Anthropic 推出全新支出分析工具：帮企业算清 AI 账本与 ROI，筑牢企业级护城河

Anthropic 本周向投资者披露的财务数据表明，其 AI 编程工具 Claude Code 取得了这家前沿实验室所期望的商业成功。该公司在第二季度录得营业利润，净收入留存率（Net Revenue Retention，NRR）达到 900%——这意味着同一批客户在 2026 年第三季度的支出相较于 2025 年第三季度增长了 9 倍。

Anthropic 一直致力于让企业客户保持高黏性，并于周三宣布推出一套全新的报告工具，旨在让这些客户更加精细地了解其 Token 支出所带来的价值。

借助内部数据分析和员工调研，管理者可以查看团队在特定时间范围内的工作情况、产出成果、消耗成本以及工作流程受阻的具体环节。Anthropic 表示，默认情况下不会将这些数据用于模型训练。这些报告还将就如何改善成果提供优化建议。

图片来源： Anthropic

图片来源： Anthropic

图片来源： Anthropic

图片来源： Anthropic

图片来源： Anthropic

图片来源： Anthropic

图片来源： Anthropic

这显然是为了展示 Anthropic 工具的投资回报率（ROI），并向客户指明如何进一步提升回报——这也是大企业客户的核心关切之一。此外，Anthropic 还推出了团队共享预算池功能，以帮助各团队更好地管理 Token 支出。

在探索如何最佳部署 AI 工具的过程中，企业一直在寻找这类数据洞察。虽然风险投资支持的初创企业的高频使用带动了 Token 消耗，但大多数《财富》500 强企业尚未从部署 AI 中看到显著收益。

软件开发者工具开发商 Jellyfish 昨天基于 27.6 万名客户对 Agent 编程工具的使用情况发布了最新报告。调查显示，Claude Code 以及基于 Anthropic 模型构建的其他工具依然是使用最广泛的工具，在提交审查的所有代码变更提案中占到了近三分之一。更为关键的是，今年年初以来，企业在这些工具上的支出增长了 10 倍，其成本目前约占整体工程研发成本的 2%。

“即使你认为 2% 的占比依然很小，但这 10 倍的增长也是大多数首席财务官（CFO）始料未及的，”Jellyfish 首席科学家 Nicholas Arcolano 在接受 TechCrunch 采访时表示，“我们确实已经到了这样一个阶段：大家开始意识到，仅靠盲目砸钱是不够的。你现在必须去优化它，让它真正发挥效用。”

Anthropic 的新报告功能正切中了这一挑战。尽管 Anthropic 在企业级客户中处于领先地位，但究竟是什么让一款 Agent 编程工具具备真正的客户黏性，目前仍难以定论。今年夏天 OpenAI 推出的 Astra 模型看似对 Anthropic 构成了潜在威胁，但迄今为止并未放缓 Anthropic 的企业采用数据增速。

“我们花了很长时间才在数据中看到 Codex 真正冒头，而随着 Astra（OpenAI 的最新模型）的推出，它的数据确实有所上升，”Jellyfish AI 产品负责人 Nikolas Albarran 对 TechCrunch 说道。他接着指出，如果浏览社交平台 X，你会明显感到竞争异常激烈。“硅谷的每个人都在说‘噢，Codex 更好用’。”但现实中，Albarran 补充道：“Anthropic 凭借出色的企业级落地策略占得了先机，且拥有自身先发优势的壁垒——他们的产品确实好用而且运行稳定，很多团队投入了大量精力围绕它建立标准化的工作流。企业目前没有足够充分的理由去切换平台。”

这不仅仅是模型能力和工作流的问题。商业现实意味着，一旦某家前沿实验室与客户建立了深度绑定，客户想要切换的难度要远大于单个程序员或小型组织。Albarran 特别提到了各大 AI 实验室目前对现场部署工程（Forward-Deployed Engineering，简称 FDE）团队的重金投入，这些团队会将 AI 顾问派遣进驻客户内部，帮助客户快速起步。对于 Anthropic 的客户而言，这些全新的报告工具可以被视为一种“轻量级 FDE（FDE-lite）”，提供了一种更轻巧的即插即用型协助。

“我们现在所处的阶段已经足够成熟，一旦你深度接入，由于前期已经投入了大量精力，后续迁移的难度会越来越大，”他表示，“即便各大模型大体相似，但它们在细节上仍各有不同，在处理 agent.md 文件等特定机制时也各有各的独特性，因此迁移成本绝不是零。”

The financials Anthropic shared with investors this week suggest that its AI coding tool, Claude Code, is every bit the business success the frontier lab wanted. The company recorded an operating profit in the second quarter, and 900% net revenue retention — that is, spending from the same customers in Q3 2025 grew nine times in Q3 2026.

Anthropic has focused on getting enterprise users to stick around and Wednesday announced a new suite of reporting tools designed to give those customers a more granular understanding of the value they get from their token spend.

Using internal analytics and employee surveys, managers can view a team’s work over a given date range, what the team produced, how much it cost, and where its workflows broke down. Anthropic says it won’t use this data for model training by default. The reports will also offer recommendations on how to improve the results.

Image Credits: Anthropic Image Credits: Anthropic Image Credits: Anthropic Image Credits: Anthropic Image Credits: Anthropic Image Credits: Anthropic Image Credits: Anthropic

This is a clear attempt to demonstrate the return on investment of Anthropic’s tools and to show customers how to improve it — a major concern among large customers. Anthropic is also creating pooled group budgets to allow teams to better manage token spending.

Businesses have been hunting for that kind of information as they figure out how best to deploy AI tools. While usage among venture-backed firms is driving token spend, most of the Fortune 500 are not seeing major gains from deploying AI.

Jellyfish, a company that makes tools for software developers, released a new report yesterday based on the use of agentic coding tools by 276,000 of its customers. Claude Code and other tools built on Anthropic models remain the most widely used in the survey, responsible for nearly a third of all proposed code changes submitted for review. More importantly, spending on those tools has grown 10x since the beginning of the year, with their cost now equal to about 2% of total engineering cost.

“Even if you’re saying 2% is still small, that 10x growth is growth that most CFOs didn’t plan for,” Nicholas Arcolano, the chief scientist at Jellyfish, told TechCrunch. We’re certainly getting to the point where people are realizing, okay, it’s not enough to just spend a bunch. You now have to optimize it, make it work.”

Anthropic’s new reports speak to that challenge. Anthropic leads among enterprise customers, but what makes an agentic coding tool truly sticky is difficult to identify. The release of OpenAI’s Astra model this summer seemed like a potential threat to Anthropic, but so far it hasn’t slowed Anthropic’s adoption numbers.

“It took a long time for us to see Codex really pop in our data at all, and it has come up with Astra [OpenAI’s latest model],” Nikolas Albarran, the AI product lead at Jellyfish, told TechCrunch. Looking at X, he continued, you’d get the distinct impression that the competition is more heated. “Everyone in the Valley is like, ‘Oh, Codex is better.’” In reality, Albarran continued, “Anthropic was just there with great enterprise motions, and they had an incumbent weight of their own — this stuff just worked and worked well, and a lot of people just put a lot of effort into standardizing workflows around it. And there’s not a big enough reason to switch.”

It’s not just a question of models and workflows, either. The realities of business mean that once a frontier lab has gotten its hooks into a customer, switching is harder than it is compared to an individual programmer or small organization. Albarran points to the investment all the major AI labs are making in forward-deployed engineering (FDE) organizations, teams that embed AI consultants inside customer organizations to help them get off the ground. These new reporting tools can be seen as a kind of FDE-lite, a lighter version of that hands-on help, for Anthropic customers.

“We’re at a point now where it’s mature enough, where once you’re in, it’s going to be harder and harder to switch because you’ve spent the effort,” he said. “Even if they’re all mostly the same, they’re all going to be slightly different, have their own slight quirks of how they handle agent.md files, etc., and so that cost to switch is not zero.”