[Oras]

Interesting the Fairseq GPT 13B pricing is cheaper than GPT-3 curie model (which is 6.7B).

For comparison:

Fairseq GPT 13B (1k tokens) = $0.00189

GPT-3 Curie 6.7B (1k tokens) = $0.0060

[IanCal]

This leads to the more interesting side that he's built a custom inference setup to run more efficiently.

> TextSynth employs a custom inference framework to get faster inference (hence lower costs) on standard GPUs and CPUs.