OpenAI 发布 GPT-6.1 Sol，称其性能接近 GPT-6 Astra、成本更低

周二，在 OpenAI 的 DevDay 活动上，该公司展示了 GPT-6.1 Sol，距离其发布 GPT-6 Sol仅过去一周。OpenAI 表示，这一新模型在 Agent 编程、电脑操作和专业工作方面的智能水平几乎与 GPT-6 Astra 相当，而标准输入和输出 token 的价格仅为后者的五分之一。

值得注意的是，该公司并未如外界原先预期的那样发布 GPT-6.1 Astra。本周，《华尔街日报》报道称，OpenAI 因研究人员在内部测试中提出的安全担忧而取消了该模型的发布。测试显示，该模型具有更高程度的欺骗行为，并倾向于在未征得用户许可的情况下继续执行任务。

OpenAI 表示，GPT-6.1 Sol 在复杂任务中的表现较前代 GPT-6 Sol 显著提升，涵盖编程与调试、理解文档以及执行多步骤工作流等场景。该公司称，在其中若干方面，这一模型的性能已接近 GPT-6 Astra。

OpenAI 还表示，新模型在面对困难提示词时提高了事实准确性。相较 GPT-6 Sol，其在这方面最大的提升似乎出现在低推理强度设置下：包含事实错误的回答占比从 11.4% 降至 7.7%。该公司称，在所有推理设置下，新模型的错误率与 GPT-6 Astra 的差距均不超过 1.9%。

此外，该公司称，GPT-6.1 Sol 会更坦率地说明自身局限，并且在遵循用户意图和安全约束方面更可靠。在高难度评测中，据称该模型在识别失效的搜索工具、遵守明确限制，以及避免在任务中产生未经授权的结果等方面，失败次数少于 GPT-6 Sol。OpenAI 称，其没有观察到该模型试图绕过自动化安全审核器的行为，这与 GPT-6 Astra 和 GPT-6 Sol 的表现一致。

GPT-6.1 Sol 自今日起面向 ChatGPT Work 和 Codex 中的所有 Plus、Pro、Business、Enterprise 与 Edu 用户开放。OpenAI 指出，该模型目前尚未在 Chat 中提供。

At OpenAI’s DevDay event on Tuesday, the company showed off GPT-6.1 Sol, a mere week after it launched GPT-6 Sol. OpenAI says the new model delivers nearly the same level of intelligence as GPT-6 Astra for agentic coding, computer use, and professional work, at one-fifth the standard input and output token prices.

Notably, the company is not launching GPT-6.1 Astra, as was originally expected. The Wall Street Journal reported this week that OpenAI scrapped the release over safety concerns raised by researchers during internal testing after the model showed higher levels of deception and a tendency to move forward with tasks without asking the user for permission.

OpenAI says GPT-6.1 Sol delivers significant improvements over its predecessor GPT-6 Sol across complex tasks, including programming and debugging, understanding documents, and executing multistep workflows. The company claims that on several of these fronts, the model approaches GPT-6 Astra’s performance.

OpenAI also says the new model improves factual accuracy when given difficult prompts. Its largest gain over GPT-6 Sol on this front appears at low reasoning effort, where the share of responses containing a factual error drops from 11.4% to 7.7%. Across all reasoning settings, the company says, the new model’s error rate stays within 1.9% of GPT-6 Astra.

Additionally, the company says GPT-6.1 Sol is more up front about its limitations and more reliable when it comes to honoring user intent and safety constraints. In challenging evaluations, it’s said to fail less often than GPT-6 Sol at flagging broken search tools, following explicit restrictions, and avoiding unauthorized outcomes during tasks. OpenAI claims it observed no attempts to circumvent the automated safety reviewer, consistent with GPT-6 Astra and GPT-6 Sol.

GPT-6.1 Sol is available starting today to all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. OpenAI notes that the model is not yet available in Chat.