Last week in AI: Astra’s math results, cheaper tokens, and higher AI spending

Hey there,

OpenAI said its next model solved ten long-standing open math problems. It then cut Luna’s price by 80%. And the we learned a lot about AI revenues from earnings last week.

Let’s get into it:

1. ​OpenAI published ten new results in math and computer science. OpenAI says an internal version of Astra, its next major model, found answers to questions that had seen no major progress for at least ten years. OpenAI estimates that the tokens used to find the answers would cost about $2,000 at current Sol prices.
2. ​OpenAI cut Luna’s price by 80% and Terra’s by 20%. The day before, OpenAI said it had used Sol to lower its own running costs by 20% and improve token output by more than 15%. OpenAI serves a vast amount of free and paid use, so it has a strong reason to improve efficiency. Its large user base may help it find gains sooner, then offer lower prices without giving up as much profit. Also, this kind of sounds like Recursive Self Improvement (RSI)?
3. The market is fine with AI spending when it produces revenue. Microsoft spent $41 billion this quarter, up 70%, and still grew net income by 31%. Meta spent more but earned 14% less. Amazon raised its annual spending plan to $220 billion and says it still cannot meet demand. The takeaway is not that investors have lost faith in AI. They want proof that the spending pays.

A few updates from me:

• ​Cheap smart models are great for builders. Now that Luna is very affordable with solid intelligence, it unlocks a whole new class of software to be built. Tasks that were previously not economical might make sense now. Go ahead, write that for loop that processes your whole 10,000 email inbox.
• ​The economics of AI-powered dev efficiency came up again. Peter joked "GPT-5.6-Sol on Cerebras will be 20x faster, so according to my calculations, I will only be working 30 minutes a day once it comes out". And this really brings back my point about why efficiency gains are taking longer to diffuse — Too large of a diff".

​

As for next week, Palantir reports tomorrow and AMD reports Tuesday. I'm curious to see if the demand for applying AI to the enterprise is insatiable or not. My bet: the music continues.

LFG,

Drew