内容来源：https://aibusiness.com/foundation-models/anthropic-launches-claude-sonnet-4-5

内容总结：

近日，人工智能公司Anthropic正式发布新一代生成式AI模型Claude Sonnet 4.5。该模型在编程、逻辑推理与计算机操作方面表现突出，能够直接操作浏览器、填写电子表格及完成网站导航等任务，适用于金融、科研和网络安全等领域。

值得注意的是，新模型采用了供应商自研的"AI安全三级防护"机制，通过过滤器和分类器技术防范有害内容输入与输出。此次更新正值AI代理市场持续升温之际，距微软宣布将Claude系列模型纳入365 Copilot平台仅隔一周，显示出行业竞争日趋激烈。

市场研究机构Gartner分析师指出，Anthropic通过提交OSWorld计算机操作、MMMLU多语言问答等基准测试结果，意在强化其技术领先地位。与此同时，企业正着力攻克AI安全与幻觉问题等系统性难题。

在商业化层面，分析师认为这家独立AI厂商面临渠道建设挑战。与法国Mistral AI类似，Anthropic正着力构建开发者生态，鼓励用户直接在其平台进行推理运算并收取费用。目前已有Glean等合作伙伴宣布将在无代码代理构建器中集成新版模型。

（注：本文内容基于公开信息客观呈现，不构成任何投资建议）

中文翻译：

由谷歌云赞助
选择首个生成式AI应用场景
开展生成式AI应用时，应首先关注能优化人类信息交互体验的领域。

在供应商AI安全三级防护机制下发布的该模型，可有效防范危险输入与输出内容。

生成式AI基础模型制造商Anthropic本周一正式推出Claude系列最新版本Claude Sonnet 4.5。该模型在编程、逻辑推理与计算机操作场景表现尤为出色。

Anthropic表示，Sonnet 4.5适用于构建复杂智能体，精通计算机操作——能直接在浏览器中工作、浏览网页、填写电子表格并完成任务。厂商称该模型特别适合金融、研究与网络安全领域的应用。

新版模型在发布时搭载了厂商的AI安全三级防护体系，包含可检测有害输入与输出内容的过滤器与分类器。

距离Opus 4与Sonnet 4及三级安全防护体系发布仅四个月，Anthropic便推出了Sonnet 4.5。本次升级正值智能体AI市场持续扩张之际，该技术不仅对厂商意义重大，对企业用户也日益重要。

此次产品发布前一周，微软选择采用Claude模型为其365 Copilot生成式AI平台提供支持（该平台同时集成Anthropic竞争对手OpenAI的模型），这一举措使Anthropic获得广泛关注。

高德纳分析师阿伦·钱德拉塞卡指出，通过Sonnet 4.5，Anthropic意在向业界展示其强大的编程模型实力。
"鉴于近期来自OpenAI等公司的竞争压力...他们实际提交了大量基准测试数据以证明其领先地位。"

厂商已将升级版模型提交至多个基准测试平台：检验计算机操作能力的OSWorld、多语言问答测试MMMLU以及视觉推理评估MMMU。

Futurum Group分析师布拉德利·希明表示，Anthropic同时致力于解决长期困扰生成式AI技术的安全性与幻觉问题。
"他们正着力解决AI领域某些更系统性的难题，这些正是当前市场...真正忽视的痛点。"

除核心模型外，Anthropic还推出了Claude Agent SDK等新工具，以及周一发布的研究预览项目"Imagine with Claude"。

据官方介绍，Imagine功能可使Claude无需预写代码即可生成软件。未来五天内，该功能将向Claude高端订阅服务Max的用户开放。

钱德拉塞卡指出，Anthropic面临的关键挑战在于为Sonnet 4.5及其他新产品构建市场化战略。
他补充说明："截至目前，企业客户拓展主要依赖其他软件厂商合作。我十分期待看到他们建立更直接的市场推广体系，从而优化商业机会变现能力，获取更高利润空间。"

但希明认为，在动荡的生成式AI市场中，Anthropic可能不得不依赖合作伙伴维持生存。
"作为独立厂商（非超大规模云服务商），Anthropic处于劣势。"他举例说明行业整合趋势，如Databricks于2023年收购MosaicML。

Anthropic的市场策略与法国Mistral AI类似，旨在构建开发者生态，使其能直接在Anthropic平台上开发智能体工具与应用。

希明补充道，该厂商似乎更注重引导用户在其平台进行推理计算并收取费用，而非获取用户数据。

部分厂商已开始集成Sonnet 4.5。AI服务商Glean周一宣布将在其无代码智能体构建器中支持该模型。

您可能还喜欢

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The model was released under the vendor's AI Safety Level 3 protection, which helps prevent dangerous inputs and outputs.
AI foundation model maker Anthropic on Monday introduced Claude Sonnet 4.5, the latest iteration of its popular Claude line of generative AI models. It is designed to perform most effectively at coding, reasoning and computer use.
Anthropic said Sonnet 4.5 is suitable for building complex agents and is proficient at using computers -- capable of working directly in a browser, navigating websites, filling out spreadsheets and completing tasks. The model is well-suited for applications in finance, research and cybersecurity, the vendor said.
Anthropic released the new model under the vendor's AI Safety Level 3 protections, which include filters and classifiers that detect harmful inputs and outputs.
The vendor unveiled Sonnet 4.5 four months after it came out with Opus 4 and Sonnet 4, along with the Level 3 safety protections. The updated model arrives in a market in which agentic AI continues to grow and hold significance not only among vendors but also among enterprises.
The product release also comes the week after the vendor drew considerable attention when Microsoft picked Claude models to power the tech giant's 365 Copilot generative AI platform, alongside models from Anthropic rival OpenAI.
With Sonnet 4.5, Anthropic is seeking to remind the world that it has a powerful coding model, said Arun Chandrasekaran, an analyst at Gartner.
"In light of some of the recent competition … obviously from OpenAI and others ... they've actually submitted a lot of benchmarks to illustrate that leadership," Chandrasekaran said.
The vendor submitted the updated model to benchmarks such as OSWorld, which tests computer use, MMMLU for multilingual question and answer, and MMMU for visual reasoning.
Bradley Shimmin, an analyst with Futurum Group, said Anthropic is also trying to address a problem that has long plagued generative AI technology: AI safety and hallucinations.
"They're trying to tackle some of these more systemic problems that we have with AI that the marketplace ... is really overlooking right now," Shimmin said.
Along with the model, Anthropic introduced new tools, including the Claude Agent SDK as well as a research preview it released Monday, Imagine with Claude.
With Imagine, Claude can generate software without the need for prewritten code, Anthropic said. The feature will be available to subscribers of Max, a high-tier subscription for Claude, for the next five days.
One key challenge for Anthropic is building a go-to-market strategy around Sonnet 4.5 and its other new releases, Chandrasekaran said.
Anthropic has relied up to now largely on other software vendors to reach enterprise customers, he noted.
"I'll be very intrigued to see how they build a more direct go-to-market approach, so that they can monetize the opportunities better, and they can have a higher margin market opportunity as well," Chandrasekaran continued.
However, Shimmin said Anthropic might have no choice but to rely on its partners to survive in the turbulent generative AI market.
"As an independent, meaning they're not a hyperscaler, Anthropic is at a disadvantage," he said, noting that some frontier model makers are getting acquired as part of an industry consolidation. For example, Databricks acquired MosaicML in 2023.
Anthropic's go-to-market strategy is similar to that of France-based Mistral AI. Anthropic is trying to build an ecosystem of developers that can build agentic tools and applications directly on Anthropic's platform.
Shimmin added that the vendor seems to be focused less on acquiring users' data and more on getting users to perform inference with its platform and charging them for it.
Some vendors have started integrating Sonnet 4.5. AI vendor Glean on Monday said it will support Sonnet 4.5 within its no-code agent builder.
You May Also Like