Ultramarine 1 Flash：前沿模型便宜100倍

Ultramarine 1 Flash：前沿模型便宜100倍
Ultramarine 1 Flash

原始链接: https://www.paperinstruments.com/ultramarine

Ultramarine 1 Flash 是一款用于金融、法律、咨询和运营等专业知识工作的 Fable 级模型，定位为比 Fable 便宜 60 倍。它在六个基准上比 GLM-5.3 Flash 少用 6.5% 的 token，同时提供更佳的工作成果。在成本-质量前沿上，每次任务成本为 0.13 美元，得分超过 Claude Fable 5 (max) 和 GPT-5.6 Sol (max)，与 Claude Opus 5.5 (max) 和 GPT-6 Sol (max) 的差距在三分以内。API 定价为每百万输入 token 0.15 美元，缓存输入 0.03 美元，输出 0.50 美元。性能基于内部 PI Knowledge Work 基准（已完成业务文档和实时系统）衡量；其得分与 Artificial Analysis Intelligence Index 强相关。该模型针对客户就绪的文本、视觉层级和专业服务规则进行了调优。模型权重即将推出。

Hacker News 上的一篇帖子强调了“Ultramarine 1 Flash”，被描述为便宜 100 倍的前沿模型。一位评论者认为，即使它无法一次性完成不常见任务，反复尝试仍然经济，建议将 AI 视作不知疲倦的实习生并用指标优化。他们引用了大规模智能体努力和 Anthropic 的工程声明。发帖者还分享了发布视频，页面还包含 Y Combinator 申请推广。

Ultramarine 1 Flash 被定位为面向专业知识工作的智能体模型，基于 GLM-5.3 Flash 通过大规模强化学习训练，据报道在六项知识工作基准测试中有所提升。Hacker News 评论者指出其 GLM 底座，称赞模型及相关工具，并质疑基准测试与成本图表中的遗漏。其他人讨论编码性能，其中一位评论者指出 DeepSWE 和 TerminalBench 结果略低于 GLM 5.3 Flash，暗示编码可能并非重点。关于“前沿”是否指成本/性能而非绝对顶级模型，也存在争论。一些评论者认为其在办公场景具有潜力，尤其在文档处理与决策工具的配合方面。

Built for professional work

Improved token efficiency

Across the six benchmarks, Ultramarine 1 Flash generates 6.5% fewer tokens than its base, GLM-5.3 Flash, taking a shorter path to better work.

Output tokens generated across all six benchmarks, reasoning included. Where GLM-5.3 Flash’s score is a public one, its tokens come from our own runs closest to that score.

Fable-class model for cents per task

Ultramarine 1 Flash sits on the cost-quality frontier at the most attractive price point. At $0.13 per task, it outscores Claude Fable 5 (max) and GPT-5.6 Sol (max), at $8.19 and $2.09 per task, and comes within three points of Claude Opus 5.5 (max) and GPT-6 Sol (max) at $6.96 and $1.26 per task. We measure this on PI Knowledge Work, a carefully built internal benchmark whose scores correlate strongly with the Artificial Analysis Intelligence Index (r = 0.81).

Input: $0.15
Cached input: $0.03
Output: $0.50

Per million tokens on the Ultramarine API, the same as GLM-5.3 Flash.

PI Knowledge Work is our internal benchmark of professional work in finance, law, consulting, and operations. Its tasks ask the model to deliver finished Excel workbooks, PowerPoint decks, and Word documents, and to work inside live business systems such as ERP and MES.

A rubric written for each task scores the result on substance, accuracy, and presentation. Every model works in the same harness, and costs are at each developer's official API pricing. Read our technical report for more on the kind of professional data and evaluations that trained this model.

Crafted with a sense of client readiness

Ultramarine 1 Flash writes clean, simple, straightforward prose and has a strong sense of visual and information hierarchy. It is built to adhere to the same rules required of analysts in professional services.