内容来源：https://www.engadget.com/2203102/openai-starts-previewing-gpt-56-and-its-three-variants/

内容总结：

OpenAI推出GPT-5.6预览版，率先面向“可信合作伙伴”小范围开放

OpenAI于今日正式发布GPT-5.6系列模型的有限预览版，首批用户仅限于一小部分“受信任的合作伙伴”。该系列共包含三个版本，并计划在未来几周内逐步扩大覆盖范围。

据OpenAI介绍，三个版本各有侧重：其中性能最强的“Sol”版本被誉为该公司迄今为止最强大的模型；面向日常使用的“Terra”版本在性能上与GPT-5.5相当，但成本降低了一半；而“Luna”版本则是价格最低的模型，主打经济实惠。

值得注意的是，OpenAI在发布前已向美国政府展示了GPT-5.6的相关能力。公司表示，此次小范围预览也是应政府要求，合作伙伴名单已与政府共享。OpenAI在公告中强调，“我们不认为这种政府审查流程应成为长期常态”，当前此举仅为短期过渡，目的是尽快将最新模型系列面向公众开放。

此前，特朗普总统于本月初签署了一项AI网络安全行政令，要求企业在公开发布最强大模型前30天，自愿提交给政府审查。据《纽约时报》报道，实际上早在行政令签署前，OpenAI、Anthropic、谷歌、xAI和微软等公司就已向政府提供早期访问权限，仅Meta公司尚未配合，美国政府正持续施压其提交模型接受评估。

在技术层面，GPT-5.6引入了“最大推理努力”功能，允许“Sol”模型有更多时间进行深度思考。该版本在网络安全领域表现尤为突出，可帮助用户发现并修复漏洞。OpenAI称，团队已花费数周时间挖掘模型弱点并加固防护，使其能够抵御真实世界攻击。

针对所有版本，OpenAI均加强了安全措施，以应对实际对抗压力。公司专门训练了GPT-5.6拒绝“被禁止的网络安全协助”请求，包括破解模型的行为。为此，OpenAI投入了70万GPU小时来寻找通用破解方法，并承诺建立“快速响应流程”，对新发现的破解手段进行复现、评估、优先级排序和修复。

业内分析认为，OpenAI此次对破解防护的极度重视，可能与竞争对手Anthropic的遭遇有关。几周前，Anthropic因政府指令暂停了其“Mythos 5”和“Fable 5”模型的全部访问权限。据称，亚马逊等公司向当局报告称，这些模型可能被破解并用于恶意目的。目前，Anthropic已获美国政府许可，恢复向特定组织重新部署“Mythos”模型。

在定价方面，GPT-5.6 Sol版本为每百万输入令牌5美元、每百万输出令牌30美元，远低于此前“Fable”模型（输入10美元、输出50美元）的价格。Terra版本为输入2.5美元、输出15美元，Luna版本则为输入1美元、输出6美元。

中文翻译：

OpenAI 面向“一小批受信合作伙伴”推出 GPT-5.6 有限预览版
该模型的三个变体将在未来几周内更广泛地发布。

OpenAI 已开始向有限数量的受信合作伙伴预览其 GPT-5.6 系列，该系列将提供三个版本。该公司表示，其变体 Sol 是迄今为止最强大的模型，而 Terra 适用于日常使用，性能与 GPT-5.5 相当，但成本却便宜一半。最后一个变体 Luna 是该公司成本最低的模型。OpenAI 计划在未来几周内的某个时间点对其进行广泛发布。

该公司在今天之前已向美国政府预览了 GPT-5.6 及其能力。同样应美国政府的要求，该公司正向一小批“其参与情况已告知政府”的受信合作伙伴预览该模型。OpenAI 在公告中写道：“我们认为这种政府准入流程不应成为长期的默认做法。”该公司表示，目前采取这一“短期措施”是为了确保能尽快向公众发布其最新模型系列。

特朗普总统本月早些时候签署了一项人工智能网络安全行政令，要求企业在公开发布其最强大模型前 30 天，将其提交给政府进行自愿审查。据《纽约时报》近期报道，OpenAI、Anthropic、Google、xAI 和微软甚至在特朗普签署该行政令之前，就已向政府提供了其最新模型的早期访问权限。Meta 是唯一拒绝的公司，据报道美国政府一直在敦促其提交 AI 模型以供评估。

GPT-5.6 引入了一种“最大”推理力度，使 Sol 能有更多时间进行深度推理。Sol 也是 OpenAI 在网络安全方面能力最强的模型，是帮助用户发现和修复漏洞的最佳选择。OpenAI 表示，Sol 针对高风险活动和敏感请求加强了保护。该公司还表示，他们已经花费了数周时间寻找其弱点，并针对现实世界攻击进行加固。

不过，该公司对所有变体都设置了安全防护措施，以确保它们能承受真实的对抗压力。此外，OpenAI 训练了 GPT-5.6 以拒绝“被禁止的网络协助”，包括试图破解模型的行为。该公司投入了 70 万 GPU 小时来寻找通用破解方法以制定应对措施，并承诺实施“快速响应流程，以重现、评估、确定优先级并修复新发现的破解方式”。

OpenAI 对防范破解问题的关注很可能源于 Anthropic 的遭遇。几周前，Anthropic 在接到政府指令后暂停了对其 Mythos 5 和 Fable 5 模型的所有访问。虽然该公司没有明确说明，但据报道亚马逊等公司已通知当局，其模型可能被破解并用于恶意目的。不过，由于美国政府刚刚批准 Anthropic 将 Mythos 重新部署给特定组织，该公司已开始解除访问限制。

OpenAI 将 GPT-5.6 Sol 的定价定为每百万输入 token 5 美元，每百万输出 token 30 美元，远低于 Fable 仍可访问时的价格（同等 token 数量，输入 10 美元，输出 50 美元）。Terra 的定价为输入 2.50 美元，输出 15 美元，而 Luna 的定价为输入 1 美元，输出 6 美元。

英文来源：

OpenAI launches a limited preview of GPT-5.6 for a 'small group of trusted partners'
The model’s three variants will be available more broadly in the coming weeks.
OpenAI has started previewing its GPT‑5.6 series, which will be available in three versions, to a limited number of trusted partners. The company says the variant Sol is its strongest model yet, while Terra is for everyday use and has a similar performance to GPT‑5.5 despite being twice as cheap. Luna, the last variant, is the company's lowest cost model. OpenAI plans to give them a broad release sometime in the coming weeks.
The company gave the US government a preview of GPT‑5.6 and its capabilities before today. It's also by the administration's request that it is previewing the model to a small group of trusted partners "whose participation has been shared" with the government. "We don't believe this kind of government access process should become the long-term default," OpenAI wrote in its announcement. It said it's taking the "short-term step," for now, because it ensures it can release its latest model series to the public soon.
President Trump signed an AI cybersecurity order earlier this month, which asks companies to present their most powerful models for voluntary government review 30 days before making them publicly available. According to a recent report by The New York Times, OpenAI, Anthropic, Google, xAI and Microsoft have been giving the government early access to their latest models even before Trump signed the order. Meta was the only holdout, and the US government has reportedly been urging it to submit its AI models for evaluation.
GPT‑5.6 introduces a "max" reasoning effort, which gives Sol more time to reason deeply. Sol is also OpenAI's most capable model for cybersecurity and is the best option to help users find and fix vulnerabilities. OpenAI says Sol comes with strengthened protections for high-risk activities and sensitive requests. It also says that the company had spent several weeks finding its weaknesses and fortifying it against real-world attacks.
The company put safeguards on all the variants, however, to make sure they hold up to real adversarial pressure. In addition, OpenAI trained GPT-5.6 to refuse "prohibited cyber assistance," including attempts at jailbreaking the model. It spent 700,000 GPU hours to find universal jailbreaks to develop measures against them, and it pledges to implement a "rapid-response process to reproduce, assess, prioritize, and remediate newly discovered jailbreaks."
OpenAI's focus on jailbreak prevention likely stems from what happened to Anthropic. A couple of weeks ago, Anthropic suspended all access to its Mythos 5 and Fable 5 models after a directive from the government. While the company didn't say it outright, Amazon and other companies had reportedly notified authorities that its models could be jailbroken and used for malicious purposes. It has started lifting its access block, though, since US government has just given Anthropic permission to redeploy Mythos to a select group of organizations.
The company has priced GPT‑5.6 Sol at $5 per million input tokens and $30 per million output, much less than what Fable cost when it was still available. ($10 for input and $50 for output for the same amount of tokens.) Terra costs $2.50 for input and $15 for output, while Luna costs $1 for input and $6 for output.