More than two months after its launch, Fable5, Anthropic's most expensive and most advanced artificial intelligence model, is struggling to win users. Amid a sense that existing Anthropic models are good enough, the company is also losing customers to cost-effective Chinese AI models.
According to the Financial Times on the 23rd, an analysis by payments firm Ramp of spending data from 70,000 companies found that businesses have devoted just 11% of their total Anthropic spending to Fable5 since the model launched in June. Instead, Anthropic's Opus5 — a smaller but powerful model released in late July — has overtaken Fable5 at the mid-10% range. The lower-spec Opus4.8 surged from around 15% in early June to more than 50% by late July.
Analysts and investors said the shift reflects the fact that Anthropic's existing models can already handle most corporate requirements without Fable5. Version 5.6 of ChatGPT, released in July by rival OpenAI, also dealt a blow with pricing far cheaper than Fable5. The FT noted that this marks a break from the established pattern in which corporate users choose the most powerful models as their default.
Even U.S. customers are increasingly adopting Chinese AI that offers strong performance for the cost. According to China's Global Times, Harvey, a U.S. legal AI company backed by OpenAI, built its own AI model on Kimi K3, an open-weight model from China's Moonshot AI. Based on Kimi K3's average token usage, users pay only about 30% of what Fable5 costs. The Global Times said the case shows U.S. companies taking interest in inexpensive, high-performing open Chinese models.
According to a report released on the 14th by Hugging Face, the world's largest open-source AI community, some U.S. models unveiled this year with more than 100 billion parameters were built on Chinese AI models or drew on the work of Chinese research labs. Airbnb, for example, said that after introducing a customer service agent using Alibaba's Qwen model in its customer service chatbot in May, its average time to resolve an issue was cut from about three hours to six seconds.
