Sakana claims its AI model router Fugu Ultra v1.1 now beats Fable 5 without even including it in the pool Sakana声称其AI模型路由器Fugu Ultra v1.1甚至未将Fable 5纳入池子便已超越它

TL;DR

• Sakana AI released Fugu Ultra v1.1, an AI model router that distributes queries across a pool of top-tier public models.
• The update claims performance gains of up to 7.9 points over version 1.0, with significant improvements on ProgramBench and TerminalBench 2.1.
• Fugu Ultra v1.1 reportedly outperforms Anthropic's Fable 5, despite Fable 5 not being included in the router's selection pool.
• All performance metrics are self-reported by Sakana AI, with no independent verification currently available.
• Pricing remains unchanged at $5 per million input tokens and $30 per million output tokens.

Why It Matters

This development highlights the growing competitive pressure on standalone large language models from routing architectures that aggregate multiple specialized models. For practitioners, it underscores the importance of evaluating whether multi-model routing offers superior cost-performance trade-offs compared to single-state-of-the-art models like Fable 5. The lack of independent verification also serves as a critical reminder for researchers and engineers to validate vendor claims before integrating such systems into production pipelines.

Technical Details

• Architecture: Fugu Ultra v1.1 functions as a router that dynamically distributes user queries across a curated pool of publicly available, top-tier AI models.
• Performance Metrics: The update reports performance increases of up to 7.9 points compared to v1.0, specifically noting major jumps on ProgramBench and TerminalBench 2.1.
• Benchmarking: The router claims superiority over Anthropic's Fable 5 across most benchmarks, achieving this without Fable 5 being part of its underlying model pool.
• Integration & Infrastructure: A new Claude Code-compatible endpoint allows direct terminal access, building on previous availability via OpenRouter and Vercel.
• Operational Constraints: Adding new models to the pool requires approximately two weeks of training and evaluation, and the service currently excludes the EU and EEA due to regulatory concerns.

Industry Insight

The success of model routers like Fugu suggests a strategic shift toward agentic orchestration layers that leverage diverse model strengths rather than relying on monolithic models. Developers should consider implementing routing mechanisms to optimize for specific tasks (e.g., coding vs. reasoning) while managing costs, provided they can verify the claimed performance gains. Additionally, the regulatory exclusion of the EU market indicates that compliance with GDPR and emerging AI regulations will remain a significant barrier to global deployment for many AI infrastructure providers.

TL;DR

• Sakana AI发布Fugu Ultra v1.1，声称在ProgramBench和TerminalBench 2.1等基准测试中性能提升高达7.9分。
• 该路由器模型在未将Anthropic的Fable 5纳入候选池的情况下，综合表现仍优于Fable 5。
• 定价维持不变（输入$5/百万token，输出$30/百万token），新增Claude Code兼容端点以支持终端直接调用。
• 所有性能数据均由Sakana AI自行发布，目前尚无独立的第三方验证结果。

为什么值得看

对于AI从业者而言，Fugu Ultra v1.1展示了“模型路由器”架构在整合多个顶级开源或闭源模型时的潜在优势，为降低单一模型依赖提供了新思路。同时，其宣称超越主流专有模型（如Fable 5）的表现，若经证实，将对现有API聚合服务市场格局产生重要影响。

技术解析

• 核心机制：Fugu Ultra是一个AI模型路由器，通过算法将用户查询动态分配给池中可用的顶级模型，旨在优化响应质量和成本效率。
• 性能基准：相比v1.0版本，v1.1在编程能力（ProgramBench）和终端操作（TerminalBench 2.1）方面取得显著进步，最大增幅达7.9分。
• 集成与接口：新增Claude Code兼容端点，允许开发者直接在终端环境中调用Fugu服务；此前已接入OpenRouter和Vercel等平台。
• 更新周期：模型池的更新需要约两周时间进行新模型的训练和评估，以确保加入池中的模型符合性能标准。

行业启示

• 路由器架构的竞争力：如果未经过微调的模型路由策略能击败特定专有模型，表明智能调度算法可能成为继大模型本身之后的下一个关键竞争壁垒。
• 信任与验证机制：由于缺乏独立验证，行业需警惕厂商自报数据的偏差，推动建立标准化的第三方基准测试流程至关重要。
• 合规与市场局限：Sakana因GDPR问题未服务欧盟市场，这凸显了全球AI服务在数据隐私法规上的碎片化挑战，限制了其潜在的市场规模。