Anthropic 发布 Opus 5

周五，Anthropic 发布了 Opus 5 模型，这是其长期以来的重型模型系列的最新版本。尽管其规模小于 Fable 5，但该模型价格更低、限制也更少，因此可能在大多数使用场景中成为更合适的选择。

值得注意的是，Opus 5 在公告所列出的多项基准测试中实际上优于 Fable 5。

Opus 5 在 Opus 4.8 于 5 月 28 日上线仅两个月后便发布。Mythos 5、Fable 5 和 Sonnet 5 均于 6 月发布，这意味着轻量级 Haiku 模型仍是唯一等待升级至 5 系列的模型。

Anthropic 在一篇宣布新模型的文章中强调，Opus 5“在验证自身工作以及审慎迭代直至成功方面强得多”。该公司援引基准测试称，除其他示例外，Opus 5 曾针对一个不完整的提示词自行编写了计算机视觉流水线。

关键在于，Opus 5 也摆脱了自 Fable 发布以来一直困扰后者的许多限制。与其前代产品一样，Opus 5 不受适用于 Fable 和 Mythos 的30 天数据保留政策约束；这一政策此前曾引发部分重视隐私的用户担忧。

Opus 仍然保留了重要的安全防护措施，尤其针对漏洞利用代码生成和渗透测试等网络安全任务。例如，Opus 5 的防护措施会阻止其被用于扫描软件二进制文件中的漏洞；但允许其在源代码中搜索漏洞，因为后者更可能被用于防御性目的。

总体而言，Anthropic 预计，相比 Fable 5，这些分类器在 Opus 5 上的触发频率将低 85%，这反映出对能力较弱模型采取了更宽松的管控方式。

Anthropic 还推出了一项新工具，以便在防护机制触发时降低其造成的干扰。用户现在可以选择加入一项名为 Automatic Fallbacks 的测试功能：当提示词触发安全分类器时，该功能会自动将请求路由至能力较低的模型。因此，启用这一设置的 API 用户将获得可用的回复，而非错误信息。

On Friday, Anthropic launched its Opus 5 model, the newest version of its long-standing heavyweight model. While smaller than Fable 5, the model will be both cheaper and less restrictive than Fable, likely making it preferable in most use cases.

Notably, Opus 5 actually outperforms Fable 5 on a number of benchmarks included in the announcement.

Opus 5 is launching only two months after Opus 4.8, which became available on May 28. Mythos 5, Fable 5, and Sonnet 5 all launched in June, leaving only the lightweight Haiku model still waiting for an upgrade to the 5 series.

In a post announcing the new model, Anthropic emphasized that Opus 5 was “much stronger at verifying its work and iterating carefully until it succeeds,” citing benchmark testing, in which Opus 5 wrote its own computer vision pipeline in response to an incomplete prompt, among other examples.

Crucially, Opus 5 is also free from many of the restrictions that have dogged Fable since its release. Like its predecessor, Opus 5 is not subject to the 30-day data retention policy that covers Fable and Mythos, which had raised concerns among some privacy-conscious users.

There are still meaningful safeguards on Opus, particularly around cybersecurity tasks like exploit generation and penetration testing. For instance, Opus 5 safeguards prevent it from being used to scan for vulnerabilities in a software binary, although it is permitted to search for vulnerabilities in source code, since the latter task is more likely to be used for defensive purposes.

Broadly, Anthropic expects these classifiers to engage 85% less often for Opus 5 than they will for Fable 5, a reflection of the lighter touch given to the less capable model.

Anthropic is also rolling out a new tool to make the safeguards less disruptive when they do engage. Users can now opt in to a beta feature called Automatic Fallbacks, which will automatically route requests to a less powerful model when a prompt triggers the safety classifier. The result is that API users with the setting engaged will get a functional response instead of an error message.