Anthropic推出了Claude Fable 5.1和Claude Mythos 5.1两款新型大语言模型123。Fable 5.1在保持原有输入输出定价不变的基础上实现了显著成本优化，缓存读取价格降至每百万令牌$0.25，相比其他Claude模型下降75%12。新版本在复杂编程任务中可节省高达45%的成本，整体成本相比前代Fable 5降低约25%1。Fable 5.1同时改进了安全防护措施，生物安全防护的良性请求触发次数减少85%，网络安全防护的平均每次会话干预次数减少约60%1。
Claude Mythos 5.1采用相同的基础模型但配置更高的安全防护等级，仅向经过验证的网络安全和生命科学专业人士提供1。在实际应用中，Mythos 5.1在蛋白质设计竞赛中表现突出，三个目标的结合亲和力比最佳设计高10倍，命中率达近50%，远超业界典型的10-15%1。该模型还通过编写自定义GPU核心将7个开源深度学习模型的速度提升至2.5倍，生物学GPU成本可降低30-60%1。在其他科学领域，Fable 5.1为金星三分之一区域创建了新的高分辨率高程图，分辨率达2-3公里（相比之前的10-20公里），高度准确性提升25%1。
两款模型均已通过广泛的化学生物风险、网络风险、代理安全和对齐性测试1，符合EU AI Act要求，并采用Anthropic的统计文本水印进行标记12。新版本引入了每消息努力级别、转向系统消息、工具调用间的进度更新和内容来源标记等功能改进2。Anthropic计划在今年秋季分阶段推出企业前沿防护措施，允许客户在自有云基础设施上存储数据1。Fable 5.1已从发布之日起在云平台和Anthropic API上提供3。
Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, positioning them as the industry's most advanced models for coding and knowledge work.1 Fable 5.1 reduces overall costs by approximately 25% compared to its predecessor, with cache read costs dropping 75% to $0.25 per million tokens, while input and output pricing remain unchanged at $10 and $50 per million tokens respectively.12 Batch processing is priced at $5 per million input tokens and $25 per million output tokens.2 The models have undergone extensive testing including chemical-biological risk assessments, cybersecurity evaluations, and alignment benchmarks.1
The new models demonstrate significant performance improvements across multiple domains. In scientific applications, Claude Mythos 5.1 achieved a 10-fold improvement in binding affinity for protein design targets with a near 50% hit rate, compared to the industry standard of 10-15%.1 The model also generated new high-resolution elevation maps covering one-third of Venus with 2-3 kilometer resolution—a substantial upgrade from the previous 10-20 kilometer standard—and improved elevation accuracy by 25%.1 Additionally, Mythos 5.1 accelerated seven open-source deep learning models up to 2.5 times faster through custom GPU kernel optimization, potentially reducing biological GPU costs by 30-60%.1
Security measures have been substantially refined in both versions.1 The Fable 5.1 implementation reduced false positive triggers in biosafety protections by 85%, while cybersecurity interventions decreased by approximately 60% per session.1 Mythos 5.1 is restricted to verified professionals and is available exclusively through two verification programs: the Cybersecurity Verification Program (CVP) and the Life Sciences Verification Program (LSVP), currently limited to U.S. organizations.1 Both models incorporate statistical text watermarking compliant with EU AI Act requirements and feature an AI-generated content transparency commitment signed July 8, 2026.12
Enterprise privacy protections are expanding with the Enterprise Frontier Safeguards feature, which allows customers to store data on their own cloud infrastructure and is set to roll out in phases beginning this fall.1 Anthropic stated that it "has never trained on enterprise data without explicit permission, and never will."3 A 30-day data retention period applies by default, with zero data retention available only with explicit Anthropic authorization.2
评论
还没有评论，欢迎留下第一条。