upvote
The GPT-6.2-Astra 20% price reduction is a match to Fable 5.2's 20% price reduction (although Astra is $0.40 MTok cache read vs. Fable 5.2's $0.20 MTok cache read).
reply
Good point, but also note that OpenAI has the same "discount rate" on subscription for Astra as the other models, while Fable's is much much lower

Opus: about 50x discount (98% off) Sonnet: about 40x Fable: about 5x All GPT: about 20x

reply
I feel like the fact that they're not public lets them play the subsidization game though. And the same playbook that Anthropic started at that.

Would be awesome if it stays this way though

reply
we've seen anthropic numbers, inference is not being subsidized by these companies
reply
> inference is not being subsidized

Inference doesn't have to be sold below direct compute cost to be subsidized. If inference pricing doesn't cover amortized capex and the cost of capital it's still using investor money to buy market share.

reply
Did you read what they wrote carefully?
reply