尽管国产大模型在开源市场表现突出，但Anthropic的商业收益并未受到显著压缩。根据数据统计，Anthropic的年化经常性收入（ARR）在5月达到470亿美元，到7月中旬增长至741亿美元，环比增速达58%[1]。OpenAI首席财务官在7月29日的内部会议上表示，该公司7月的ARR"超过了整个二季度之和"[1]，表明OpenAI才是Anthropic面临的主要竞争压力源。
从Token消耗结构看，国产模型的市场份额确实快速扩张。在OpenRouter平台上，美国用户在7月13日至19日期间消耗的Token中有64%来自中国开源大模型，而Claude、GPT和Gemini等头部商业模型的联合份额从一年前的70%下降至现在的30%左右[1]。但这种市场格局反映的是不同层次的市场细分：国产大模型以低廉价格在开源和性价比市场获得认可，而Claude仍维持较高定价策略[1]。
从性能成本权衡来看，差异更为明显。根据Kilo Code 2026年6月的测试数据，开源大模型的平均单次成本为0.22美元、成功率为46.7%，而Claude Opus等商业模型的单次成本为0.79美元、成功率达65.6%[1]。这表明Anthropic在复杂任务处理上保持的技术优势，使其能够在高端应用市场维持溢价能力，国产大模型对其商业收入的直接冲击相对有限。
Anthropic has demonstrated robust financial momentum despite intensifying competition from Chinese open-source large language models, according to market analysis. The company's annualized recurring revenue (ARR) reached $47 billion in May and grew to an estimated $74.1 billion by mid-July, representing a 58 percent sequential increase [1]. This expansion trajectory suggests that OpenAI, rather than domestic competitors, remains Anthropic's primary competitive pressure in the market [1].
While Chinese open-source models have captured significant market share by token consumption on the OpenRouter platform—accounting for 60 to 70 percent of tokens consumed by U.S. users between July 13 and 19—they have not substantially eroded Anthropic's commercial revenue [1]. The combined market share of Claude, GPT, and Gemini declined from approximately 70 percent one year ago to roughly 30 percent currently on the same platform [1]. However, this token-based metric does not directly correlate with revenue performance. Domestic models have gained traction in price-sensitive segments and open-source markets through significantly lower costs, yet Claude maintains premium pricing that reflects its superior performance on complex tasks [1].
Performance benchmarks illustrate the value differentiation between the two approaches. Testing conducted by Kilo Code in June 2026 showed that open-source large models averaged $0.22 per task with a 46.7 percent success rate, while Claude Opus and comparable enterprise-grade models averaged $0.79 per task with a 65.6 percent success rate [1]. Meanwhile, OpenAI's financial trajectory also accelerated sharply, with the company's CFO announcing in an internal meeting on July 29 that July's ARR "exceeded the entire second quarter combined" [1].