OpenAI 发布了 GPT-6.1 Terra，该公司称其将成本与性能的边界推向了前所未有的高度。其 API 价格仅为 GPT-5.6 Terra 的一半。
Terra 是 OpenAI 在 Anthropic 中端 Sonnet 系列模型中最具竞争力的产品，且价格减半。输入 token 价格为每百万个 1 美元，输出 token 为 5 美元，相比之下 Claude Sonnet 5.5 分别为 2 美元和 10 美元，GPT-5.6 Terra 分别为 2 美元和 12 美元。缓存输入的费率降至每百万 token 仅 0.05 美元，仅为此前费率的四分之一。
根据 OpenAI 的说法，该模型结合了快速的视觉推理能力、精确的计算机控制能力以及更自然的写作风格，旨在应对从起草提案到解读技术图纸再到填写表格等日常工作任务。GPT-6.1 Terra 现已在 ChatGPT Work 和 Codex 中面向 Plus、Pro、Enterprise 和 Edu 用户开放，同时也通过 API 提供。
分析师称成本效率达到新高
Artificial Analysis 将 GPT-6.1 Terra 称为“每美元智能”的新基准。在最大推理力度下，该模型在智能指数上比 GPT-5.6 Terra 高出 6 分，而每项任务的成本却不到三分之一（0.40 美元对比 1.40 美元）。在较低的努力水平下，提升幅度更大——高出 7 到 13 分——使该模型处于帕累托前沿（Pareto frontier），即成本和性能无法同时进一步提升的点。
在智能指数（上图）中，GPT-6.1 Terra 得分为 48，介于 GPT-6.1 Sol（52）和 GPT-5.6 Terra（42）之间。成本性能图表（下图）显示，在较低的努力水平下，该模型提供了当前最佳的每项任务智能与成本比率。| 图片来源：Artificial Analysis
在新的 Coding Agent Index v2.0 中，Terra 得分为 48，低于 GPT-6.1 Sol 的 54 分，但高于 GPT-6 Luna 的 31 分 17 分。幻觉现象的减少更为显著。在 AA-Omniscience 私有基准测试中，当缺乏答案时，该模型选择猜测的比例仅为 38%，而 GPT-5.6 Terra 为 88%。准确率也提升了两个百分点。
计算机控制速度比最大规模模型快三倍
OpenAI 将 GPT-6.1 Terra 定位为在自动化软件控制方面特别强大的模型。在 Agents' Last Exam 基准测试中，它以十分之一的 token 价格达到了与 GPT-6 Astra 相当的性能。与最大努力下的 Claude Opus 5.5 相比，Terra 的 token 效率高出三倍以上，且 token 价格仅为四分之一。这意味着每项任务的成本大约降低了九倍。公司还表示，它的完成速度也比 GPT-6 Astra 和 GPT-6.1 Sol 快三倍以上。
在用于视觉推理的 MMMU-Pro 测试中，GPT-6.1 Terra 在所有推理层级上都优于其前身，并几乎追平了 GPT-6 Astra 和 GPT-6.1 Sol。在桌面代理基准测试 OSWorld 2.1 中，它在中等努力下以极低的成本击败了 GPT-5.6 Terra 的最佳结果。
Terra 在编码方面也显示出显著的改进。在 Terminal-Bench 4.0 中，GPT-6.1 Terra 在最大推理力度下解决了 50.3% 的任务，而 GPT-5.6 Terra 仅为 21.5%。据 Artificial Analysis 称，这 20 个百分点的提升使该模型跻身与 GPT-6.1 Sol 相同的梯队。在科学编码基准测试 Terminal-Bench-Science 0.1 中，Terra 以更低的成本超越了 GPT-6.1 Sol，并以不到五分之一的预估成本接近 Claude Opus 5.5 的水平。然而，对于最具挑战性的编码任务，OpenAI 仍推荐 GPT-6.1 Sol 和 GPT-6 Astra。
Terra 落入 OpenAI 的“关键”安全阈值之下
GPT-6.1 Terra 符合 OpenAI 根据其准备框架定义的“关键”网络安全阈值，并继承了 GPT-6 Astra 的安全保障措施。与其前身相比，该模型更有可能坚守其分配的范围并向用户披露限制条件。
根据 系统卡片，GPT-6.1 Terra 在面对针对 指令层级 的直接攻击时，表现与 GPT-6 Astra 和 GPT-6.1 Sol 大致相当，这些模型在此类测试中的得分均为 99.99%。对于间接提示注入攻击（恶意指令隐藏在第三方内容中），Terra 的防御率为 98.8%，高于 GPT-5.6 Terra 的 96.2%。
无炒作 AI 新闻 – 由人工策划
订阅 THE DECODER，享受无广告阅读、每周 AI 通讯、每年六次的独家“AI Radar”前沿报告、完整档案访问权限以及评论板块访问权限。
立即订阅
查看原文
OpenAI has released GPT-6.1 Terra, a model it says pushes the cost-to-performance boundary further than any of its predecessors. API prices are half those of GPT-5.6 Terra.
Terra is OpenAI's closest competitor to Anthropic's mid-tier Sonnet series at half the price. Input tokens cost $1 per million, output tokens $5, compared to $2 and $10 for Claude Sonnet 5.5 and $2 and $12 for GPT-5.6 Terra. Cached inputs drop to just $0.05 per million tokens, a quarter of the previous rate.
According to OpenAI, the model combines fast visual reasoning with precise computer control and a more natural writing style, targeting everyday work tasks from drafting proposals to interpreting technical drawings to filling out forms. GPT-6.1 Terra is available now for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, as well as through the API.
Analysts say cost efficiency has hit a new high
Artificial Analysis calls GPT-6.1 Terra a new benchmark for intelligence per dollar. At maximum reasoning effort, the model gains 6 points on the Intelligence Index over GPT-5.6 Terra while costing less than a third per task ($0.40 versus $1.40). At lower effort levels the gains are even larger—7 to 13 points—placing the model on the Pareto frontier, the point where cost and performance can no longer both improve simultaneously.
On the Intelligence Index (top), GPT-6.1 Terra scores 48, between GPT-6.1 Sol (52) and GPT-5.6 Terra (42). The cost-performance chart (bottom) shows that at lower effort levels, the model delivers the best current ratio of intelligence to cost per task. | Image: Artificial Analysis
In the new Coding Agent Index v2.0, Terra scores 48, six points below GPT-6.1 Sol (54) and 17 points above GPT-6 Luna (31). The hallucination drop is even bigger. On the private AA-Omniscience benchmark, the model resorts to guessing only 38 percent of the time when it lacks the answer, down from 88 percent for GPT-5.6 Terra. Accuracy also rose by two points.
Computer control three times faster than the largest models
OpenAI positions GPT-6.1 Terra as particularly strong at automated software control. On the Agents' Last Exam benchmark, it matches GPT-6 Astra's performance at one-tenth the token price. Compared to Claude Opus 5.5 at maximum effort, Terra is more than three times as token-efficient at a quarter of the token price. That translates to roughly nine times lower cost per task. It also finishes more than three times faster than both GPT-6 Astra and GPT-6.1 Sol, the company says.
On MMMU-Pro for visual reasoning, GPT-6.1 Terra outperforms its predecessor at every reasoning level and nearly matches GPT-6 Astra and GPT-6.1 Sol. On the desktop agent benchmark OSWorld 2.1, it beats GPT-5.6 Terra's best result at medium effort for a fraction of the cost.
Terra also shows a sharp improvement in coding. On Terminal-Bench 4.0, GPT-6.1 Terra solves 50.3 percent of tasks at maximum reasoning effort, up from 21.5 percent for GPT-5.6 Terra. According to Artificial Analysis, that 20-point jump puts the model in the same tier as GPT-6.1 Sol. On the scientific coding benchmark Terminal-Bench-Science 0.1, Terra outperforms GPT-6.1 Sol at lower cost and comes within five points of Claude Opus 5.5 at less than a fifth of the estimated cost. For the most demanding coding tasks, however, OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra.
Terra falls under OpenAI's "critical" safety threshold
GPT-6.1 Terra meets OpenAI's "Critical" cybersecurity threshold as defined by the company's Preparedness Framework and inherits the safeguards of GPT-6 Astra. Compared to its predecessor, the model is more likely to stay within its assigned boundaries and disclose restrictions to users.
According to the system card, GPT-6.1 Terra holds up about as well as GPT-6 Astra and GPT-6.1 Sol against direct attacks on the instruction hierarchy, with those models hitting 99.99 percent on such tests. For indirect prompt injection attacks, where malicious instructions are buried in third-party content, Terra's defense rate sits at 98.8 percent, up from 96.2 percent with GPT-5.6 Terra.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
原文版权归 The Decoder 及原作者所有，本站转载仅供阅读。如需删除，请联系 support@aisa.one。
Studio