the performance gap between us and chinese models is at a record-low 6% per bloomberg

@rachelmetz, ai reporter at @technology, broke down the cost math:

the industry is quietly moving from cost per token to cost per task. not how cheap the model is per million tokens – but how much it costs to actually get the job done.

companies and individual users are becoming cost-conscious fast. if you're doing computationally heavy tasks with a frontier model, it gets expensive quickly. chinese models offer what metz calls a really good bang for the buck.

bloomberg tested this directly: they asked models to build a website for a fictional coffee shop. fable 5 was the most expensive. chinese alternatives built passable websites at a 75% discount to claude.

kimi k3 is performing at benchmark levels comparable to the priciest us options on math, science and coding – built on a much humbler budget and with chips generations behind nvidia's lineup. the us-china performance gap narrowed from 9% in may to 6% in june per bloomberg intelligence.

chinese models now account for 41.4% of generative model downloads on hugging face – 5 points ahead of us models.

the old question was: which model is smarter? the new one: which model gets the job done for less?

full segment via link. source: @technology