5.6 Terra score under the medium reasoning intensity and the cost of individual tasks was significantly lower.
II. Reduced factual error rate to 4.3 per cent under high reasoning intensity and doubled programming task resolution rate
Open AI claims that they have done a lot of work this year to reduce model hallucinations and factual errors. In one of its internal benchmarking tests using easily induced errors prompt, the error rate for GPT-61 Terra fell from 53.6 per cent of GPT-5.6 Terra to 16.7 per cent, a decline of almost 37 percentage points, with low reasoning intensity.
At the highest level of reasoning, GPT-6.1 Terra ' s error rate fell by 18 percentage points to 4.3 per cent from the previous generation, close to GPT-6 Astra and GPT-6.1 Sol.
For programming, in Terminal-Bench 4.0, the mission resolution rate of GPT-6.1 Terra reached 50.3 per cent, which is double the achievement of GPT-5.6 Terra 21.5 per cent. The test examined the ability of Agent to solve problems through command line, such as the layout deviations that occurred when the web page was repaired for loading.
However, Open AI states that GPT-6.1 Sol and GPT-6 Astra remain the preferred options for the most complex programming tasks.
Terminal-Bench-Science 0.1 examines the ability of Agent to use codes to perform scientific research tasks, such as tracking cell splits in delayed photographic videos. At the highest level of reasoning, GPT-61 Terra scored more than GPT-61 Sol and at a lower cost. The score gap between GPT-6.1 Terra and Claude Opus 5.5 is only 5 percentage points, and the estimated cost is less than one fifth of the latter.
III. Continuation of GPT-6 series alignment improvements to enhance mission boundaries and prompt injection protection
Open AI states that while improving the efficiency and accessibility of AI, they are continuously training, testing and improving protective measures. GPT-6.1 Terra, under its “frontier risk-prevention framework”, meets the CCIL threshold for network security capabilities and retains the protective measures developed for GPT-6 Astra.
GPT-6.1 Terra also inherited the GPT-6 series alignment improvements. According to Open AI ' s assessment, the new model is more capable of operating within a user-mandated mandate than GPT-5.6 Terra, and provides users with an accurate account of their own limitations and possible impediments to mission completion.
In the prompt injection test, GPT-6.1 Terra also demonstrated greater resistance to malicious instructions seeking to divert it from a user ' s mission.
Conclusion: Automation of computer missions at lower cost
GPT-6.1 Terra focuses on improving visual understanding, computer operation and factual accuracy, while reducing the time and cost of completing tasks. The model has a higher value-for-money advantage compared to GPT-6.1 Sol and GPT-6 Astra, but some gaps remain in complex programming tasks.
Lower individual task costs for jobs requiring repeated reading of information, operating software and data entry help to expand the application of Agent missions such as computer operations.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.