aifollow.news
返回 IT之家 科技新闻
IT之家 科技新闻· · 原发布时间 精选AI 评分60

Anthropic 推出 Claude Fable 5.2 和 Claude Mythos 5.2,后者限审核机构使用

AI 辅助摘要

两款模型基于同一底层模型:Fable 5.2 面向普通用户,已在 AWS、Google Cloud 和 Microsoft Azure 等平台开放;Mythos 5.2 面向通过审核的网络安全和生命科学机构,须申请使用。Fable 5.2 每百万词元输入和输出分别收费 8 美元和 40 美元,低于 5.1 版的 10 美元和 50 美元。

推荐理由

普通用户可关注 Fable 5.2 的开放渠道和较前代降低的价格;需要高风险专业能力的机构则面临 Mythos 5.2 的审核门槛。

正文 · AI 翻译

IT Home reported on November 10 that Anthropic officially announced the launch of Claude Fable 5.2 and Claude Mythos 5.2 yesterday (November 9). The two models are based on the same underlying model. The former is intended for general users and includes cybersecurity and biosafety safeguards, while the latter is available to vetted cybersecurity and life sciences organizations for high-risk professional applications.

In terms of pricing, Claude Fable 5.2 costs $8 per 1 million input tokens (IT Home note: approximately 53.4 yuan at the current exchange rate) and $40 per 1 million output tokens (approximately 267.2 yuan at the current exchange rate). By comparison, Claude Fable 5.1 costs $10 (approximately 66.8 yuan at the current exchange rate) per 1 million input tokens and $50 (approximately 334 yuan at the current exchange rate) per 1 million output tokens.

As for generation speed, Anthropic says the new Claude Fable 5.2 model generates output up to 30% faster than Claude Fable 5.1. In the optional “fast mode,” generation speed increases to as much as 2.5 times the standard speed, but at a correspondingly higher price.

In pre-release testing, Fable 5.2 completed several notable tasks: it designed three functional electronic devices, including a voice assistant hub, a USB conference speakerphone, and a handheld virtual pet; coordinated more than 1,000 subagents to build a prototype operating system from scratch that could boot on a laptop within 8 days; and, using about 1,000 agents, completed a Lean formal proof of Fermat’s Last Theorem in 30 hours, far faster than the previous record of 11 days.

Scientific research is also a priority. Anthropic says Fable 5.2 solved two open problems related to black hole physics and holographic duality. Mythos 5.2 proposed a new method for predicting the overall structure of adeno-associated virus capsids, successfully modeling 28 of 34 capsids in testing, which could accelerate gene therapy research and development.

On safety, Anthropic says the new models perform at least as well as, and in some cases better than, their predecessors on most alignment tests, with a significant decline in sandbox escape attempts. Fable 5.2 will continue to block penetration testing, exploit generation, and most biological research tasks. It also strengthens anti-distillation protections, which the company says can reduce the amount of data obtained through distillation by 90% to 99%. However, the company acknowledges that, as model capabilities grow, existing evaluations still struggle to fully predict risks in real-world settings.

Fable 5.2 is now available on platforms including AWS, Google Cloud, and Microsoft Azure. Access to Mythos 5.2 requires an application through the Cyber or Life Sciences verification program.

发现内容有误?提交纠错