- 研究
- 产品
- 企业
- 开发者
- 公司
- 基金会（在新窗口中打开）
- 研究
- 产品
- 企业
- 开发者
- 公司
- 基金会（在新窗口中打开）
OpenAI
0 0 . 0 . . 0 . . .
. 0 . 0 0 0 0 . 0
. . . 0
. . .. 0 0 .
. 0 . .
0 0 .
. . . . .. . . . . . 0
0 . . . . 0 . 0
. . . . 0 . . . . .. .. . . . . 0 . . . . 0 . 0 0 . . . . . . 0 . 0 0 . . . 0 . . . . . 0 . . . . 0 . . . 0 0 0 . . . .. . . . . 0 0. 0 0 . . . 0 . .0 . 00 . 0 . . 0 . . . . . . . 0 . . . . . . . 0 0 0 . . . 0 . . . 0 0 . . . . . . 0 . . . . . . .. . 0 .. . . . 0 . .0 0 . . . . .. 0 . . . 0 0 . 0 . 0 . .. . . . .. . . . . . . . . 0 . 0 . 0 . .0. . . . . 0 . . 0 . . . . .. . . . . . .. . . . 0 . . 0 . . . 0 . . .. . . . 0 . . . . .
655655555555
55555666655555.665
555556565.55555655555665
566656655555.565665556565.65
555.55555655556565.56656.5555556
5555.555555.55565655555.5665555655
56.6.565.5.66..55655665.55555555.556
56.566555.556556556.665556556565665.5.55
55555.55555655555.55565.565.66555555.5.5
65.555566555...565555..5.555.565555.56.655
556.5555656556.56555.55556556555555565665565
5556.55555555555566555.5566.5555.556656.655556
6.555555655556565565.55.5555655555555555.66555
.66555.5.65555555655566555555.665556.5555.565566
5555.655555655.65.5.5565.55.655..55656.555555555
5665.55555565656.55556.5..6656565555656555.65555
56.55655566555566565...5.5556.56.5655.555555655555
6555555555556555555555555.555565565556.55.55565556
5555.65555556565555556555.56665556566556.5556.5655
.5665.5.55555555565556566556565655.6655556.5555555
56656.565565555.5556.555555655555.5.55655.65.66.56
56.55.655555555555555555.565..55666555555555.55555
5555.556555665.5555555.555655555.5556565.555566555
.665565656566.5566655556565655656.5..55.6.65556655
55.565.55666555555566.5555555555566..55555556..6
55.555556655.56566.6565655.555.555655.5655555555
665555555.555556665.6566555565565.55565565565556
655.55555555555.5.55.65655566.5566565.55555.55
56555565.556.5666666555.65665555565556655.5565
5.555656555555555555655..5.5.655555655565555
56655665556565655555.665555556555565566655
.6655665.665555.55655655565665565555.555
5555555655555565565555555.5.55.66656555.
5565555556665.5.6555665556.665565555
6.66566555555..5555555555555655555
556656555555.55555555555555.5556
56..5556665655.5565555555556
.55555655655..6665555655
6565656.5555555555
565555555665 565 5
6 5
5 5 5 5
6 5 5 5
. 5 5 5 6 5
55 5 6 55 5 55
6 . 5
5 5 5 5 5 5 6 565 5
5 . 55 5 5
5 5 . 5 . 5 56 555 6
6 6 5 .5 65 5 6
5 6 5 5 5 5 5 .6
5 65 5.55 5 .
66 5 55 65 6 .
5 5 5 5.6
. 6 55 . 6 6 5
5 5 5 56 5 5
6 5 55 56 5 65 5
5 5 5 .56 5 6 5 6 6 5 5
5 6 5 5
6 5 . 5 5 5 5
5 5 55 55 5
5 5 6 5 6 . 5
. 5 6 5 5 . 5
5. . 6 5 . 5 . 5
.5 .6 5 5 5 5
5 6 5 6 5 55 6 5 5
5 5 5 5
5 6 6 6 6 5 6
5 55 . 5 5
6 55 5 5 5 5
. 6 5 5 5
55 5 5 5 5 . 5 5
5 5 65 5
. 5 555
55 5 55 5
. 5 5 5 5
. 5 ..6 55 5
5 55 555
6 5 5 5
55
5 6 65 5
5.
5 6 5 5
5 5 6
.5 6 5 5
. 5 5 5 6 66
55 5 6 6 .
6 5 5 6 5 5 5 5
5 56 6 6 5 5 5
5 5 5 6.55 5 66 5 5
5 6 55 5 5 55
. . 5 . 6 5 5 5
5 6 55 . 5 5
5 65 5 5 6
6 .. . 55 5 5
55 5 5 5 6 5 5 56
5 65 5 5
6 5 5 56 . 55 6 5
6 5 55 5 . . 66
6 . 5 5 5 5
5 5 5 5 55 5 5
6 6 . 5 5 5 5 6 6
5 6 5 5 5 5 5 5 . .
5 5 5 5 6 .5 5 5.5 5
6 6 . 55 6
6 . 55 5 6 5
6 5 5 6 6 5 65 5 6
6 5 5 6 6 5
6 5 6 5 5
6 566 .5 5 6 65 .
5 55 65 6 5
5 55 . 5 6 55 6 6 5 5
5 55
565 55 5 5
5 5 5 5 5
6
6 6 5
5 5
5 6 .6
5 .5 5 5 5 5
5
5 55 5 6 5.
5 6 5 55
5 5 5 6
5 6 5
5 . 5
66 5 . 6 5
. 5 . 5 6 5 5 65
. 55 5 6 5 5 6 5
6 55
5 55 5 5 . 5
5 5.6 . 5 5 555
6 5 56 65 . 5 55 5 5 5
6 5 6 . . 5 5 6 . 5
55 5 55 5 5 . 55 6
5 5 6 5
. .5 5 55 55 5 6 5 6 55
5 5 5 6 5 . 5 . 6
5
.5 .55
5 5 . 5 55 . 5 6 5 6
6 5 5 6 .
6 5 5 6 555 6
55 6 5 5 5
55 5.55 6 5 6
5 55 5 .6 6
5 . 6 5
65 5 55 . 5 5 555
6 56 6 5 5 6665
6 6 55
6 5 . 6
55 . 5 6 5
6 6 5 5 5 5
6 . 5 5
5 6
55 5 6 65
5 5
5 5 6 5
55 6
5 6
66 5 5 5 6 6 5.
55 5 5 6. 5 5 6
5 5 6
6 6 5 55 6 5 5
5 . 5 5 .
. 5 6 . 6 5
5 5 . .5 . 5
55 5 5 5 5
5 5 6 5 5
6 5 5 5 5 65 5 5 6 5
. 6 6 5 555 5 66
5 5 5 . . 5
6 5 .5 5 5
6.5 6 5 5 6 5 65
5 55 55 5 5 556 5
6 556 55 5 6 5 5 6
5 5 6 5 6 5 5 5
56 5 6 5 5 5 5 6 5
5 . 55 5 5 5 655 5
5 5 6 5 5.5 5
5 6 5 5 6 6 5 6
5 5 6 5 555 6 5 6
5 6 5 5 5
5 5 6 5 5 6
5 5 5 6 5 5 .5
5 6 5 5 5 5 5 5
. 5 5 5 6 5
55 5 6 5 5
. 5 5 5 5
5 5 56 55 5 6 5
5 6 6 5
6 55 . 5 5 5 6 5
6 5
6 5.
5 6 . 5 5
55 6 5 5 5
5 5 6 5 5 5
5 5 5 6 5
6 55 5
5 6 5
5 5 55 5 5 5 5
5 . 56 6 5 5 5
6 6 . 6 6 6 6 5 5.55
55 5 5 5 5 5 5 6 5
. 5 . 5 5 .
5 5 6
5 5 5 5 6 . 5 6
. 5 6 5 .5 5
55 5 5 6 5 55 6
56 5 5 56 5 5
5 55 5 5 . 6 6
5 6 . 5 5 5
5 55 55
555 5 5 6 5 . 5
65 . 5 5 55
56 5 5 5 . 5
5 5 55 5 5. 6 5 5
. 5 5 65 5 5 6 5
6 65 6 6 5 . 55 5
5 6 55 .
. 5 5 5 6
66 5 5 5 5
55 565 5 6 . 5
5 5 5 6 6 5 66 5
5 5 5 5 . . 5
56 6 5 65 5
5 5 5 5
5 5 5 . 5 6
56 6
.6 5 5 55 5 5
5 5 55
. 6 6 5 5 5
5 5 5
6
5 5
6
6
66 5 6 5 56 6
6 5 5
. 5 .5 . 6
.5. 6. . 5 .
6 5
5 55 55 5 5 . 5
5 5 .
55 5 5 5 5 5 6 5
5 . 5
5 55 5 5 5 5 6
5 5 5 5 6
5 . 5 5.5 6 . 56 5
5 6 5 56 5.
5 5 6 5 5 5 5 5 5
. 5.
. 6555 56 .
656 5 6 . 5
. 5 . 5 5 5 6 6
5 6 5 565. 6 5 5
555 5 65 5
5 6 65 56 . 5. 5 6
5 55 5 6 5
5 6 6 6 6 5 . 5 6 55 5
5 5 5 .65 6 5 5 5 5
55 5 5 . . 5 . 6 5 5
5. . 55 5 5.
55 55 5 5 5 6
5 55. 5 56
5 5 6 5 65 655 65
5 5 5 5 .
5 66 5 . 5
6 5 5
55 5 5 5 55. 6
6 56 6 5 5 5
5 6 5 6 5 5
5
抱歉，您提供的内容似乎不完整，只显示了“5 6”。请提供完整的科技文章原文，以便我为您进行忠实翻译。
5 6 5 6 6 6 5 5 5 5 5 5 5 5 6 5..5.5 5 5 5 5 5 5..5 5..5 5 5 5 5 6 5 6 5 5 5 5.6 5 5.6 5 5 5 5 6 5 5.5 6 5 5.5 5 5 5 5 6 5 5 6 5 6.5 5 6 5 5 5 5.5 5 5 5 5.5 5 5 5 5 5 6 6 6 5 5.5 5 5 5 5.6 6 5 5 5 5 6 6 5 5 5.5 5 6 6 5 5 5 6 5 5 5 6 5 5 5 6 5 5 6 5 5 5.5 5.6 6 6 5 5 5 5 5 5 6 6 5.5 5 5 5
抱歉，您提供的内容似乎包含大量无意义的数字和空白字符，没有可翻译的科技文章。请提供有效的英文科技文本，我将忠实翻译为简体中文，并保留所有 Markdown 格式。
抱歉，您提供的内容似乎是一段包含大量空白和数字“5”的文本，没有实际的科技文章内容可供翻译。请提供需要翻译的科技文章原文，我将忠实为您翻译为简体中文，并保留所有 Markdown 格式。
抱歉，您似乎没有提供需要翻译的科技文章内容。请将原文粘贴在消息中，我会忠实为您翻译为简体中文，并保留所有 Markdown 格式。
565...65
6555.5..6..565
.5....566.5.555655
56.55.6555.5566656.5
5565.5556.5...6..656.6
.65.56.566555555.5656..5
656555565.65.6.5555566.556
66.565.55566565555.5.555..
5555556555.6565.6555556.6565
5...56.55555...5..655.55..56
5655.666665.5566.5655..56665
55655566665.65555555655555..
5.66556.5...55.56.6.5.5.55.6
.6.5.65555..565.5555.55666..
565.5565..66556..56.555.655.
5.55556....5.6555.65..5..6
56566556.5665556556555.6.6
55.5555555..6..565655.65
665.56565556555565.555
.555566.66..566..655
56.65655566.655...
.65665556.6656
56.55555 65 . 6
5 5 5
. . 5 6. 6 5
5 55. 5 5 5 56
6 6 . .. 6.6
6 55 5. 56 5
5 6 5 5 .
6 5 5 66 5 .
55 55 5 . 5. 6
. .5 5 . 5.. .
66 . 5 6
5 55 6 6 5
. . 5 5 5 5
6 6 5 65 5 5 6
6 5 . 5 5 6
5 6 . 5 5 65 .
6 5 6 56 6
555 55 ..
. 6 6 .
5 6 .5 6
. 6. 5 .
. 5 . 5
6 5 5
5 5
. 55 5
. 6 5 6
. 5
5 5 5
55 . 5
6 5 5.
5 65 5
5 . . . 5
5 6 6 . 6 5 . 6
56 5 6 5 55 5
66 56 .. 5 6. . . 5
5 5 55 5 6
5 . 65 . 56 55 .
. 6 6
56 . 5 6
5 .5 6 .6
6 5 5 5 55
6 6 . 55
6 56 6 65 .
5 6 6 6 6
55 5 . .
. .. .. 6
.. 5
5
5 . 5 . 6
6 . .5 .
6 5 . 5 6. 5 6
555 6 5 5 5
6 5 5 65
6 5 5 65 .5 56
5 . 6 5 6 5 5
6 6 5 55 .
5. 5 . 6 . 6
. . .5 5 5 . ..
5 66 56. . 5
. 5. .5 .
5 56655 5 5. .
5 5
6 65 55 555
.5 5 6. . . 6
5 555
6 55
5 .5 5
6 5 6
5. 6 5
6 . 6 6 .5
5 5 55 5. 6
. 56 6 5 56 .
65 5 6 5 6 556
6. . 5 5 . 5 .
5 5 56 655 5 6.
5 5. 5.
65 5 5 56 . 6
5 6 . 5 5 5 5 .
5 5 .
. . . 5 6
5 . 5 .
5555 . .. 5 .
5 6 655
5 ..6 65 55 5
5 6 5 5
5 6
6 .
6 6
.5.5 6 5..6 6 5 5.5 5 5 6 5 6.6 6.5 6...6..5 6 5 5.5 6.6 6 5 6
556565
56.5666.66
5555.6.6....
.5.5...56.556.
55656.55..655.
5....555555.55
.6.5.5566565.5
6.65.566.5..6.
556.555.5.6.
66.55.6555
6.5.55 5 6
6 66
5 . .
. . . .
6 5
5 . 5 5 5
6.5. 66 5.
6.65 .6
5
6 5 5
5 5
5 .
5 6. .
5
5 5 . 5.
. . 5 .
. 56 5
.
5 . 5. .
6 .
5 565
6 5 6
5 5 . .
.5 5 5
5 6
5
5
566 5.
.
5 5
6.
6 .
6
5.. 6 6
6 5 . 5.
. 55 5
5
. .
56 5 5 6
. 65
5.
5.6 5..5 5.6 5 5 5 5 5.5..5.5 6
2026年6月26日
预览 GPT‑5.6 Sol：下一代模型
加载中…
分享
能力
我们开始对 GPT‑5.6 系列进行有限预览：Sol，我们的旗舰模型；Terra，适用于日常工作的均衡模型；以及 Luna，快速且经济实惠的模型。Terra 的性能与 GPT‑5.5 相当，但价格便宜 2 倍，而 Luna 则以最低成本提供强大能力。
GPT‑5.6 Sol 以我们迄今为止最强大的安全堆栈推出。我们加强了对高风险活动、敏感网络请求和重复滥用的保护，并花费数周时间寻找弱点、压力测试我们的系统，并使其能够抵御现实世界的攻击。
我们坚信广泛的可及性，并计划在未来几周内将 GPT‑5.6 Sol、Terra 和 Luna 全面开放。作为我们与美国政府持续合作的一部分，我们在今天发布前预先展示了我们的计划及模型能力。应其要求，我们首先向一小批已与政府共享参与信息的受信任合作伙伴提供有限预览，随后再更广泛地发布。在预览期间，我们将继续测试并与合作伙伴密切协调，以推进更广泛的可用性。我们并不认为这种政府准入流程应成为长期默认模式。它会使最优秀的工具远离那些真正需要它们的用户、开发者、企业、网络防御者和全球合作伙伴。我们采取这一短期步骤，是因为我们相信这是在未来几周内实现更广泛可用性的最有力途径，同时我们将与政府合作制定网络安全行政令框架，并为未来模型发布建立可重复的流程。
能力
GPT‑5.6 Sol 是我们迄今为止最强大的模型。为预览模型性能，我们分享了一组评估结果，突出展示了在编程、生物学和网络安全方面增强的智能体能力，更多安全与准备评估可在我们的系统卡(在新窗口中打开)中查看。当模型广泛可用时，我们将分享更全面的评估结果。
借助 GPT‑5.6，我们引入了新的 max 推理努力级别，让 Sol 拥有最充足的时间进行深度推理。此外，我们还推出了新的 ultra 模式，该模式通过利用子智能体加速复杂工作，超越了单一智能体的能力。
在编程工作流方面，GPT‑5.6 Sol 在 Terminal‑Bench 2.1 上树立了新的行业标杆，该基准测试需要规划、迭代和工具协调的命令行工作流。
GPT‑5.6 Sol 在生物学工作流方面也展现出广泛改进。在 GeneBench v1（评估长周期基因组学和定量生物学分析）上，它使用更少的 token 就取得了比 GPT‑5.5 更强的结果。
GPT‑5.6 Sol 是我们迄今为止在网络安全方面能力最强的模型。它改变了长周期安全任务（包括漏洞研究和利用）的性能效率边界。在 ExploitBench² 上，GPT‑5.6 Sol 仅使用约 1/3 的输出 token 就与 Mythos Preview 不相上下。在 ExploitGym(在新窗口中打开)³（由加州大学伯克利分校研究人员与 OpenAI 及其他前沿实验室合作创建的基准测试）上，随着推理能力的提升，GPT‑5.6 Sol、Terra 和 Luna 模型均在网络能力方面展现出显著改进。
更强的网络能力与更强的安全保障
我们开发了 GPT‑5.6 Sol、Terra 和 Luna，并配备了迄今为止最强大的安全保障，其配置与每个模型的能力相匹配。随着模型能力增强，我们设计的安全保障能更好地抵御现实世界中的对抗性压力，同时保留对合法工作（如代码审查、漏洞研究、补丁开发、调试、安全教育和防御性测试）的访问权限。我们的目标是让被禁止的攻击性活动变得更困难、更不确定且更易被检测，同时不必要地限制那些有益用途。基于我们对模型和安全保障的评估，我们预计合法防御性工作将获得显著收益，同时有效约束被禁止的攻击性使用。
GPT‑5.6 Sol 在帮助人们发现和修复漏洞方面，比可靠地执行端到端攻击更为出色。随着这些能力的持续进步，我们的优先事项是确保它们能够惠及防御者，他们可以利用这些工具发现弱点、开发补丁并更广泛地强化系统。
根据我们的准备框架，GPT‑5.6 Sol 并未跨越网络关键阈值。在涉及 Chromium 和 Firefox 的评估中，它识别了漏洞和利用原语（即利用的构建块），但在测试条件下并未自主生成功能完整的全链利用。然而，基准测试阈值无法涵盖模型可能被使用或与其他工具结合的所有方式。这种不确定性，加上模型能力的更广泛阶跃变化，正是我们将模型增强能力与更强安全保障及分阶段发布相结合的原因。我们在 GPT‑5.6 预览系统卡(在新窗口中打开)中分享了更多关于安全保障的细节。
分层安全保障体系
没有任何单一保障足以应对坚决或适应性的滥用。在 GPT‑5.6 预览中，我们采用分层安全保障，各模型的具体配置有所不同，并针对现实世界攻击进行压力测试。这些保障包括训练到模型中的保护措施、生成过程中的实时检查、账户级信号、差异化访问、监控、执行和持续测试。
GPT‑5.6 经过训练，会拒绝被禁止的网络协助，包括当用户试图掩饰其意图或越狱模型时。这些模型级保障确立了模型应协助和不应协助的第一道边界。
实时网络和生物学滥用分类器通过评估生成时的输出来提供另一层保障。对于高风险情况，如果检测到潜在违规，生成可能会暂停，同时一个更大的推理模型会审查对话及其上下文。如果输出被评估为不允许，则在到达用户之前被拦截。
被标记的活动还可能触发跨相关对话和风险信号的账户级审查，这与我们关于内容保留和审查的条款和政策一致。超越单一对话的审查有助于我们的系统区分持续性恶意行为与合法的双重用途安全工作，其中类似的技术概念可能出现在截然不同的上下文中。
这些层次共同使整体方法比任何单一保障更为稳健。模型行为降低了有害响应的可能性，实时系统可以在生成过程中进行干预，账户级审查可以识别更广泛的模式，差异化访问则保留了重要的防御性工作，而不会默认广泛提供最敏感的能力。
特别是在预览期间，用户可能会遇到阻止或拒绝某些请求的安全保障。其他请求可能需要更长时间，因为生成过程会暂停以进行额外审查。安全保障有时可能会干预合法工作，尤其是在防御性和攻击性活动最初看起来相似的双重用途领域。
这正是预览版旨在测试的一部分。我们不仅想了解安全措施能否限制滥用，还想确认合法用户是否仍能可靠、高效地完成正常工作。预览期间的反馈将帮助我们减少不必要的拦截和延迟，改进安全措施对上下文的解读能力，并在更广泛发布前创造更流畅的体验。
我们也在与企业客户合作，探索更长期的方案——包括隐私保护检测、客户操作的安全控制，以及根据客户、用户或工作负载的风险级别调整访问权限——在支持企业隐私要求的同时提升安全性。
通过自动化红队测试提升鲁棒性
当攻击者调整策略时，安全措施也需要保持有效。仅能应对一组固定已知攻击的保护措施，对于前沿模型而言是不够鲁棒的。
因此，我们在安全方面投入了比以往更多的智能和算力，利用我们自己的模型来更快地发现弱点并改进安全措施。我们投入了超过70万A100等效GPU小时用于自动化红队测试，旨在寻找通用越狱方法：即那些能在多种提示或上下文中生效的攻击，而不仅限于单一狭窄场景。专注于这些更困难、更通用的攻击，使我们能够超越一组固定的已知故障来测试安全措施。这也让我们能够探索比仅靠人工测试多得多的攻击模式，更早地识别故障模式，并缩短从发现弱点到解决弱点之间的路径。
除了自动化红队测试，我们还与第三方测试人员合作，进行了广泛的人类专家红队测试，这一工作将在预览期间持续进行。人类红队测试通过让富有创造力的专家尝试以我们系统可能无法预料的方式滥用模型，来补充自动化工作。
没有任何评估能够代表每一种产品配置、多步骤攻击或真实世界的工作流程。因此，我们维持一个快速响应流程，以复现、评估、确定优先级并修复新发现的越狱方法，然后将它们纳入我们持续的评估中，以便未来能够针对类似的故障进行测试。
可用性与定价
在预览期间，GPT‑5.6 模型将首先通过 API 和 Codex 向一组经过筛选的受信任合作伙伴和组织提供。我们计划很快将其更广泛地提供给使用 ChatGPT、Codex 和 API 的用户。
在 GPT‑5.6 引入的这个新命名系统中，数字标识模型的代际，而 Sol、Terra 和 Luna 则标识持久的能力层级，这些层级可以按照自己的节奏进行升级。整个系列为用户和开发者在智能、速度和成本之间提供了更清晰的选择。
GPT‑5.6 按每百万 Token 定价，涵盖三种模型尺寸：Sol 为 5 美元输入 / 30 美元输出；Terra 为 2.50 美元输入 / 15 美元输出；Luna 为 1 美元输入 / 6 美元输出。GPT‑5.6 还引入了更可预测的提示缓存功能，包括支持显式缓存断点和至少 30 分钟的缓存生命周期。对于 GPT‑5.6 及后续模型，缓存写入按模型未缓存输入费率的 1.25 倍计费，而缓存读取则继续享受 90% 的缓存输入折扣。
我们还将于 7 月在 Cerebras 上推出 GPT‑5.6 Sol，速度高达每秒 750 个 Token，以前所未有的速度将前沿智能带给客户。随着我们扩展容量，初期访问将仅限于特定客户。
我们很高兴能继续从本次预览期中学习，并尽快将 GPT‑5.6 Sol、Terra 和 Luna 带给更多人。
我们通过观察模型的线上行为并进行离线模拟来估算延迟和 API 成本。这些估算考虑了工具调用细节、采样 Token 和输入 Token。实际结果可能会有很大差异，并取决于我们模拟中未包含的许多因素。我们以快速 API 速度模拟延迟，并以常规 API 定价模拟成本。
所有模型均使用 ExploitBench API 工具进行评估，采用 5 个种子和推理连续性。
我们在 alpha API 上运行了 ExploitGym，该 API 输出响应的速度快于我们的公开 API，然后我们重新缩放以匹配公开 API。当将延迟重新缩放到我们公开 API 的预期速度时，这会导致某些估算延迟超过 2 小时和 6 小时的时间限制，尽管在评估运行中这些限制被正确遵守。为了在时间敏感的工作中获得更快的速度，我们在 API 中提供优先处理，在 Codex 中提供快速模式。
未报告输出 Token、延迟或成本的模型以水平虚线绘制。
作者
OpenAI
继续阅读
GPT-5.6 现已成为 Microsoft 365 Copilot 中的首选模型 产品 2026年7月9日
GPT-5.6：与您雄心相匹配的前沿智能 产品 2026年7月9日
ChatGPT 现已成为您最具雄心工作的合作伙伴 产品 2026年7月9日
研究
最新进展
安全
产品
- ChatGPT（在新窗口中打开）
- ChatGPT Business（在新窗口中打开）
- ChatGPT Enterprise（在新窗口中打开）
- ChatGPT for Education（在新窗口中打开）
- Codex
- 发布说明
API 平台
商业
开发者
公司
支持
更多
条款与政策
（在新窗口中打开）（在新窗口中打开）（在新窗口中打开）（在新窗口中打开）（在新窗口中打开）（在新窗口中打开）（在新窗口中打开）
OpenAI © 2015–2026 您的隐私选择
英语 美国
Log inTry ChatGPT(opens in a new window)
- Research
- Products
- Business
- Developers
- Company
- Foundation(opens in a new window)
Try ChatGPT(opens in a new window)Login
OpenAI
0 0 . 0 . . 0 . . .
. 0 . 0 0 0 0 . 0
. . . 0
. . .. 0 0 .
. 0 . .
0 0 .
. . . . .. . . . . . 0
0 . . . . 0 . 0
. . . . 0 . . . . .. .. . . . . 0 . . . . 0 . 0 0 . . . . . . 0 . 0 0 . . . 0 . . . . . 0 . . . . 0 . . . 0 0 0 . . . .. . . . . 0 0. 0 0 . . . 0 . .0 . 00 . 0 . . 0 . . . . . . . 0 . . . . . . . 0 0 0 . . . 0 . . . 0 0 . . . . . . 0 . . . . . . .. . 0 .. . . . 0 . .0 0 . . . . .. 0 . . . 0 0 . 0 . 0 . .. . . . .. . . . . . . . . 0 . 0 . 0 . .0. . . . . 0 . . 0 . . . . .. . . . . . .. . . . 0 . . 0 . . . 0 . . .. . . . 0 . . . . .
655655555555
55555666655555.665
555556565.55555655555665
566656655555.565665556565.65
555.55555655556565.56656.5555556
5555.555555.55565655555.5665555655
56.6.565.5.66..55655665.55555555.556
56.566555.556556556.665556556565665.5.55
55555.55555655555.55565.565.66555555.5.5
65.555566555...565555..5.555.565555.56.655
556.5555656556.56555.55556556555555565665565
5556.55555555555566555.5566.5555.556656.655556
6.555555655556565565.55.5555655555555555.66555
.66555.5.65555555655566555555.665556.5555.565566
5555.655555655.65.5.5565.55.655..55656.555555555
5665.55555565656.55556.5..6656565555656555.65555
56.55655566555566565...5.5556.56.5655.555555655555
6555555555556555555555555.555565565556.55.55565556
5555.65555556565555556555.56665556566556.5556.5655
.5665.5.55555555565556566556565655.6655556.5555555
56656.565565555.5556.555555655555.5.55655.65.66.56
56.55.655555555555555555.565..55666555555555.55555
5555.556555665.5555555.555655555.5556565.555566555
.665565656566.5566655556565655656.5..55.6.65556655
55.565.55666555555566.5555555555566..55555556..6
55.555556655.56566.6565655.555.555655.5655555555
665555555.555556665.6566555565565.55565565565556
655.55555555555.5.55.65655566.5566565.55555.55
56555565.556.5666666555.65665555565556655.5565
5.555656555555555555655..5.5.655555655565555
56655665556565655555.665555556555565566655
.6655665.665555.55655655565665565555.555
5555555655555565565555555.5.55.66656555.
5565555556665.5.6555665556.665565555
6.66566555555..5555555555555655555
556656555555.55555555555555.5556
56..5556665655.5565555555556
.55555655655..6665555655
6565656.5555555555
565555555665 565 5
6 5
5 5 5 5
6 5 5 5
. 5 5 5 6 5
55 5 6 55 5 55
6 . 5
5 5 5 5 5 5 6 565 5
5 . 55 5 5
5 5 . 5 . 5 56 555 6
6 6 5 .5 65 5 6
5 6 5 5 5 5 5 .6
5 65 5.55 5 .
66 5 55 65 6 .
5 5 5 5.6
. 6 55 . 6 6 5
5 5 5 56 5 5
6 5 55 56 5 65 5
5 5 5 .56 5 6 5 6 6 5 5
5 6 5 5
6 5 . 5 5 5 5
5 5 55 55 5
5 5 6 5 6 . 5
. 5 6 5 5 . 5
5. . 6 5 . 5 . 5
.5 .6 5 5 5 5
5 6 5 6 5 55 6 5 5
5 5 5 5
5 6 6 6 6 5 6
5 55 . 5 5
6 55 5 5 5 5
. 6 5 5 5
55 5 5 5 5 . 5 5
5 5 65 5
. 5 555
55 5 55 5
. 5 5 5 5
. 5 ..6 55 5
5 55 555
6 5 5 5
55
5 6 65 5
5.
5 6 5 5
5 5 6
.5 6 5 5
. 5 5 5 6 66
55 5 6 6 .
6 5 5 6 5 5 5 5
5 56 6 6 5 5 5
5 5 5 6.55 5 66 5 5
5 6 55 5 5 55
. . 5 . 6 5 5 5
5 6 55 . 5 5
5 65 5 5 6
6 .. . 55 5 5
55 5 5 5 6 5 5 56
5 65 5 5
6 5 5 56 . 55 6 5
6 5 55 5 . . 66
6 . 5 5 5 5
5 5 5 5 55 5 5
6 6 . 5 5 5 5 6 6
5 6 5 5 5 5 5 5 . .
5 5 5 5 6 .5 5 5.5 5
6 6 . 55 6
6 . 55 5 6 5
6 5 5 6 6 5 65 5 6
6 5 5 6 6 5
6 5 6 5 5
6 566 .5 5 6 65 .
5 55 65 6 5
5 55 . 5 6 55 6 6 5 5
5 55
565 55 5 5
5 5 5 5 5
6
6 6 5
5 5
5 6 .6
5 .5 5 5 5 5
5
5 55 5 6 5.
5 6 5 55
5 5 5 6
5 6 5
5 . 5
66 5 . 6 5
. 5 . 5 6 5 5 65
. 55 5 6 5 5 6 5
6 55
5 55 5 5 . 5
5 5.6 . 5 5 555
6 5 56 65 . 5 55 5 5 5
6 5 6 . . 5 5 6 . 5
55 5 55 5 5 . 55 6
5 5 6 5
. .5 5 55 55 5 6 5 6 55
5 5 5 6 5 . 5 . 6
5 .5 .55
5 5 . 5 55 . 5 6 5 6
6 5 5 6 .
6 5 5 6 555 6
55 6 5 5 5
55 5.55 6 5 6
5 55 5 .6 6
5 . 6 5
65 5 55 . 5 5 555
6 56 6 5 5 6665
6 6 55
6 5 . 6
55 . 5 6 5
6 6 5 5 5 5
6 . 5 5
5 6
55 5 6 65
5 5
5 5 6 5
55 6
5 6
66 5 5 5 6 6 5.
55 5 5 6. 5 5 6
5 5 6
6 6 5 55 6 5 5
5 . 5 5 .
. 5 6 . 6 5
5 5 . .5 . 5
55 5 5 5 5
5 5 6 5 5
6 5 5 5 5 65 5 5 6 5
. 6 6 5 555 5 66
5 5 5 . . 5
6 5 .5 5 5
6.5 6 5 5 6 5 65
5 55 55 5 5 556 5
6 556 55 5 6 5 5 6
5 5 6 5 6 5 5 5
56 5 6 5 5 5 5 6 5
5 . 55 5 5 5 655 5
5 5 6 5 5.5 5
5 6 5 5 6 6 5 6
5 5 6 5 555 6 5 6
5 6 5 5 5
5 5 6 5 5 6
5 5 5 6 5 5 .5
5 6 5 5 5 5 5 5
. 5 5 5 6 5
55 5 6 5 5
. 5 5 5 5
5 5 56 55 5 6 5
5 6 6 5
6 55 . 5 5 5 6 5
6 5
6 5.
5 6 . 5 5
55 6 5 5 5
5 5 6 5 5 5
5 5 5 6 5
6 55 5
5 6 5
5 5 55 5 5 5 5
5 . 56 6 5 5 5
6 6 . 6 6 6 6 5 5.55
55 5 5 5 5 5 5 6 5
. 5 . 5 5 .
5 5 6
5 5 5 5 6 . 5 6
. 5 6 5 .5 5
55 5 5 6 5 55 6
56 5 5 56 5 5
5 55 5 5 . 6 6
5 6 . 5 5 5
5 55 55
555 5 5 6 5 . 5
65 . 5 5 55
56 5 5 5 . 5
5 5 55 5 5. 6 5 5
. 5 5 65 5 5 6 5
6 65 6 6 5 . 55 5
5 6 55 .
. 5 5 5 6
66 5 5 5 5
55 565 5 6 . 5
5 5 5 6 6 5 66 5
5 5 5 5 . . 5
56 6 5 65 5
5 5 5 5
5 5 5 . 5 6
56 6
.6 5 5 55 5 5
5 5 55
. 6 6 5 5 5
5 5 5
6
5 5
6
6
66 5 6 5 56 6
6 5 5
. 5 .5 . 6
.5. 6. . 5 .
6 5
5 55 55 5 5 . 5
5 5 .
55 5 5 5 5 5 6 5
5 . 5
5 55 5 5 5 5 6
5 5 5 5 6
5 . 5 5.5 6 . 56 5
5 6 5 56 5.
5 5 6 5 5 5 5 5 5
. 5.
. 6555 56 .
656 5 6 . 5
. 5 . 5 5 5 6 6
5 6 5 565. 6 5 5
555 5 65 5
5 6 65 56 . 5. 5 6
5 55 5 6 5
5 6 6 6 6 5 . 5 6 55 5
5 5 5 .65 6 5 5 5 5
55 5 5 . . 5 . 6 5 5
5. . 55 5 5.
55 55 5 5 5 6
5 55. 5 56
5 5 6 5 65 655 65
5 5 5 5 .
5 66 5 . 5
6 5 5
55 5 5 5 55. 6
6 56 6 5 5 5
5 6 5 6 5 5
5
5 6
5 6 5 6 6 6 5 5 5 5 5 5 5 5 6 5..5.5 5 5 5 5 5 5..5 5..5 5 5 5 5 6 5 6 5 5 5 5.6 5 5.6 5 5 5 5 6 5 5.5 6 5 5.5 5 5 5 5 6 5 5 6 5 6.5 5 6 5 5 5 5.5 5 5 5 5.5 5 5 5 5 5 6 6 6 5 5.5 5 5 5 5.6 6 5 5 5 5 6 6 5 5 5.5 5 6 6 5 5 5 6 5 5 5 6 5 5 5 6 5 5 6 5 5 5.5 5.6 6 6 5 5 5 5 5 5 6 6 5.5 5 5 5
5
5
5.6...6.5..6..5..
5
5
5
5
5
5
565...65
6555.5..6..565
.5....566.5.555655
56.55.6555.5566656.5
5565.5556.5...6..656.6
.65.56.566555555.5656..5
656555565.65.6.5555566.556
66.565.55566565555.5.555..
5555556555.6565.6555556.6565
5...56.55555...5..655.55..56
5655.666665.5566.5655..56665
55655566665.65555555655555..
5.66556.5...55.56.6.5.5.55.6
.6.5.65555..565.5555.55666..
565.5565..66556..56.555.655.
5.55556....5.6555.65..5..6
56566556.5665556556555.6.6
55.5555555..6..565655.65
665.56565556555565.555
.555566.66..566..655
56.65655566.655...
.65665556.6656
56.55555 65 . 6
5 5 5
. . 5 6. 6 5
5 55. 5 5 5 56
6 6 . .. 6.6
6 55 5. 56 5
5 6 5 5 .
6 5 5 66 5 .
55 55 5 . 5. 6
. .5 5 . 5.. .
66 . 5 6
5 55 6 6 5
. . 5 5 5 5
6 6 5 65 5 5 6
6 5 . 5 5 6
5 6 . 5 5 65 .
6 5 6 56 6
555 55 ..
. 6 6 .
5 6 .5 6
. 6. 5 .
. 5 . 5
6 5 5
5 5
. 55 5
. 6 5 6
. 5
5 5 5
55 . 5
6 5 5.
5 65 5
5 . . . 5
5 6 6 . 6 5 . 6
56 5 6 5 55 5
66 56 .. 5 6. . . 5
5 5 55 5 6
5 . 65 . 56 55 .
. 6 6
56 . 5 6
5 .5 6 .6
6 5 5 5 55
6 6 . 55
6 56 6 65 .
5 6 6 6 6
55 5 . .
. .. .. 6
.. 5
5
5 . 5 . 6
6 . .5 .
6 5 . 5 6. 5 6
555 6 5 5 5
6 5 5 65
6 5 5 65 .5 56
5 . 6 5 6 5 5
6 6 5 55 .
5. 5 . 6 . 6
. . .5 5 5 . ..
5 66 56. . 5
. 5. .5 .
5 56655 5 5. .
5 5
6 65 55 555
.5 5 6. . . 6
5 555
6 55
5 .5 5
6 5 6
5. 6 5
6 . 6 6 .5
5 5 55 5. 6
. 56 6 5 56 .
65 5 6 5 6 556
6. . 5 5 . 5 .
5 5 56 655 5 6.
5 5. 5.
65 5 5 56 . 6
5 6 . 5 5 5 5 .
5 5 .
. . . 5 6
5 . 5 .
5555 . .. 5 .
5 6 655
5 ..6 65 55 5
5 6 5 5
5 6
6 .
6 6
.5.5 6 5..6 6 5 5.5 5 5 6 5 6.6 6.5 6...6..5 6 5 5.5 6.6 6 5 6
556565
56.5666.66
5555.6.6....
.5.5...56.556.
55656.55..655.
5....555555.55
.6.5.5566565.5
6.65.566.5..6.
556.555.5.6.
66.55.6555
6.5.55 5 6
6 66
5 . .
. . . .
6 5
5 . 5 5 5
6.5. 66 5.
6.65 .6
5
6 5 5
5 5
5 .
5 6. .
5
5 5 . 5.
. . 5 .
. 56 5
.
5 . 5. .
6 .
5 565
6 5 6
5 5 . .
.5 5 5
5 6
5
5
566 5.
.
5 5
6.
6 .
6
5.. 6 6
6 5 . 5.
. 55 5
5
. .
56 5 5 6
. 65
5.
5.6 5..5 5.6 5 5 5 5 5.5..5.5 6
June 26, 2026
Previewing GPT‑5.6 Sol: a next-generation model
Loading…
Share
Capabilities
We're beginning a limited preview of the GPT‑5.6 series: Sol, our flagship model; Terra, a balanced model for everyday work; and Luna, a fast and affordable model. Terra has competitive performance to GPT‑5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost.
GPT‑5.6 Sol launches with our most robust safety stack to date. We strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, and spent multiple weeks finding weaknesses, pressure-testing our system, and hardening it against real-world attacks.
We believe in broad access, and we plan to make GPT‑5.6 Sol, Terra, and Luna generally available in the coming weeks. As part of our ongoing engagement with the U.S. government, we previewed our plans and the models’ capabilities ahead of today’s launch. At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly. During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability. We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks, while we work with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases.
Capabilities
GPT‑5.6 Sol is our strongest model yet. To give a preview of model performance, we share a set of evaluations highlighting improved agentic capabilities in coding, biology, and cybersecurity, with additional safety and preparedness evaluations available in our system card(opens in a new window). We will share an expanded suite of evaluation results when we make the model broadly available.
With GPT‑5.6, we’re introducing a new max reasoning effort to give Sol the most time to reason deeply. Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
For coding workflows, GPT‑5.6 Sol sets a new state of the art on Terminal‑Bench 2.1, which tests command-line workflows requiring planning, iteration, and tool coordination.
GPT‑5.6 Sol also shows broad improvements in biology workflows. On GeneBench v1, which evaluates long-horizon genomics and quantitative-biology analyses, it achieves stronger results than GPT‑5.5 while using fewer tokens.
GPT‑5.6 Sol is our most capable model yet for cybersecurity. It shifts the performance-efficiency frontier for long-horizon security tasks including vulnerability research and exploitation. On ExploitBench², GPT‑5.6 Sol is competitive with Mythos Preview using only ~1/3 of the output tokens. On ExploitGym(opens in a new window)3, a benchmark created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs, GPT‑5.6 Sol, Terra, and Luna models all demonstrate strong improvements in cyber capabilities as we increase reasoning.
Stronger cyber capabilities with stronger safeguards
We developed GPT‑5.6 Sol, Terra and Luna with our most robust safeguards to date, with configurations matched to each model’s capabilities. As the model becomes more capable, we design safeguards to increasingly hold up to real-world adversarial pressure while preserving access to legitimate work such as code review, vulnerability research, patch development, debugging, security education, and defensive testing. Our goal is to make prohibited offensive activity more difficult, uncertain, and detectable without unnecessarily limiting those beneficial uses. Based on our assessment of the model and safeguards, we expect substantial benefit for legitimate defensive work, while meaningfully constraining prohibited offensive use.
GPT‑5.6 Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks. As these capabilities continue to advance, our priority is to make sure they reach and benefit defenders, who can use these tools to find weaknesses, develop patches, and strengthen systems more broadly.
GPT‑5.6 Sol does not cross the Cyber Critical threshold under our Preparedness Framework. In evaluations involving Chromium and Firefox, it identified bugs and exploitation primitives—the building blocks of an exploit—but did not autonomously produce a functional full-chain exploit under the conditions tested. Still, benchmark thresholds cannot capture every way a model may be used or combined with other tools. That uncertainty, along with the model’s broader step change in capabilities, is why we are pairing the model’s increased capabilities with stronger safeguards and a phased release. We share more details about our safeguards in the GPT‑5.6 Preview system card(opens in a new window).
A layered safeguard stack
No single safeguard is sufficient against determined or adaptive misuse. Across the GPT‑5.6 preview, we use layered safeguards, with exact configurations varying across models, and pressure-test them for real-world attacks. These include protections trained into the model, real-time checks during generation, account-level signals, differentiated access, monitoring, enforcement, and continued testing.
GPT‑5.6 is trained to refuse prohibited cyber assistance, including when users attempt to disguise their intent or jailbreak the model. These model-level safeguards establish the first boundary around what the model should and should not help with.
Real-time cyber and biology misuse classifiers provide another layer by evaluating output as it is generated. For higher risk cases, if they detect a potential violation, the generation may be paused while a larger reasoning model reviews the conversation and its context. If the output is assessed as disallowed, it is withheld before it reaches the user.
Flagged activity can also trigger account-level review across relevant conversations and risk signals, consistent with our terms and policies around content retention and review. Looking beyond a single conversation helps our systems distinguish persistent malicious behavior from legitimate dual-use security work, where similar technical concepts may appear in very different contexts.
Together, these layers make the overall approach more robust than any one safeguard on its own. Model behavior reduces the likelihood of harmful responses, real-time systems can intervene during generation, account-level review can identify broader patterns, and differentiated access preserves important defensive work without making the most sensitive capabilities broadly available by default.
Especially during the preview, users may encounter safeguards that block or refuse some requests. Other requests may take longer because generation is paused for additional review. Safeguards may occasionally intervene on legitimate work, particularly in dual-use areas where defensive and offensive activity can initially look similar.
That is part of what the preview is designed to test. We want to understand not only whether the safeguards constrain misuse, but whether legitimate users can still complete normal work reliably and efficiently. Feedback during the preview will help us reduce unnecessary blocks and delays, improve how the safeguards interpret context, and create a smoother experience before wider release.
We are also working with enterprise customers on longer-term approaches—including privacy-preserving detection, customer-operated safety controls, and access calibrated to the risk of a customer, user, or workload—to advance safety while supporting enterprise privacy requirements.
Improving robustness with automated red-teaming
Safeguards also need to remain effective when attackers adapt their tactics. A protection that works only on a fixed set of known attacks is not robust enough for a frontier model.
That’s why we are applying more intelligence and compute than ever before to safety, using our own models to find weaknesses and improve safeguards faster. We dedicated over 700,000 A100-equivalent GPU hours to automated red teaming aimed at finding universal jailbreaks: attacks that can work across many prompts or contexts, not just one narrow setting. Focusing on these harder, more general attacks let us test the safeguards beyond a fixed set of known failures. It also lets us explore far more attack patterns than human testing alone could cover, identify failure patterns earlier, and shorten the path from finding a weakness to addressing it.
In addition to automated red-teaming, we worked with third-party testers to conduct extensive human expert red teaming, which will continue in the preview period. Human red-teaming complements the automated work by testing safeguards against creative experts trying to misuse the model in ways our systems might not anticipate.
No evaluation can represent every product configuration, multi-step attack, or real-world workflow. We therefore maintain a rapid-response process to reproduce, assess, prioritize, and remediate newly discovered jailbreaks, then add them to our ongoing evaluations so we can test against similar failures in the future.
Availability and pricing
During the preview, GPT‑5.6 models will initially be available through the API and Codex to a select group of trusted partners and organizations. We plan to make them more broadly available to people using ChatGPT, Codex, and the API soon.
In this new naming system introduced with GPT‑5.6, the number identifies a model’s generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence. Together, the family gives people and developers clearer choices across intelligence, speed, and cost.
GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount.
We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed. Access will initially be limited to select customers as we expand capacity.
We’re excited to continue learning from this preview period, and to bring GPT‑5.6 Sol, Terra and Luna to more people soon.
We estimate latency and API cost by looking at the production behavior of our models, and simulating offline. These estimates account for tool call details, sampled tokens, and input tokens. Real-world results may vary substantially, and depend on many factors not captured in our simulation. We simulate latency at fast API speeds, and cost at regular API pricing.
All models are evaluated using the ExploitBench API harness with 5 seeds and reasoning continuity.
We ran ExploitGym on our alpha API, which outputs responses faster than our public API, and then rescaled to match our public API. When rescaling latencies to the speeds expected for our public API, this causes some estimated latencies to exceed the 2h and 6h hour time limits, despite being correctly obeyed in the evaluation run. To get faster speeds for time-sensitive work, we offer priority processing in the API and fast mode in Codex.
Models without reported output tokens, latency or cost are plotted as horizontal dotted lines.
Author
OpenAI
Keep reading
GPT-5.6 is now the preferred model in Microsoft 365 Copilot Product Jul 9, 2026
GPT-5.6: Frontier intelligence that scales with your ambition Product Jul 9, 2026
ChatGPT is now a partner for your most ambitious work Product Jul 9, 2026
Research
Latest Advancements
Safety
Products
- ChatGPT(opens in a new window)
- ChatGPT Business(opens in a new window)
- ChatGPT Enterprise(opens in a new window)
- ChatGPT for Education(opens in a new window)
- Codex
- Release Notes
API Platform
Business
Developers
- Apps SDK(opens in a new window)
- Open Models
- Docs(opens in a new window)
- Resources(opens in a new window)
- Developer Forum(opens in a new window)
Company
Support
More
Terms & Policies
(opens in a new window)(opens in a new window)(opens in a new window)(opens in a new window)(opens in a new window)(opens in a new window)(opens in a new window)
OpenAI © 2015–2026 Your privacy choices
English United States
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文