The Falling Price of Thought
0:00 / 1:00
The Falling Price of Thought
This week's cue is a bright, restless instrumental about a price falling fast: one clipped piano figure steps out alone, doubles, then doubles again as glassy arpeggios answer from further out over a dry breakbeat, until the layers cut away and a single held chord refuses to resolve. It suits a commute, a coffee break, or the minute before a focused work block.
Listener notes
- OpenAI halved its token prices. On September 22 the company added GPT-6 Sol and GPT-6 Luna to its GPT-6 line at half the promotional rates of their predecessors: Sol at $2 per million input tokens and $10 per million output tokens, Luna at $0.10 and $0.50, with both borrowing some of the flagship Astra model's capabilities. OpenAI credits caching and inference improvements for the reduction, so these are the company's own published prices rather than an independently measured cost. 1
- Alibaba marked its whole voice stack down by up to 95%. Its Qwen team released Qwen-Audio-3.1, five models covering speech recognition, synthesis and real-time conversation, and cut prices across the line at the same time: text-to-speech down about 70%, the real-time model about 85%, and speech recognition by as much as 95% against Alibaba's previous list prices. 2
- Google open-sourced the machinery for running agents at scale. Released on September 21 under an Apache 2.0 licence, AX treats an AI agent as a workload of its own kind: a sandboxed, stateful process that is suspended while it waits on a model or a person and resumed in under a second. Google's repository describes it as built to run "billions of autonomous agent workloads in a cluster", which is the project's own claim, and early commenters split between those who call idle-agent compute the real cost it solves and those who object to the Kubernetes-sized operational overhead. 34
- Anthropic says roughly 950 Claude agents found a biological system in 21 hours. Working through more than 200,000 reverse transcriptases and 210 million tokens, the agents produced reports on 20 candidates and flagged a repeat array beside one of them; the lab then confirmed an uncharacterized system in bacteriophage DNA that Anthropic calls ART. What ART actually does is still unknown. Lucas Harrington, a CRISPR researcher, argues that this kind of genome-neighbourhood mining has been standard for decades and that the hard part — working out what a system does — has not been done, while Feng Zhang of the Broad Institute calls the pattern genuinely intriguing. The finding rests on a preprint that has not been peer reviewed. 56
- The argument over control got louder, not clearer. At the White House on September 24, Xi Jinping said the two countries have a "responsibility" to keep AI "under human control", while Donald Trump said the United States "totally rejects any attempt to construct a globalist scheme" and that the Justice Department is the only guardrail he wants. A day earlier, Bernie Sanders and Greg Casar introduced the Ban Artificial Superintelligence Act, which would pause advanced AI development until a cabinet-level Department of Artificial Intelligence exists and would bar recursive self-improvement. Washington analysts told reporters it is unlikely to pass this year. 78
The shared thread is that the machinery of AI got cheaper and more numerous this week while the argument over who controls it only got louder: a thousand agents are already affordable, and the rules for them are still being drafted in opposite directions.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.