今天真是个热闹日子。那边Anthropic 提交 IPO，这边 OpenAI 发布了最新模型 GPT-6.1 Terra。它的定位是“以更低价格，提供快速、强大的电脑操作能力”，将细致的视觉推理、精准的电脑操作能力与更清晰、更自然的写作表现结合在一起。
Sam Altman 兴奋地表示，“GPT-6.1 Terra 将开启电脑操作的新时代。”
与 GPT-5.6 Terra 相比，GPT-6.1 Terra 能力有了显著提升，尤其擅长处理日常工作中的沟通和电脑操作任务，从起草提案、解读技术图纸，到填写表格、剪辑视频，都能胜任。
Terra 继续拓展各模型系列在智能表现与成本之间的平衡，为用户提供速度、能力和价格之间更好的组合。它的标准 API 定价如下：
- 输入 Token：每百万 1 美元
- 缓存输入 Token：每百万 0.05 美元
- 输出 Token：每百万 5 美元
标准 API 价格比 GPT-5.6 Terra 和 GPT-6.1 Sol 低 50%。这让开发者可以用更低的成本自动处理大量工作，例如读取收到的发票，并将相关信息录入会计软件。
从今天起，所有 Plus、Pro、Enterprise 和 Edu 用户都可以在 ChatGPT Work 和 Codex 中使用 GPT-6.1 Terra。开发者也可以通过 OpenAI API 调用该模型，模型 ID 为 gpt-6.1-terra。
并且，在 ChatGPT Work 和 Codex 中，使用 GPT-6.1 Terra 也能让你的套餐额度发挥更大作用。同样的使用额度，可以完成比 GPT-5.6 Terra 更多的任务；与 GPT-6.1 Sol 或 GPT-6 Astra 相比，可完成的任务数量则会多出很多。
从视觉理解到实际操作
OpenAI 使用 Agents' Last Exam 基准测试评估 GPT-6.1 Terra 的电脑操作能力。该基准涵盖多种专业任务，例如根据建筑图纸创建 3D 模型，以及逐镜头审查动画影片，为修复团队制作带时间码的记录。
在中等推理强度下，GPT-6.1 Terra 的得分比 GPT-5.6 Terra 高出 15 个百分点，成本仅为后者的约三分之一。即使与 GPT-6 Astra 相比，它的表现也不相上下，但每个 Token 的价格只有后者的十分之一。
在最高推理强度下，Terra 的 Token 使用效率是 Claude Opus 5.5 的三倍以上，每个 Token 的价格则只有后者的四分之一，因此，完成同一任务的成本可以降低到约九分之一。
Terra 的运行速度也快得多，完成任务的速度超过 GPT-6 Astra 和 GPT-6.1 Sol 的三倍。
在视觉推理基准 MMMU-Pro 上，GPT-6.1 Terra 在所有推理强度设置下的表现都优于 GPT-5.6 Terra。它的得分甚至已接近 OpenAI 能力最强的前沿模型 GPT-6 Astra 和 GPT-6.1 Sol，但成本要低得多。
OSWorld 2.1 用于测试 AI 智能体能否操作桌面应用完成多步骤任务，例如根据技术图纸创建 3D 模型。在中等推理强度下，GPT-6.1 Terra 的成绩已经超过 GPT-5.6 Terra 的最佳成绩，而每项任务的成本却低得多。
事实准确性与编程能力显著提升
今年，OpenAI 投入了大量精力，减少模型产生的幻觉和事实错误。
在其内部设计的、容易诱发错误的提示词基准测试中，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%，下降了近 37 个百分点。
在最高推理强度下，错误率也下降了 18 个百分点，降至 4.3%，与 GPT-6 Astra 和 GPT-6.1 Sol 的错误率已非常接近。
GPT-6.1 Terra 的软件工程能力也比 GPT-5.6 Terra 有了大幅提升。
Terminal-Bench 4.0 用于测试编程智能体通过命令行解决问题的能力，例如修复一个页面加载时布局会发生偏移的网页。在最高推理强度下，GPT-6.1 Terra 能够解决 50.3% 的任务，超过 GPT-5.6 Terra 已公布的 21.5% 得分的两倍。对于最复杂的编程任务，GPT-6.1 Sol 和 GPT-6 Astra 仍然是更好的选择。
Terminal-Bench-Science 0.1 用于测试智能体能否通过编写和运行代码完成科学研究任务，例如追踪延时摄影视频中的细胞分裂过程。
在最高推理强度下，GPT-6.1 Terra 的表现超过了 GPT-6.1 Sol，同时成本更低。与 Opus 5.5 相比，它的得分差距不到 5 个百分点，而估算成本不到后者的五分之一。
安全性与对齐
在让强大的 AI 变得更高效、更容易使用的同时，我们也在持续训练、测试和完善安全防护机制。
根据 OpenAI 的准备度框架（Preparedness Framework），GPT-6.1 Terra 在网络安全领域达到了“关键”（Critical）级别的能力阈值，同时保留了为 GPT-6 Astra 开发的安全防护措施。
GPT-6.1 Terra 也继承了 GPT-6 系列在对齐方面的改进。根据评测，与 GPT-5.6 Terra 相比，GPT-6.1 Terra 更有可能将操作限制在用户指定的范围内，并如实告知用户自身的能力局限和可能遇到的障碍。
在提示词注入测试中，GPT-6.1 Terra 也表现出较强的抵抗能力，能够抵御那些试图通过恶意指令将模型引离用户原定任务的攻击。
模型细节请参阅模型卡：
- 模型卡地址：https://deploymentsafety.openai.com/gpt-6-1-terra/gpt-6-1-terra.pdf
- 博客地址：https://openai.com/index/introducing-gpt-6-1-terra
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。
What a day. Anthropic over there submits IPO, and Open AI publish here has the latest model GPT-6.1 Terra. Its positioning is “to provide faster, robust computer operations at lower prices”, combining detailed visual reasoning, precision computer operations with clearer, more natural writing performance.
Sam Altman was excited to say, "GPT-61 Terra will open a new era of computer operations."
Compared to GPT-5.6 Terra, there has been a significant improvement in GPT-6.1 Terra ' s capacity, especially in communications and computer-operated tasks in day-to-day work, ranging from the drafting of proposals, the interpretation of technical drawings and the filling out of forms and clipping videos.
Terra continues to expand the balance between smart performance and cost of model series, providing users with a better mix of speed, capacity and prices. Its standard API pricing is as follows:
- Enter Token: $1 per million
- cache input Token: $0.05 per million
Token: $5 per million
Standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. This allows developers to automatically process large volumes of work at lower cost, such as reading invoices received and entering relevant information into the accounting software.
From today, all Plus, Pro, Enterprise and Edu users can use GPT-6.1 Terra in ChatGPT Work and Codex. The developer can also call the model through Open AI API. Model ID is gpt-61-terra.
Also, in ChatGPT Work and Codex, the use of GPT-61 Terra would make your quota more effective. The same use of quota allows for more tasks than GPT-5.6 Terra; the number of tasks that can be accomplished is much more than GPT-6.1 Sol or GPT-6 Astra.
From visual understanding to practicalization
Open AI uses the Ages' Last Exam benchmark test to assess GPT-6.1 Terra's computer performance. The benchmark covers a wide range of professional tasks, such as the creation of a 3D model based on building drawings and the review of animated videos by shot to produce time-coded records for repair teams.
At the medium level of reasoning, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra, at about one third of the cost. Even when compared to GPT-6 Astra, its performance is not the same, but the price per Token is only one tenth.
At the highest level of reasoning, Terra’s Token efficiency is more than three times that of Claude Opus 5.5, and the price of each Token is only one quarter of the latter, so the cost of completing the same task can be reduced to about one ninth.
Terra is also operating much faster, with the completion of tasks more than three times as fast as GPT-6 Astra and GPT-6.1 Sol.
GPT-6.1 Terra performed better than GPT-5.6 Terra in all reasoning intensity settings on MMU-Pro. Its scores are even close to the front-line models of Open AI, GPT-6 Astra and GPT-6.1 Sol, but at a much lower cost.
OSWorld 2.1 Tests whether AI AI agent can operate desktop applications to complete multistep tasks, such as creating 3D models based on technical drawings. At the moderate level of reasoning, GPT-6.1 Terra has outperformed GPT-5.6 Terra ' s best, while the cost of each mission is much lower.
Significant improvement in factual accuracy and programming capacity
This year, Open AI invested a lot of energy in reducing the hallucinations and factual errors that the models generated.
In its internally designed prompt benchmark test, which is prone to error, 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.
At the highest level of reasoning, the error rate also declined by 18 percentage points to 4.3 per cent, which is very close to GPT-6 Astra and GPT-6.1 Sol.
The software engineering capacity of GPT-6.1 Terra has also improved significantly over GPT-5.6 Terra.
Terminal-Bench 4.0 is used to test the ability of programming AI agent to solve problems through command line, e.g. to repair a page where the layout will shift when the page is loaded. At the highest level of reasoning, GPT-6.1 Terra was able to solve 50.3 per cent of the task, more than double the 21.5 per cent mark published by GPT-5.6 Terra. GPT-6.1 Sol and GPT-6 Astra remain better options for the most complex programming tasks.
Terminal-Bench-Science 0.1 is used to test the ability of AI agent to perform scientific research tasks through the preparation and operation of codes, such as tracking cell splits in delayed photographic videos.
At the highest level of reasoning, GPT-6.1 Terra outperformed GPT-6.1 Sol with lower costs. Compared to Opus 5.5, the score gap is less than 5 percentage points, while the estimated cost is less than one fifth.
Security and alignment
While making powerful AI more efficient and user-friendly, we are continuously training, testing and improving safety protection mechanisms.
According to the Open AI Preparedness Framework (Preparedness Platform), GPT-6.1 Terra met the “critical” level capability threshold in the area of cybersecurity, while retaining the safety precautions developed for GPT-6 Astra.
GPT-6.1 Terra also inherited improvements in alignment of the GPT-6 series. According to the assessment, GPT-6.1 Terra is more likely than GPT-5.6 Terra to limit the operation to the extent specified by the user and to inform the user of its own capacity limitations and possible obstacles.
During the prompt injection test, GPT-6.1 Terra also demonstrated greater resistance to attacks that attempted to divert the model from the user ' s original mission through malicious instructions.
For model details, see model cards:
- Model card address: https://deploymentsafety.openai.com/gpt-6-1-terra/gpt-6-1-terra.pdf
- Blog address: https://openai.com/index/international-gpt-6-1-terra
Auto-fetched and translated by AI Atlas. All rights belong to the original author.