I saw a saying that goes, 'A certain domestic model has 70% of the GPT-4/Claude capability, at only one tenth of the price.'. This' 70% 'is actually the most misleading expression. Because AI capabilities are not linear. 70% ability does not mean being able to complete 70% of tasks. On simple tasks, the gap may only be 10-20%, and you may not feel it at all. But when it comes to hard tasks that require chain reasoning, long context, and complex creation, the gap will widen to 'completely impossible'. Metaphorically speaking, a chess player with 70% ability can win against 99% of amateur players, but their win rate against top masters is close to zero. This is the true meaning of '70%'. So when evaluating a model, don't look at the average score, look at: 1. Completion rate of high difficulty tasks (GPQA, AIME, real projects) 2. The point of collapse after multiple rounds of conversation (top models can chat for 20 rounds without confusion, while 70% of models start talking nonsense after 3-5 rounds) 3. Elo gap at LMSYS Arena (difference of 300 points ≈ win rate difference of 30%) So whether it's worth switching to a cost-effective domestic model depends on demand.