ChatGPT made itself cheaper

Welcome back! OpenAI pointed its own model at its own plumbing and came out with a price cut, and Google taught robots to use their legs. Meanwhile Amazon booked $53 billion from owning a piece of Anthropic while Meta spent almost every dollar it earned and watched its stock fall.

In today's Generative AI Newsletter:

OpenAI cut the price of GPT-5.6 Luna by 80% yesterday, from $1 per million input tokens down to 20 cents. Terra came down 20%.

Your subscription price didn't move. What changed is that Luna and Terra now eat fewer credits inside Codex and ChatGPT Work, so the same plan stretches further.

Here's who found the savings. Inside a human-led process, GPT-5.6 Sol rewrote and optimized OpenAI's production kernels, then designed and ran hundreds of experiments on token generation.

That work cut the cost of serving the model by 20% and improved token-generation efficiency by more than 15%. OpenAI turned it into a price cut the next day.

Luna is the one to watch. On professional work it beats Fable 5 at an estimated cost per task nearly 99% lower, and runs year-ago frontier performance at roughly 6 cents on the dollar.

Replit's president called it the closest anyone has come to intelligence too cheap to meter.

Sol got Fast mode instead, 2.5 times quicker at double the price.

HubSpot switched to Mintlify and saw 3x faster builds with 50% fewer eng resources. Beautiful, AI-native documentation that scales with your product — no custom infrastructure required.

Simplify Your Docs Today

Google's earlier robotics models drove a humanoid's upper body for tabletop work. Gemini Robotics 2 drives all of it, feet to fingertips.

Ask Apptronik's Apollo 2 to put the watering can in the green bin on the bottom shelf and it walks to the table, picks the can up, walks to the shelves and puts it away.

Two robots can now split one job. A reasoning model called ER 2 works as the high-level brain, watching the room, planning sequences that run for minutes and correcting itself when a step fails.

There's an on-device version too, and it adapts to a completely new robot body in a few hours on fewer than 200 examples.

The success rates say where this actually is. On Apollo 2, picking off a shelf works 76.3% of the time and off the floor 45.7%. The five-fingered hand unscrews a lightbulb 92% of the time and manages a dustpan 32%.

ER 2 is on Google AI Studio now. The models that move the body are limited to early-access partners.

Amazon reported $62.6 billion in quarterly profit, and $53.4 billion of it was non-operating income coming primarily from its investments in Anthropic.

The same quarter last year, the entire company earned $18.2 billion.

The core business is doing fine on its own. Net sales rose 20% to $200.6 billion, AWS grew 37% in its fastest quarter in four and a half years, and the AWS AI business and Amazon's chips business each passed a $25 billion run rate.

Amazon is selling to both sides too. Anthropic and OpenAI have each committed multi-year and multi-gigawatt to Trainium, Amazon's own AI chip.

What the building costs shows up further down. Free cash flow swung to an outflow of $7.6 billion over the trailing year after property and equipment purchases climbed $66.1 billion.

Microsoft ran the same trade this week, booking a $3.2 billion gain on its Anthropic stake, nearly as much in one quarter as its OpenAI investment returned all year.

Meta's revenue rose 28% and beat expectations. The stock fell as much as 10% in after-hours trading anyway, then settled around 7% down.

Costs ballooned 55%. Operating income fell 8%, net income fell 14%, and free cash flow came in at $784 million against the roughly $12 billion Meta has averaged over the past eight quarters.

Capital expenditure hit $31.1 billion for the quarter, nearly double a year earlier, against $31.9 billion of operating cash flow. Almost every dollar the business earned went into servers, data centers and chips.

Meta also issued $24.9 billion of long-term debt and bought back no stock, after repurchasing more than $10 billion the year before.

Full-year spending is now guided at $130 to $145 billion, and $50.9 billion is already gone. Run that math and this was the last positive cash-flow quarter of the year.

Microsoft, Amazon and Google rent their infrastructure out through cloud businesses. Zuckerberg confirmed Meta has cloud plans coming, then called it foolish to sell compute for a short-term profit when selling intelligence carries the higher margin.

📸 Follow us on Instagram for fast, visual AI updates in 30 seconds. 

📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.

🐦 Follow us on X for breaking AI news and real-time industry updates.

🧠 Learn how to build your next AI application with practical resources and expert guidance.