AI billing costs, spoken instructions in ChatGPT, new graduate careers, Meta’s audio glasses and the facts behind $7,000-a-day token use.
 |
| Friday, September 25, 2026 · Edition 319 |
AI is changing the price of getting work done. Insurers point to rising bills, finance leaders report spending above plan, and new graduates are finding jobs helping smaller companies use the tools. Also inside: spoken assignments in ChatGPT, Meta’s audio glasses and the source behind those $7,000-a-day token figures. |
|
TODAY’S THROUGHLINE
As AI takes on more work, its costs and its effects on people are becoming easier to see.
|
|
|
TOP 5
Five stories worth your time.
|
|
01
Reuters / KELO · Sept. 24 study
A new Blue Cross Blue Shield Association analysis identifies $942 million in extra inpatient spending during 2024–25 compared with 2023, associated with billing for more complex care. It attributes part of the increase to AI tools finding additional conditions in patient records.
|
|
|
02
OpenAI / release notes · Sept. 23
You can ask ChatGPT Work aloud to create a document, presentation or spreadsheet, use a connected app, or work in a browser. The update brings spoken instructions into tasks that produce a usable result, with written responses available in the conversation.
|
|
|
03
Associated Press / BizTimes · Sept. 24 reporting
After hundreds of unsuccessful applications, one graduate landed work by demonstrating an AI search project to a civil engineering firm. AP’s reporting follows students and colleges adapting to a difficult hiring market, where smaller employers need help applying AI to their own operations.
|
|
|
04
Meta / product announcement · Sept. 23
The new Ray-Ban Meta Audio glasses weigh 43 grams and promise up to 12 hours of use on a charge. Meta is also bringing its Muse assistant to AI glasses, extending tasks and planning beyond the phone. Audio frames start at $349; the separate, camera-equipped Gen 3 starts at $449.
 Styles from Meta’s expanded glasses lineup. Image: Meta.
|
|
|
05
Anthropic / research and preprint · Sept. 23
Anthropic’s biology team used Claude agents to search genetic data and flag an unusual pattern for investigation. Human scientists performed the laboratory work and published the early findings. The result shows a concrete division of labor: AI narrows the search, and researchers test what survives review.
|
|
|
AI Workflows | Friday: Rapid new-skill acquisition
Learn pricing by building a decision you can explain.
In an OpenAI Forum replay posted September 24, restaurant owner Ted Taylor demonstrates using AI to compare menu prices with local competitors, portion sizes and his own costs. The discussion lands on a useful learning problem: when AI helps you do work outside your usual expertise, how do you learn enough to judge its answer?
Suppose you run a small shop and need to review the price of one popular item. The new skill is understanding what a price change does after the costs of making and selling that item. You can learn it while producing a decision you actually need this week.
Give an agent such as ChatGPT Work or Claude Cowork read access to approved sales and cost records. Let it research public competitor menus and build an editable worksheet for that one item. It should show the current price, comparable offers, your item costs and applicable selling fees, then calculate what would remain at two possible prices. Every outside price needs a source; every missing cost or uncertain comparison needs a label. If connected access is unavailable, use one approved export as the starting data.
|
The learning assignment
“Build the worksheet and explain the biggest reason your recommendation changes between in-store and delivery sales. Show the calculation using our records. Flag competitor items with different sizes. Leave prices unchanged and give me the questions I must answer before deciding.”
|
Now check the calculation against an actual cost record. Ask what happens if a fee rises or a competitor’s portion is larger than yours. Make the agent correct any bad comparison and save that mistake beside the worksheet. You are learning which inputs move the answer and which apparent bargains are misleading. What remains after these item costs is not the business’s final profit.
The second attempt proves more than the first. Choose a different item with a different cost structure. Before the agent analyzes it, predict whether the same price change would help or hurt. Have it run the numbers, then explain any disagreement in your own words. If you cannot account for the result using the source records, you have found the next part to learn.
A manager can give an employee one product category and review both worksheets together. The owner approves any real price change after considering customer response and costs the worksheet may have missed.
Pick two items before the week ends. Make one worksheet with AI, then test your understanding on the second before you look at its recommendation.
|
|
|
|
NEW AI TOOL
Adobe’s September 24 update adds Acrobat tools to its Claude plugin. You can combine or split PDFs, reorder pages, convert files and open an interactive editor for text changes, comments and highlights. A document-review feature can add recommendations directly to the file.
Try it with an approved meeting packet: combine the documents, remove superseded pages and inspect the finished PDF. The plugin is rolling out globally and supports guests and free or paid accounts. Adobe says it is not optimized for enterprise use and does not support education accounts.
Adobe announcement · Sept. 24 · Setup and availability
|
|
|
THE DAILY ROUNDUP
Headlines
|
Work, skills and education
Associated Press / BizTimes · Sept. 24 reporting
Colleges are adding business-specific internships and AI training as traditional software hiring weakens. Employers increasingly want evidence that graduates can solve a working business’s problems.
Indeed Hiring Lab / economist survey · Sept. 24
More than 120 U.S. academics and economists surveyed by Indeed grew more concerned about AI’s effect on degree holders’ wages over the next year. These are forecasts; the survey does not measure actual AI-related pay cuts.
Pearson / research release · Sept. 24
Pearson identifies gaps in job-specific AI knowledge, human judgment and practical experience across trades, technical roles and services. Retiring staff make the problem harder: employers must preserve hands-on knowledge while teaching new tools.
Center for Democracy & Technology / toolkit · Sept. 24
The guide covers school leaders’ own use and their responsibility for what happens across a building. It emphasizes access to training, student privacy and preventing greater risks for already disadvantaged pupils.
CNA / launch reporting · Sept. 24
Twenty-three financial institutions plan to train their Singapore workforces by 2028. The newly launched collaboration will also test changes to jobs, giving the initiative a practical task beyond course completion.
|
|
Tools people can use
OpenAI / release notes · Sept. 23
Connected apps follow the account’s existing permissions and limits. Voice in Work requires access to both features; an unfinished assignment can continue in text after the call ends.
Meta / product announcement · Sept. 23
Ray-Ban Meta Audio is available to preorder and is scheduled to ship October 13. The new frames support calls, music and spoken assistance, with prescription options.
Google / expansion announcement · Sept. 23
The three-dimensional video meeting system is shipping in six countries. Starting in October, select Industrious locations in four U.S. cities will offer sessions, giving smaller teams a way to try it before buying specialized hardware.
Framer / product update · Sept. 22
Skills preserve a project’s design rules, writing style and content workflow for repeated use. The instructions can travel with a remixed template, helping a team maintain the site after its original designer hands it over.
|
|
Trust and public decisions
ABC7 / university response · Sept. 24 reporting
The university acknowledged that altering the dining-hall photo violated its policy. The incident gives communications teams a concrete reason to check whether an apparently authentic image still depicts the people who agreed to be photographed.
YouTube / product announcement · Sept. 23
The platform will combine speaking-voice detection with facial matching and bring likeness alerts and reporting to mobile. These are announced additions; creators should check availability before relying on them to monitor copies.
Australian prime minister / transcript · Sept. 24 disclosure
The June incident involved unauthorized access to public and nonpublic files in a Medicare statistics portal. The prime minister criticized delayed notification; no personal information was believed accessed at the time of the announcement, and the investigation continues.
The National News Desk / summit reporting · Sept. 24
Xi called for AI to remain under human control as the leaders met Thursday. A proposed system for notifying each other about serious AI incidents remains under discussion; the report does not establish that a binding safety agreement was reached.
Illinois / executive order · Sept. 22
The new group brings state officials and outside expertise into a review of AI’s effects on safety, privacy and the economy. Its remit is to assess risks and recommend further state action.
Government Technology / Digital States Survey · Sept. 24
In the 2026 survey, 76% of states reported needing additional staff for AI modeling, while 84% needed more cybersecurity workers. Public services adopting AI still depend on people who can maintain and protect the systems.
|
|
Health and scientific research
Reuters / KELO · Sept. 24 study
The insurer association argues that more billable diagnoses did not come with a comparable increase in treatment. Its findings do not establish that every added charge was improper.
Anthropic / research and preprint · Sept. 23
The work concerns an enzyme system in viruses that infect bacteria. Its function remains under investigation, and a useful medical application has not been demonstrated.
Chief Healthcare Executive / Ardent interview · Sept. 24 reporting
After about a year of use, the health system reports less documentation time and work after hours. The estimate comes from Ardent’s own monitoring; clinicians still review the notes before signing them.
eClinicalWorks / practice case study · Sept. 24
Goodtime Family Care reports that healow Genie handled 77% of daytime calls and scheduled more than 700 appointments in its first four months. The practice’s claimed staff-time savings come from a vendor-published account.
Google / Project Suncatcher update · Sept. 24
A prototype satellite will test how the chips handle launch, radiation and cooling in space. This is an early hardware experiment in a project announced last year; working orbital data centers remain a longer-term goal.
|
|
Budgets and business operations
OpenAI / research report · Sept. 6 · background
By mid-August, the median researcher’s daily use exceeded $600 at public API prices, while the 90th-percentile user exceeded $7,000. These figures value internal usage at published rates; they are not researchers’ personal bills or a statement of OpenAI’s actual compute costs.
Esker / survey release · Sept. 24
Among 338 global finance leaders surveyed, 72% said their organizations spent more than planned on AI during the past year. Many still reserve execution or approval for people, making cost control and decision authority part of the same purchasing conversation.
UiPath / platform announcement · Sept. 23
Delegate can gather context, move information between systems and save repeated tasks as routines that run on demand or on a schedule. The update also adds controls for models, permissions and actions, while people retain final testing decisions.
Shield / product announcement · Sept. 24
Forge handles eligible helpdesk work and supplies technicians with context for other cases. Shield reports faster resolution in its own network; access comes through a Shield partnership, with deployment tailored to the provider.
Rocket Software / release via BigDATAwire · Sept. 24
EVA helps teams investigate failures and performance problems on mainframes. A new permission checkpoint separates an agent’s proposed action from execution; customers are testing the platform in pilots with their own operational data.
|
|
|
mAIn Street gives nontechnical readers useful AI news they can put to work.
mAIn Street #319 · Friday, September 25, 2026
Reporting checked through September 24, 2026.
|
|
|
|
