Anthropic Fable 5 Training Cost is Three Times Higher than Open Source Models, Facing Future Challenges

In a recent analysis, Wojciech Gryc explored the market competition between Anthropic's Fable 5 and two new Chinese models. Kimi K3 launched by Moonshot on July 16th and Qwen 3.8 released by Alibaba on July 19th have achieved performance similar to Fable 5, and both models will open-source their weights.

Gryc pointed out that the main costs of training cutting-edge AI models come from researchers and computing power, but after the model is trained, the real expense is the inference cost. Every time a user calls an API, it consumes a large amount of GPU resources and electricity. He believes that how to pay for inference costs will directly affect the company's business model.

He categorized AI companies into three types: the first type are companies that rent data centers and pay electricity bills, such as Anthropic, Moonshot, and GLM 5.2; the second type are companies that build their own data centers, such as Meta and Alibaba; the third type are companies that have self-power generation capabilities and built their own data centers, such as SpaceX. Gryc emphasized that companies with infrastructure can still make profits by renting servers even if their model performance is not the best.

Additionally, Gryc noted that the inference cost of Anthropic's Fable 5 is almost three times that of OpenAI and open-source models for a single task. Whether users can accept such high costs in a competitive market is a matter worth attention. He believes that current benchmark tests in the market are becoming saturated, and users' perception differences of models are far less than cost differences, which may lead to the risk of Anthropic being abandoned by the market.

Compared to this, OpenAI demonstrates stronger market competitiveness with its comprehensive product line and data center layout. Gryc expects that this wave of competition will continue to develop, and multiple laboratories may continuously surpass traditional American tech giants in the future.

Key Points:

🌟 The inference cost of Fable 5 is three times that of OpenAI and open-source models, facing market pressure.  

🔍 Gryc categorizes AI companies into three types, emphasizing the importance of infrastructure.  

📈 Competition is intensifying, and several Chinese laboratories are rapidly catching up or even surpassing American companies.