刚刚，GPT-6.1 Terra 骨折价发布，Agent 时代的打工模型来了
OpenAI 又卷起了大模型的性价比。
就在刚刚，OpenAI 正式发布新一代模型 GPT-6.1 Terra，主打更快的响应速度、更强的视觉理解和电脑操作能力，以及更低的使用成本。
相比上一代 GPT-5.6 Terra，新模型的标准 API 价格直接降低 50%，但在复杂任务执行、事实准确率和编程能力上均有明显提升。尤其是在部分专业任务测试中，GPT-6.1 Terra 已经能够接近旗舰模型 GPT-6 Astra 的表现，Token 价格却只有后者的十分之一。
OpenAI 将其定位为一款面向高频工作场景的高效模型，希望让更多企业和开发者能够以更低的成本，将 AI Agent 真正应用到日常工作中。
性能直追 Sol，执行任务速度快 3 倍
GPT-6.1 Terra 最值得关注的变化，是视觉理解与电脑操作能力的进一步结合。
过去，AI 可以理解图片、分析文档，但要让它真正操作电脑、跨应用完成复杂任务，依然需要较高的模型能力和推理成本。
如今，GPT-6.1 Terra 已经能够处理从阅读技术图纸、填写表格，到编辑视频、操作桌面软件等一系列专业任务。
例如，用户可以让它根据建筑设计图纸创建 3D 模型，或者逐帧检查动画影片，为修复团队制作带有时间码的镜头记录。企业还可以利用它读取收到的发票，并将相关信息自动录入财务软件。
在衡量 AI Agent 完成专业工作能力的 Agents’ Last Exam 测试中，GPT-6.1 Terra 的表现相当突出。
在中等推理强度下，新模型的得分比 GPT-5.6 Terra 提高了 15 个百分点，单次任务成本却降低至约三分之一。
更值得关注的是，同样在这一测试中，GPT-6.1 Terra 已经能够达到 GPT-6 Astra 的水平，而标准 API 的 Token 单价仅为后者的十分之一。
与 Anthropic 的 Claude Opus 5.5 相比，在最高推理强度下，GPT-6.1 Terra 的 Token 使用效率超过其 3 倍，Token 单价只有其四分之一，最终完成同类任务的成本约低 9 倍。
速度方面，GPT-6.1 Terra 完成任务的速度也超过 GPT-6 Astra 和 GPT-6.1 Sol 的 3 倍。
有开发者晒出的测试结果显示，GPT-6.1 Terra 单次请求的生成速度可以达到每秒 450 Token，低推理强度下甚至有望达到每秒 600 Token，不过后者尚未得到验证。
第三方评测机构 Artificial Analysis 的智力指数榜单也给出了一个有意思的参照，GPT-6.1 Terra（Max）拿到 48 分，超过除 Kimi K3.1 Pro（55 分）之外的中国开放权重模型。
另外，在考察视觉推理能力的 MMMU-Pro 测试中，GPT-6.1 Terra 在所有推理强度下均超过上一代模型，并以明显更低的成本，取得接近 GPT-6 Astra 和 GPT-6.1 Sol 的成绩。
在专门测试 AI 操作桌面应用能力的 OSWorld 2.1 中，新模型仅使用中等推理强度，就已经超过 GPT-5.6 Terra 在最高推理强度下的成绩，同时显著降低了单次任务成本。
实际体验也印证了部分提升，DesignArena 让两代 Terra 重现日落时分的里斯本贝伦塔，6.1 版本生成了更精细的建筑与水面光影，虽然耗时接近上一代的 4 倍、Token 消耗增加 48%，单次成本却从 0.077 美元降至 0.047 美元。
而 atomic.chat 使用相同提示词让两款模型分别以 8 个 Agent 开发游戏，Terra 只用 6 分钟、花费 1.06 美元就完成任务，而 Claude Sonnet 5.5 花费约一小时和 18.10 美元，却交出了明显更好的游戏作品。
此外，从现有测试来看，GPT-6.1 Terra 在高难度任务的最终交付质量上，恐怕仍然比不上 Claude Opus 5.5，但考虑到它更低的价格和更快的执行速度，依然有不小的生存空间。毕竟，现实中有大量工作只需要 AI 把事情做完，并不值得每次都请出最贵的旗舰模型。
幻觉率大幅下降，编程成绩翻倍
除了视觉理解和电脑操作，GPT-6.1 Terra 在事实准确率与编程方面的进步同样明显。
OpenAI 表示，今年一直在重点解决模型产生幻觉和事实性错误的问题。
在其内部设计的高难度事实准确率测试中，GPT-5.6 Terra 在低推理强度下的错误率高达 53.6%，而 GPT-6.1 Terra 已将其降低至 16.7%，下降 36.9 个百分点。
当推理强度提升至最高档时，新模型的错误率进一步下降至 4.3%，接近 GPT-6 Astra 和 GPT-6.1 Sol 的水平。
对于需要处理合同、技术文档和企业内部数据的 AI Agent 来说，事实错误率的下降尤其重要。毕竟，模型能够完成任务只是基础，执行过程中能否持续保持准确，直接决定了用户能否放心将更多工作交给 AI。
编程能力方面，GPT-6.1 Terra 也迎来大幅提升。
在测试 AI Agent 命令行编程能力的 Terminal-Bench 4.0 中，GPT-6.1 Terra 在最高推理强度下取得了 50.3% 的任务解决率，上一代 GPT-5.6 Terra 的公开成绩仅为 21.5%。
换句话说，新模型的编程测试成绩已经达到上一代的两倍以上。
在科学研究编程测试 Terminal-Bench-Science 0.1 中，GPT-6.1 Terra 甚至超过了定位更高的 GPT-6.1 Sol。
该测试包含利用代码分析延时摄影视频中的细胞分裂等科学研究任务。OpenAI 表示，GPT-6.1 Terra 的得分距离 Claude Opus 5.5 不足 5 个百分点，估算成本却低于后者的五分之一。
当然，OpenAI 也明确表示，对于最复杂的软件工程和深度研究任务，GPT-6.1 Sol 和 GPT-6 Astra 依然是更合适的选择。
API 价格腰斩，正式进入 ChatGPT Work 和 Codex
相比性能提升，GPT-6.1 Terra 的价格调整可能对开发者更有吸引力。
根据 OpenAI 公布的信息，新模型的标准 API 定价如下。
- 输入价格为每百万 Token 1 美元
- 缓存输入价格为每百万 Token 0.05 美元
- 输出价格为每百万 Token 5 美元
相比 GPT-5.6 Terra 和 GPT-6.1 Sol，GPT-6.1 Terra 的标准 API 价格降低了 50%。如果与旗舰模型 GPT-6 Astra 相比，其输入和输出 Token 的价格均只有后者的十分之一。
在产品层面，GPT-6.1 Terra 已经面向 ChatGPT Plus、Pro、Enterprise 和 Edu 用户开放，可以在 ChatGPT Work 和 Codex 中使用。
OpenAI 表示，在相同的使用额度下，用户能够通过 GPT-6.1 Terra 完成比 GPT-5.6 Terra 更多的任务，相较 GPT-6.1 Sol 和 GPT-6 Astra，额度使用效率的提升则更加明显。
开发者也可以通过 OpenAI API 调用该模型，模型名称为 gpt-6.1-terra。
安全方面，OpenAI 表示，GPT-6.1 Terra 在其 Preparedness Framework 中达到网络安全能力的 Critical 风险等级，因此沿用了 GPT-6 Astra 等模型的相关安全防护措施。
与此同时，新模型继承了 GPT-6 系列的对齐改进，在遵守用户授权范围、主动说明自身能力限制，以及抵御恶意提示词注入方面，相比上一代有所提升。
回看 OpenAI 今年的产品更新，从旗舰模型 Astra 到强调综合能力的 Sol，再到如今主打高效执行的 Terra，产品之间的分工正在变得越来越明确。
尤其是在 AI Agent 加速进入真实工作场景的背景下，模型竞争也开始从单纯追求更高的 Benchmark 分数，延伸到完成同一项任务究竟需要花多少钱、多少时间，以及能否持续可靠地执行。
过去，企业想让 AI 自动处理大量办公任务，往往需要在模型能力、速度和成本之间反复权衡。
GPT-6.1 Terra 的出现，正在降低这种权衡的代价。
当一个能够看懂屏幕、操作软件、理解复杂文档的模型，已经具备接近旗舰级的能力，却只需要十分之一的 Token 成本，AI Agent 也就有了从少数高价值任务走向大规模日常应用的现实条件。
对于 OpenAI 来说，让越来越强的 AI 变得越来越便宜，或许与发布下一款最强模型同样重要。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Just now, GPT-6.1 Terra's fracture price publish, Agent's working model.
Open AI has another LLM value for money.
Just recently, Open AI, the official publish new-generation model GPT-6.1 Terra, which focuses on faster response speed, greater visual understanding and computer operation, and lower usage costs.
Compared to the previous GPT-5.6 Terra, the standard API price of the new model was reduced by 50 per cent directly, but there was a marked improvement in the performance of complex tasks, factual accuracy and programming capacity. In particular, in some of the professional mission tests, GPT-6.1 Terra has been able to approach the performance of the flagship model GPT-6 Astra, with only one tenth of Token ' s price.
Open AI has positioned it as an efficient model for a high-frequency (HF) work scenario, with the hope that more businesses and developers will be able to apply AI Agent to their daily work at a lower cost.
Sol, three times faster.
GPT-6.1 Terra ' s most noteworthy change is the further integration of visual understanding with computer-operated capabilities.
In the past, AI has been able to understand pictures and analyse documents, but it still requires higher modelling capabilities and reasoning costs to enable it to actually operate computers and cross-application complex tasks.
Today, GPT-6.1 Terra has been able to handle a range of professional tasks ranging from reading technical drawings, filling out forms, editing videos, operating desktop software, etc.
For example, the user can allow it to create a 3D model based on the architectural drawings, or to screen animated films on a frame-by-frame basis to produce shot records with time code for the repair team. It can also be used by businesses to read invoices received and automatically record relevant information in financial software.
GPT-6.1 Terra performed fairly well in the Agents ' Last Exam test, which measured the ability of AI Agent to complete professional work.
Under the medium-intensity of reasoning, the score of the new model increased by 15 percentage points to GPT-5.6 Terra, while the cost of a single mission was reduced to about one third.
Of even greater concern is the fact that, in this same test, GPT-6.1 Terra has been able to reach GPT-6 Astra, while the standard API Token unit price is only one tenth of the latter.
Compared to Claude Opus 5.5 in Anthropic, the GPT-6.1 Terra Token is more than three times more efficient than it is in the highest degree of reasoning, with only one quarter of the unit price of Token, and the cost of completing the same kind of task is about nine times lower.
In terms of speed, GPT-6.1 Terra also completed tasks three times faster than GPT-6 Astra and GPT-6.1 Sol.
The test results, which were tanned by the developers, showed that GPT-6.1 Terra single requests could be generated at a rate of 450 token per second, and that even 600 token per second was expected under low reasoning intensity, although the latter had not yet been validated.
The third-party rating agency, Artificial Analysis, also gave an interesting reference to GPT-61 Terra (Max), which received 48 points, exceeding the Chinese open weight model except for Kimi K3.1 Pro (55).
In addition, during the MMMU-Pro test, which examined the capability of visual reasoning, GPT-6.1 Terra exceeded the previous generation ' s model at all levels of reasoning and achieved near GPT-6 Astra and GPT-6.1 Sol at significantly lower cost.
In OSWorld 2.1, which is dedicated to testing AI ' s ability to operate desktop applications, the new model has exceeded GPT-5.6 Terra ' s achievement at the highest level of reasoning, while significantly reducing the cost of a single task.
The experience is also partly evidence of the increase, with two generations of Terra re-emerging Lisbon Belenta, version 6.1 producing a more sophisticated image of buildings and water, which, while taking time close to four times the previous generation, token consumption increased by 48 per cent, the single cost fell from 0.077 to 0.047.
While atomic.chat uses the same prompt to allow two models to develop the game in eight Agents, Terra performs the task in only six minutes at a cost of $1.06, while Claude Sonet 5.5, at a cost of about an hour and $18.10, handed over clearly better.
Moreover, according to the available tests, GPT-61 Terra still does not match Claude Opus 5.5 in terms of the quality of the final delivery of a difficult mission, but there is still considerable room for survival, given its lower price and faster implementation. After all, a lot of work in reality just needs AI to finish things, and it's not worth the most expensive flagship model ever.
The hallucinogenic rate has dropped and the programming has doubled.
Apart from visual understanding and computer operations, GPT-61 Terra advances in factual accuracy as well as programming.
Open AI states that this year has been focusing on the problem of illusions and factual errors in models.
The GPT-5.6 Terra error rate at low reasoning strength was as high as 53.6 per cent in its in-house design, while GPT-6.1 Terra has reduced it to 16.7 per cent, a decrease of 36.9 percentage points.
The error rate of the new model further decreased to 4.3 per cent, close to GPT-6 Astra and GPT-6.1 Sol, when the reasoning intensity rose to the highest level.
The decline in the incidence of factual errors is particularly important for AI Agent, which is required to process contracts, technical files and enterprise internal data. After all, the ability of the model to fulfil its mandate is only a basis, and the continued accuracy of the implementation process directly determines whether the user is confident that more work will be done by AI.
In terms of programming capacity, GPT-6.1 Terra has also seen a significant increase.
During the testing of AI Agent command line programming capability Terminal-Bench 4.0, GPT-6.1 Terra achieved a task resolution rate of 50.3 per cent at the highest level of reasoning and the previous generation of GPT-5.6 Terra achieved only 21.5 per cent in public.
In other words, the programming test results of the new model have been more than double those of the previous generation.
In the scientific research programming test Terminal-Bench-Science 0.1, GPT-61 Terra exceeded even the higher GPT-6.1 Sol.
The test included scientific research tasks such as cytology in delayed photographic videos using code analysis. Open AI states that the GPT-6.1 Terra score range Claude Opus 5.5 is less than 5 percentage points, while the estimated cost is less than one fifth of the latter.
Of course, Open AI has made it clear that GPT-61 Sol and GPT-6 Astra are still more appropriate options for the most complex software engineering and in-depth research missions.
API price cut, official entry to ChatGPT Work and Codex
GPT-61 Terra price adjustments may be more attractive to developers than performance improvements.
Based on information published by Open AI, the standard API pricing for the new model is as follows.
- The input price is $1 per million Token.
- cache input price per million Token 0.05
- Output price is $5 per million Token.
Compared to GPT-5.6 Terra and GPT-6.1 Sol, GPT-6.1 Terra standard API prices decreased by 50 per cent. If compared to the flagship model GPT-6 Astra, the input and output of Token is only one tenth of the price of the latter.
At the product level, GPT-6.1 Terra is already open to ChatGPT Plus, Pro, Enterprise and Edu users and can be used in ChatGPTWork and Codex.
Open AI states that under the same use of quota, users are able to perform more tasks than GPT-5.6 Terra with GPT-6 Sol and GPT-6 Astra, with quota increasing efficiency.
The developer can also call the model through Open AI API, which is called gpt-61-terra.
In terms of security, Open AI states that GPT-6.1 Terra has reached the Critical risk level of network security capability in its Preparedness Framework, and therefore follows relevant safety protection measures such as GPT-6 Astra.
At the same time, the new model inherited GPT-6 series alignment improvements, which were higher than the previous generation in terms of compliance with the scope of the user ' s mandate, self-described capacity constraints and resistance to malicious prompt injections.
Looking back at Open AI's product update this year, from the flagship model, Astra, to Sol, which emphasizes integrated capabilities, to Terra, which is now the main executive, the division of labour between products is becoming clearer.
Especially in the context of AI Agent's accelerated access to real work, model competition has also begun from the simple pursuit of higher Benchmark scores, extending to the extent of how much, how much time it takes to complete the same task, and whether it can be carried out reliably and continuously.
In the past, an enterprise has made it necessary for AI to deal automatically with a large number of office tasks, often with a trade-off between model capabilities, speed and cost.
The advent of GPT-6.1 Terra is reducing the cost of this trade-off.
When a model that can read the screen, operate the software, understand the complex documents, has the ability to approach the flag class, but only needs one tenth of the Token cost, and AI Agent has the realistic conditions to move from a few high-value tasks to a large-scale daily application.
For Open AI, getting stronger and stronger is becoming cheaper, perhaps as important as publish's next strongest model.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.