OpenAI 发布 GPT-6.1 Terra：价格腰斩、任务速度提升超 3 倍
当大模型竞争从回答问题转向实际完成工作，成本和效率正在成为与模型能力同样重要的指标。
刚刚，OpenAI 正式发布 GPT-6.1 Terra。作为 GPT-5.6 Terra 的升级版本，新模型重点提升了视觉推理、计算机操作、事实准确性及编程能力，同时进一步降低了 API 调用成本。
与过去主要强调推理能力和基准测试成绩的模型更新不同，此次 GPT-6.1 Terra 更侧重日常工作场景，包括撰写商业提案、理解技术图纸、填写表格、操作软件，以及辅助完成视频编辑等任务。
OpenAI 表示，新模型在速度、能力与成本之间取得了新的平衡。其标准 API 价格较 GPT-5.6 Terra 和 GPT-6.1 Sol 下降 50%，而在部分专业任务测试中，Terra 的表现已接近旗舰模型 GPT-6 Astra。
这意味着，开发者不一定需要调用最昂贵的模型，就有机会完成过去需要高性能模型处理的部分复杂任务。
老规矩，先来看看新模型各项基准测试表现如何。
更值得关注的是，OpenAI 称，在该测试设定下，GPT-6.1 Terra 的表现已经与 GPT-6 Astra 大致相当，但其每 Token 价格仅为后者的十分之一。
与竞争对手相比，官方数据显示，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 Token 价格则约为后者的四分之一。综合计算，其单任务成本降低约 9 倍。
此外，Terra 完成测试任务的速度比 GPT-6 Astra 和 GPT-6.1 Sol 均快三倍以上。
在另外两项测试中，Terra 同样取得进展。视觉理解基准 MMMU-Pro 显示，新模型在各档推理强度下均优于上一代。
而测试桌面应用多步骤操作的 OSWorld 2.1 则显示，Terra 在中等推理强度下，已经超过 GPT-5.6 Terra 在最高成绩设定下的表现，同时降低了任务成本。
除了计算机操作，GPT-6.1 Terra 的另一项明显改进是事实准确性。
OpenAI 表示，过去一年，公司持续投入资源降低模型幻觉及事实错误。在其内部设计的、旨在诱发模型错误的提示词测试中，GPT-6.1 Terra 的错误率明显下降。
具体而言，在低推理强度下，错误率从上一代的 53.6% 降至 16.7%，下降 36.9 个百分点，相对降幅约为 69%。
当推理强度提高至最高档时，Terra 的错误率进一步降至 4.3%，较上一代下降约 18 个百分点，接近 GPT-6 Astra 和 GPT-6.1 Sol 的水平。
这类改进对于实际办公场景具有重要意义。无论是提取合同条款、处理业务数据，还是依据技术文档执行操作，模型如果频繁产生错误，即便调用价格再低，也可能增加人工核查和返工成本。
编程方面，Terra 的升级同样较为明显。
在 Terminal-Bench 4.0 测试中，GPT-6.1 Terra 在最高推理强度下完成了 50.3% 的任务，而 GPT-5.6 Terra 此前公布的成绩为 21.5%。Terminal-Bench 4.0 主要评估编程智能体通过命令行解决实际软件工程问题的能力，例如修复网页加载过程中发生的布局偏移。
这意味着，新模型的任务解决率提高了 28.8 个百分点，达到上一代的 2.34 倍左右。
此外，在科学编程任务基准 Terminal-Bench-Science 0.1 中，Terra 在最高推理强度下的表现超过 GPT-6.1 Sol，并在显著更低的成本下，将成绩差距缩小至 Claude Opus 5.5 的 5 个百分点以内。
但 OpenAI 也明确表示，对于最复杂的软件工程任务，GPT-6.1 Sol 和 GPT-6 Astra 仍然是更合适的选择。
API 价格下降 50%，百万输入 Token 仅需 1 美元
定价是此次 GPT-6.1 Terra 更新的重要看点。
根据 OpenAI 公布的标准 API 价格，新模型每百万输入 Token 收费 1 美元，每百万输出 Token 收费 5 美元，缓存输入则为每百万 Token 0.05 美元。
单说 GPT-6.1 Terra 的价格可能不够直观，下面是 OpenAI 几款最新模型价格的对比：
按照这一价格，GPT-6.1 Terra 的标准输入和输出 Token 单价均比 GPT-6.1 Sol 低 50%，仅为 GPT-6 Astra 的十分之一。
但 Token 单价并不等于最终任务成本。实际执行一次任务需要花费多少，还取决于模型消耗的推理 Token 数量、任务步骤、工具调用次数，以及失败后的重试成本。
因此，相比单纯降价，Terra 此次更值得关注的是其在部分评测中同时实现了能力提升、Token 使用效率改善和任务执行速度提高。
在产品使用方面，OpenAI 宣布 GPT-6.1 Terra 已向 ChatGPT Work 和 Codex 中的 Plus、Pro、Enterprise 及 Edu 用户开放。开发者也可以通过 API 调用 gpt-6.1-terra。
OpenAI 表示，相同的使用配额在 GPT-6.1 Terra 上能够覆盖比 GPT-5.6 Terra 更多的任务，也明显多于 GPT-6.1 Sol 或 GPT-6 Astra。
根据官方技术文档，GPT-6.1 Terra 支持约 105 万 Token 上下文窗口，最大输出长度为 12.8 万 Token，支持文本和图像输入。但这里值得注意的是，其 API 本身并不提供原生视频输入或输出，因此视频相关工作主要依赖与外部工具及应用的协同操作。
模型发多了，开发者也选择困难
GPT-6.1 Terra 发布后，x、Reddit 等平台已经出现了不少讨论。
和以往新模型发布时集中讨论“性能是否屠榜”不同，这次开发者的关注点更多落在三个问题上：价格是否真的更便宜、速度是否足够快，以及新模型是否值得替换现有工作流。
有意思的是，社区也出现了两种截然不同的情绪：一部分网友对性能与成本的改善表示兴奋，另一部分则对越来越密集的新模型发布感到麻木，甚至拿 OpenAI 的模型命名体系开起了玩笑。
在 Reddit 社区，一位名为 sillybluejayway 的网友认为，新模型实际上已经接近 GPT-6.0 Astra 的能力水平，但运行成本仅为后者的十分之一。在他看来，Terra 或许没有进一步突破最前沿模型的能力上限，但如果能够以如此低的成本提供接近上一代旗舰模型的性能，本身就是相当可观的进步。
这位网友还提出了一个更值得关注的观点：真正推动 AI 广泛应用的，未必只是最强模型不断刷新性能纪录，而是让已经具备较强能力的模型变得足够便宜。
他认为，无论是 AI 对就业岗位的替代，还是具身智能的发展，都与这类成本下降密切相关。
不过，也有网友对这种判断提出质疑。
网友 AzorAhai1TK 表示，自己仍然不够信任基准测试，尤其不认为仅凭测试成绩，就能断言一个规模小得多的模型已经全面追平大型模型。他指出，较小模型可能缺少一些额外的世界知识，但这些知识在特定任务中可能十分关键。
网友 NeedleworkerNarrow56 表示，自己对 GPT-6.1 Sol 进行了大量测试，发现其在编程工作中与 GPT-6 Astra 的差距已经很小，但在视觉和文本等通用推理任务中，两者仍然存在明显差距。
“在真实世界中运行的智能体，并不总是需要最前沿的顶级能力，但需要在成本与智能水平之间取得良好平衡。”另一位网友 Material-Watercress1 评论道。
这也点出了 Terra 的产品定位：对于高频调用模型的 Agent 应用来说，每次请求是否都需要使用最强模型，本身就是一个值得重新考虑的问题。
但当 OpenAI 不断推出不同系列和版本的模型时，开发者也遇到了一个新麻烦：究竟应该选哪个？
网友 NotAFishEnt 提出了一个颇有代表性的问题：面对 GPT-5.6、GPT-6.0、GPT-6.1，以及 Luna、Sol、Terra 等不同系列，再加上从 Light 到 Ultra 的多档推理强度，到底怎样才能找到最适合自己工作的组合？
他表示，自己日常需要处理复杂的编程任务，真正想弄清楚的是，怎样才能让花出去的每一美元获得最大的回报。
这一问题引起其他开发者共鸣。有网友建议，只能根据实际使用经验不断试错，例如先从 Medium 推理强度开始，如果无法完成任务，再逐步提高档位。
也有人认为，可以参考 Artificial Analysis 的性能与成本帕累托前沿，筛选那些在价格和能力之间具有竞争力的模型。
随着模型版本越来越多，开发者似乎又多出了一项工作：不仅要完成开发任务，还要持续评估哪款 AI 模型最划算，想想也是头大！
有意思的是，就在 GPT-6.1 Terra 发布的同一天，OpenAI 的老对手 Anthropic 刚传来了一则重磅消息。
当地时间 10 月 26 日，Anthropic 正式向美国证券交易委员会（SEC）公开提交 S-1 注册文件，启动 IPO 的公开申报程序。按照文件披露的信息，Anthropic 计划登陆纳斯达克，股票代码为“ANT”，但具体发行规模及发行价格区间尚未确定。
几乎在同一天，OpenAI 就发布新模型。一边是 Anthropic 向公开资本市场迈出重要一步，一边是 OpenAI 继续加快产品迭代。两家公司虽然动作不同，却都在回答同一个问题：如何让大模型能力真正转化为可持续的商业价值？
事实上，OpenAI 与 Anthropic 的竞争早已不局限于模型排行榜。尤其在 AI 编程和企业级 Agent 领域，两家公司不仅要争夺开发者，还需要证明自己的模型能够以更低成本完成更多实际工作。
这也解释了为什么 OpenAI 近期的新模型越来越强调成本效率。对于大规模部署 Agent 的企业而言，模型能力固然重要，但每项任务究竟花费多少钱、执行多长时间、能否稳定完成，正在成为同样关键的选型指标。
更微妙的是，当 Anthropic 开始向公开市场展示其商业化能力和财务状况时，OpenAI 也在通过密集的产品更新，展示模型能力与运行效率的持续进步。可见，大模型的竞争，正在从技术排行榜延伸到一场更加复杂的商业较量。
参考链接：
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。
Open AI publish GPT-6.1 Terra: Price cut, task speed up 3 times
As LLM competes from answering questions to actually completing work, costs and efficiency are becoming equally important indicators of modelling capacity.
Just, Open AI official publish GPT-6.1 Terra. As an upgraded version of GPT-5.6 Terra, the new model focuses on enhancing visual reasoning, computer operations, factual accuracy and programming capabilities, while further reducing API call costs.
This time GPT-6.1 Terra focuses more on day-to-day work scenarios, including the preparation of business proposals, understanding of technical drawings, filling in forms, operating software, and supporting video editing, as opposed to previous model updates, which focused mainly on reasoning capabilities and benchmarking performance.
Open AI states that the new model strikes a new balance between speed, capacity and cost. Its standard API prices are 50 per cent lower than GPT-5.6 Terra and GPT-6.1 Sol, while in some of the professional mission tests Terra performance is close to the flagship GPT-6 Astra model.
This means that developers do not necessarily need to call on the most expensive models and have the opportunity to perform some of the complex tasks that used to require high performance models.
The old rule, let's see how the new model's benchmarks are tested.
Of even greater concern is the fact that, according to Open AI, under the test, GPT-6.1 Terra is already performing roughly the same way as GPT-6 Astra, but its price per Token is only one tenth of the latter.
In comparison to competitors, official data show that, at the highest level of reasoning, Terra Token was more than three times more efficient than Claude Opus 5.5, with a price of about one quarter per Token. The combined calculation shows a reduction of about nine times the cost of the single job.
In addition, Terra completes the test task more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
Terra also made progress in two other tests. The visual understanding baseline MMMU-Pro shows that the new model is better than the previous generation at the level of reasoning in the various tranches.
OSWorld 2.1, which tested the desktop application for multi-step operations, showed that Terra had exceeded GPT-5.6 Terra's performance at the highest achievement level, while reducing the cost of the mission, at the medium reasoning strength.
In addition to computer operations, another significant improvement in GPT-61 Terra is factual accuracy.
Open AI states that over the past year, companies have consistently invested resources to reduce model hallucinations and factual errors. In its internal prompt test designed to induce model errors, GPT-6.1 Terra showed a significant decrease in error rates.
Specifically, under the low degree of reasoning, the error rate fell from 53.6 per cent to 16.7 per cent in the previous generation, a decrease of 36.9 percentage points, or a relative decrease of approximately 69 per cent.
When the intensity of the reasoning rose to the highest level, Terra's error rate was further reduced to 4.3 per cent, which was about 18 percentage points lower than the previous generation, close to GPT-6 Astra and GPT-6.1 Sol levels.
Such improvements are important for the actual office landscape. Whether it is to extract contract terms, process business data, or perform operations on the basis of technical files, the model may increase the cost of manual verification and back-to-work if it is frequently subject to errors, even if the call price is lower.
In terms of programming, the upgrading of Terra is equally evident.
In the Terminal-Bench 4.0, GPT-6.1 Terra performed 50.3 per cent of its tasks at the highest level of reasoning, while GPT-5.6 Terra had previously published an achievement of 21.5 per cent. Terminal-Bench 4.0 assesses mainly the ability of the AI agent programming to solve practical software engineering problems through command line, such as layout deviations during the restoration of web-based loading.
This means that the task resolution rate of the new model has increased by 28.8 percentage points to about 2.34 times that of the previous generation.
Furthermore, in the scientific programming mission benchmark Terminal-Bench-Science 0.1, Terra performed more than GPT-6.1 Sol at the highest level of reasoning and reduced the achievement gap to within 5 percentage points of Claude Opus 5.5 at significantly lower cost.
Open AI, however, also makes it clear that GPT-6.1 Sol and GPT-6 Astra remain the more appropriate options for the most complex software engineering tasks.
API prices are down 50%. Token takes only $1
Pricing is an important point of view for this GPT-61 Terra update.
According to the standard API price published by Open AI, the new model will charge Token $1 per million, Token $5 per million, and cache $0.05 per million.
The price of GPT-61 Terra alone may not be intuitive, and the following is a comparison of the most recent model prices of Open AI:
At this price, both GPT-61 Terra standard input and output Token unit prices were 50 per cent lower than GPT-61 Sol, only one tenth of GPT-6 Astra.
But the unit price of Token is not equal to the final mission cost. The actual cost of a mission also depends on the number of Tokens consumed by the model, the mission steps, the number of tool calls, and the cost of retrys after failure.
As a result, Terra deserves more attention this time than simply reduced prices, because of its capacity enhancement, token efficiency improvement and speed of mission implementation in part of the assessment.
In terms of product use, Open AI announced that GPT-6.1 Terra has been opened to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Developer can also call gpt-61-terra via API.
Open AI states that the same use quota can cover more tasks than GPT-5.6 Terra and significantly more than GPT-6.1 Sol or GPT-6 Astra.
According to official technical documentation, GPT-6.1 Terra supports approximately 1.05 million Token context window, with a maximum output length of 128,000 Token, and supports text and image input. It is noteworthy, however, that its API does not itself provide raw video input or output, so video-related work relies mainly on synergetic work with external tools and applications.
There's a lot of models and it's hard for developers to choose.
After GPT-6.1 Terra publish, there has been much discussion on platforms such as x and Reddit.
Unlike the previous new model publish, which focused on “Performance Scattered or Not”, the developmenter's focus was more on three issues: whether prices were really cheaper, whether they were fast enough, and whether the new model was worth replacing the existing workflow.
Interestingly, there are two distinct emotions in the community: Some netizens are excited about improvements in performance and costs, while others are numb to the increasingly intensive new model publish, even making fun of Open AI’s model naming system.
In the Reddit community, a netizens named Sillybluejayway believe that the new model is actually close to GPT-6.0 Astra's capacity level, but running costs are only one tenth of the latter. In his view, Terra may not have reached the upper limit of the ability to break the front-line model further, but it would be a considerable advance if performance close to the previous generation of flagship models could be provided at such a low cost.
The online user also raised a more interesting point: what really drives AI's wide application is not necessarily just that the strongest models keep updating their performance records, but that they make models that already have greater capacity cheap enough.
In his view, both AI ' s replacement of jobs and smart development were closely related to such cost reductions.
However, there are also netizens who question this judgement.
Netizen AzorAhai1TK states that he still does not trust the benchmark test, especially since he does not believe that a much smaller model can be assumed to have fully paced the large model. He noted that smaller models might lack some additional world knowledge, but that such knowledge could be critical in a given task.
Netleworker Narrow56 stated that he had conducted extensive tests of GPT-6.1 Sol and found that the gap between programming and GPT-6 Astra was already small, but that there was still a clear gap between the two in generic reasoning tasks such as visual and text.
“AI agent, which operates in the real world, does not always require a top-level capability, but a good balance between cost and intelligence. Another netizens comment.
This also points to Terra's product positioning: For the Agent application of the HF call model, whether the strongest model is required for each request is in itself a matter of reconsideration.
But when Open AI continues to produce different series and versions of models, developers also encounter a new problem: which should be chosen?
Netizen NotAFishEnt raises a very representative question: How can we find the most appropriate combination of jobs in the face of different series, such as GPT-5.6, GPT-6.0, GPT-6.1, and Luna, Sol, Terra, and so on, together with the multilayered reasoning from Light to Ultra?
He stated that he had to deal with complex programming tasks on a daily basis and that what he really wanted to know was how to maximize the return on each dollar spent.
This issue resonates with other developers. It has been suggested by netizens that mistakes can be made only on the basis of practical experience, starting with the intensity of Medium reasoning and, if it is not possible to complete the task, gradually increasing the slot.
It was also suggested that reference could be made to the performance and cost front of Artificial Analysis to screen models that were competitive between prices and capabilities.
With more models coming up, the developers seem to have done more: not only do development tasks, but also continuously assess which AI models are the most cost-effective, and also the biggest!
Interestingly, just on the same day as GPT-61 Terra publish, Open AI's old opponent, Anthropic, just sent a heavy message.
On October 26th, local time, Anthropic officially submitted S-1 registration documents to the United States Securities and Exchange Commission (SEC) and started the IPO public filing process. According to the information disclosed in the document, Anthropic plans to land in NASDAQ, with the stock code ANT, but the exact size of the issue and the range of the price of issuance have not been determined.
Almost the same day, OpenAi was on publish's new model. On the one hand, Anthropic took an important step towards open capital markets, and on the other hand, Open AI continued to accelerate product overlay. Both companies, although acting differently, answered the same question: How can LLM be truly translated into sustainable commercial values?
In fact, the competition between Open AI and Anthropic is no longer limited to model rankings. In particular, in the areas of AI programming and enterprise-level Agent, both companies need not only to compete for developers but also to prove that their models can do more practical work at lower cost.
This explains why Open AI's recent new model is increasingly emphasizing cost efficiency. Modelling capabilities are important for the large-scale deployment of Agent, but the amount of money spent on each task, the duration of its implementation and its sustainability are becoming equally critical selectional indicators.
Even more subtle, when Anthropic began to demonstrate to the open market its commercial capacity and financial position, Open AI was also demonstrating continuous progress in modelling capacity and operating efficiency through intensive product upgrading. As can be seen, LLM competition is extending from a technology ranking to a more complex business competition.
References:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.