Latest / Minds, Bodies, and Terawatts / Google Just Cut the Price of Its Coding AI in Half, and the Whole Industry Is Racing Down With It
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, the show where the AI race stopped being about genius and started being about the invoice. Reuters reported Thursday, August thirteenth, that Google launched Gemini three point seven Flash and cut the price of its coding model in half.
- 0:16[Guest] Seventy-five cents per million input tokens, three seventy-five per million out. That rate holds only through December.
- 0:23[Host] Tokens are the chunks of text a model reads and writes. A million written out, a shelf of paperbacks' worth of code, costs less than a sandwich.
- 0:33[Guest] Koray Kavukcuoglu, who runs Gemini model development at Google DeepMind, called it a workhorse model on X. A workhorse, the animal you rent by the hour.
- 0:42[Host] It shipped three weeks after the model it replaces. That's not a product cycle, that's a price cut with a new name. Aimed at whom?
- 0:52[Guest] Google published a table with its rivals' list prices in it. Claude Sonnet five at two dollars in, ten out. Google put itself at seventy-five cents.
- 1:02[Host] That's not a capability pitch. That's undercutting.
- 1:04[Guest] OpenAI answered the same day with a different lever, an invite-only tier of GPT five point six Sol running up to fourteen times faster on Cerebras chips. Both are chasing agents, software that runs for hours with nobody watching.
- 1:19[Host] And an agent that waits is an agent you pay to wait. Does this reach me, or is it two companies trading discounts?
- 1:27[Guest] Read what Google says the model is for. Software engineering, web development, document-heavy knowledge work. That isn't a category, that's a job description.
- 1:36[Host] A one-person shop just got frontier-grade coding help for pocket change.
- 1:40[Guest] That part is real. That same rate is also on the desk of whoever decides whether to hire that shop at all.
- 1:48[Host] Then why is my company's AI bill going up instead of down?
- 1:53[Guest] Gartner's Will Sommer, a senior director analyst, said in March that an agentic model burns five to thirty times more tokens per task than a chatbot. Cheap doesn't shrink the bill, it widens the work you hand over.
- 2:07[Host] Except we've watched a company make that bet and walk it back. Klarna replaced most of its customer service with AI, then started hiring people again.
- 2:17[Guest] Its CEO, Sebastian Siemiatkowski, admitted the cost-cutting went too far and quality suffered. The machine kept the routine queue, the humans came back for the hard cases.
- 2:27[Host] Is there research on where that line falls?
- 2:30[Guest] Generative AI at Work, the Brynjolfsson, Li, and Raymond paper for the National Bureau of Economic Research, gave an AI assistant to five thousand support agents. Issues resolved per hour rose fourteen percent overall, more than thirty percent among the newest agents, and close to zero for the most experienced.
- 2:48[Host] So it lifts the floor and leaves the ceiling alone.
- 2:52[Guest] Which is what a seventy-five-cent model does to a payroll. It buys the routine tier outright and stops where reliability runs out. Klarna rehired the people that paper says the machine could not copy.
- 3:05[Host] Then what decides whether a task flips over?
- 3:08[Guest] Not the demo. The invoice. The best model wins the bake-off, the cheapest good-enough one wins the volume.
- 3:15[Host] Patrick Deglon calls that line the Substitution Threshold in Unscarcity, the moment the cheapest reliable provider of your work stops being human. It moved this week on a rate that expires in December.
- 3:27[Guest] For more on what happens to work when the machines get cheap, head to unscarcity dot ai. The book is on Amazon, the audiobook on Spotify. This has been Minds, Bodies, and Terawatts.