关键要点
- 据《华尔街日报》报道，Anthropic 预计第二季度营收达 109 亿美元，运营利润为 5.59 亿美元。
- 主要驱动力：编码工具需求旺盛，以及 Claude 的智能体任务处理能力。算力短缺促使该公司签署了新的数据中心协议。
- 新的分词器使得 Opus 4.7 对用户而言更加昂贵。与此同时，Anthropic 将其自身算力成本降至每美元营收 56 美分。
据《华尔街日报》报道，Anthropic 正接近其首个盈利季度，第二季度营收跃升 130%，达到 109 亿美元。
就在去年夏天，Anthropic 还告诉投资者不要指望在 2028 年之前实现年度盈利。如今，该公司预计 6 月季度的运营利润为 5.59 亿美元。《华尔街日报》报道称，这使得 Anthropic 的增长速度超过了疫情期间的 Zoom，或上市前的谷歌和 Facebook。
主要驱动力是 Anthropic 的编码工具，自今年年初以来，全球企业已大规模采用这些工具。除此之外，还有 Claude 的“智能体”式使用，即模型在较长时间内自主完成任务的场景。需求有时超过了算力容量，迫使 Anthropic 限制部分用户的访问权限，并签署新的数据中心协议，其中包括与埃隆·马斯克的 SpaceX 达成的协议。
AI 使用成本越来越高，Anthropic 从中受益
在成本方面，智能体 AI 是 token 消耗大户，尤其是在编码领域。新旗舰模型中还存在一个隐藏的价格驱动因素：Anthropic 的 Opus 4.7 每个 token 的成本与其前代产品相同，但每次请求的成本却显著增加。原因在于新的分词器将同一文本分割成的 token 数量最多增加了 47%。对于一轮 80 次交互的示例会话，开发者 Abhishek Ray 计算出的成本增加约为 20% 到 30%。OpenRouter 的分析证实，对于超过 2000 个 token 的提示词，实际成本增加了 12% 到 27%。
OpenAI 也在推行类似的激进定价策略：GPT-5.5 的标价比 GPT-5.4 翻了一番，输入 token 现为每百万个 5 美元，输出 token 为每百万个 30 美元。OpenAI 辩称其回复更短，但 OpenRouter 的一项研究显示，实际成本上涨了 49% 到 92%。
Anthropic 的效率也在提升。第一季度，该公司每获得 1 美元收入，在算力上的支出为 71 美分。本季度，这一数字预计将降至 56 美分。CEO Dario Amodei 在开发者大会上开玩笑说，收入增长已经变得难以应对。与 OpenAI 不同，Anthropic 主要通过投资者与 Google 和 Amazon 的交易来依赖更便宜的芯片，并且无需补贴一个充满免费用户的大型消费者业务。
Key Points
- Anthropic expects $559 million in operating profit for Q2 on $10.9 billion in revenue, the WSJ reports.
- The main drivers: coding tools seeing massive demand and Claude's agentic task processing. Compute shortages pushed the company to sign new data center deals.
- A new tokenizer makes Opus 4.7 more expensive for users. At the same time, Anthropic cut its own compute costs to 56 cents per dollar of revenue.
Anthropic is approaching its first profitable quarter, with revenue jumping 130 percent to $10.9 billion in Q2, according to the Wall Street Journal.
Just last summer, Anthropic told investors not to expect annual profits before 2028. Now the company projects an operating profit of $559 million for the June quarter. That puts Anthropic's growth ahead of Zoom during the pandemic or Google and Facebook before their IPOs, the WSJ reports.
The main driver is Anthropic's coding tools, which companies worldwide have adopted en masse since the start of the year. On top of that, there's "agentic" use of Claude, tasks where the model works on its own over longer periods. Demand at times exceeded compute capacity, forcing Anthropic to throttle access for some users and sign new data center deals, including one with Elon Musk's SpaceX.
AI usage is getting pricier, and Anthropic benefits
On the cost side, agentic AI is a token hog, especially for coding. There's also a hidden price driver in the new flagship model: Anthropic's Opus 4.7 costs the same per token as its predecessor, but significantly more per request. The culprit is a new tokenizer that breaks the same text into up to 47 percent more tokens. For an example session of 80 rounds, developer Abhishek Ray calculated cost increases of roughly 20 to 30 percent. An OpenRouter analysis confirms real-world cost increases of 12 to 27 percent for prompts over 2,000 tokens.
OpenAI is running a similarly aggressive pricing strategy: GPT-5.5's list price doubled compared to GPT-5.4, with input tokens now at $5 and output tokens at $30 per million. OpenAI argues its responses are shorter, but an OpenRouter study shows real costs rose 49 to 92 percent.
Anthropic's efficiency is improving, too. In Q1, the company spent 71 cents per dollar of revenue on compute. This quarter, that figure is expected to drop to 56 cents. CEO Dario Amodei joked at the developer conference that revenue growth had become hard to handle. Unlike OpenAI, Anthropic relies mostly on cheaper chips through investor deals with Google and Amazon and doesn't have to subsidize a large consumer business full of free users.