Anthropic 于 2026 年 6 月 9 日发布 Claude Mythos 5 与 Claude Fable 5，称两者能力高于 4 月限量推出的 Mythos Preview。Claude Mythos 5 仍仅供有限产业伙伴使用，且 Anthropic 表示正与美国政府协作推进。Claude Fable 5 则公开发布，使用相同底层模型，但在发布时加入防护栏，会阻挡许多与网路安全、生物学、化学相关的请求，并将它们改路由至较旧的 Claude Opus 4.8。
Anthropic 的策略反映出能力释放与风险控制之间的比例取舍：Mythos 级模型可设计骇客工具、发现并利用新旧软体漏洞，因此 Project Glasswing 先让防御方取得准备时间。Diane Penn 表示，目前保护机制偏向谨慎，代表部分无害查询也可能被导向较弱模型。Anthropic 也会把疑似蒸馏 Claude Fable 5 的请求转至 Claude Opus 4.8，并计划在可信存取计划成熟后扩大不受限版本的取得范围。
价格与测试数据凸显商业压力。Claude Fable 5 与 Claude Mythos 5 对开发者收费为每百万输入 token 10 美元、每百万输出 token 50 美元，是 Anthropic 公开可用模型的 2 倍，但低于 Mythos Preview。Anthropic 称超过 1,000 小时红队测试未发现通用 jailbreak；然而，在 OpenAI 也于 4 月私下推出高阶网路安全模型、两家公司均已秘密申请 IPO 的背景下，Fable 5 的防护能否在公开环境中维持有效仍未确定。
Anthropic released Claude Mythos 5 and Claude Fable 5 on June 9, 2026, saying both are more capable than the limited Mythos Preview launched in April. Claude Mythos 5 remains available only to a limited group of industry partners, and Anthropic says it is coordinating the rollout with the US government. Claude Fable 5 is public and uses the same underlying model, but launches with guardrails that block many cybersecurity, biology, and chemistry requests by rerouting them to the older Claude Opus 4.8.
Anthropic’s strategy shows a trade-off between releasing capability and controlling risk: Mythos-level models can design hacking tools and discover and exploit vulnerabilities in both new and legacy software, so Project Glasswing first gives defenders time to prepare. Diane Penn said the current protection system errs on the side of caution, meaning some benign queries may also be sent to the weaker model. Anthropic will also reroute suspected attempts to distill Claude Fable 5 to Claude Opus 4.8, and it plans to widen access to unrestricted versions after a trusted access program matures.
Pricing and testing numbers highlight the business pressure. Claude Fable 5 and Claude Mythos 5 cost developers $10 per million input tokens and $50 per million output tokens, twice the price of Anthropic’s publicly available models but cheaper than Mythos Preview. Anthropic says more than 1,000 hours of red-teaming found no universal jailbreaks; however, with OpenAI also privately launching an advanced cybersecurity model in April and both companies confidentially filing for IPOs, whether Fable 5’s safeguards hold up in public use remains uncertain.