| |
Since last week: Google released Gemini 4 Argon, its most capable model. Anthropic's AI now writes code so fast that Anthropic replaced Git, the industry-standard tool software teams have relied on for two decades. Seven engineers built the replacement with Claude's help and moved Anthropic's main code base onto it in just five months. Also: AI stocks fell, eight states sued OpenAI, Meta released new models, and big banks moved to keep outside AI agents out of customer accounts. |
The week in three numbers
| $16.8B | Anthropic's revenue for July to September, about 45% above the quarter before, the FT reported. Four quarters at that level would be about $67 billion. OpenAI's yearly pace was about $50 billion in early October, roughly double its $25 billion at the end of February. |
| 2x | How much higher Google's Gemini 4 Argon scored on a well-known reasoning test in a setup like Google's own (97.5%) than in a generic setup every model gets (49.8%). OpenAI's GPT-6 Astra showed a similar gap in September: 99.9% vs 54.8%.Why it matters: two leading models now show the same pattern. Google's shared-setup run completed only 42 of 55 tasks, so the gap isn't a clean measure of the setup's effect. But the point for buyers holds: a single published score won't tell you how a model will perform in your own setup. ( ARC Prize: Argon · ARC Prize: Astra) |
| $3.5T | The AI data-center spending Gartner forecasts could be exposed to write-down risk by 2030 if building continues at today's pace.Why it matters: Gartner says owners would need $8.5 trillion of revenue through 2030 to cover their costs and a 12% return, and it expects no more than $3.6 trillion. ( Gartner, Oct 14) |
The week in product
Meta released its most capable models (Oct 12) and a very large one anyone can download (Oct 15). Muse Spark Open runs on the major clouds or on two high-end Apple desktops, under a license with "limited commercial terms." Meta's own comparisons admit its runs of rival models "may not be optimally tuned." ( Meta, Oct 12 · Meta model card · Meta, Oct 15)
ChatGPT can now read your Chrome tabs and act on sites where you're signed in (Oct 15). It asks before using each new site, unless someone turns on a setting OpenAI labels "Elevated Risk." ( OpenAI · The Verge)
Workday made its AI agent, Sana, the new front door to Workday for companies that switch it on (Oct 13). Staff ask it to approve an expense or process leave instead of clicking through screens. ( Workday · Workday)
Anthropic's Claude Code Review is now available on Team and Enterprise plans (Oct 15). Once an admin picks the code projects, it reviews each new change request automatically. ( Anthropic)
|
In this issue
01 · The One Thing · 02 · Do This Week · 03 · The Signal · Skim ends here.
04 · The Margin-Proof Tracker · 05 · The Synthesis · 06 · Where the Minds Disagree
Then: What We're Watching · Worth Your Time · Corrections
|
|
01 · The One Thing
Google released Gemini 4 Argon, its most capable AI model, on October 13, to developers, paying subscribers and businesses at once. It is available in Google's developer tools and Google Cloud, in the Gemini app and Search for paying subscribers, and in Gemini for business.
The introductory price is $2 per million tokens in and $10 out (a token is a piece of a word) until November 30. Then it doubles. Google says Argon ranks first on the Vals Index, a test of finance, legal, tax and coding work. (Google, Oct 13 · Google pricing)
|
| |
The executive shift: Google is back among the top model makers, and the leaders are close. On Epoch AI's index, which combines many tests, Argon scored 165, against 166 for OpenAI's GPT-6 Astra and 167 for Anthropic's Opus 5.5, The Decoder reported. Yet on one reasoning test, Argon scored 97.5% in one setup and 49.8% in another. For a buyer, the question is not how close the published scores are. It is how the model performs where your work gets done. Test it in your own setup, at the December price. (The Decoder · ARC Prize) |
1 · Test the model and its setup as one system, and change one at a time. Owner: whoever runs your AI platform. Our read: when you switch models, keep your tools and instructions fixed and rerun real tasks. When you change the tools, keep the model fixed. Stakes: Argon scored about twice as high in Google's setup as in a generic one, and almost half of the bigger paying teams on one developer gateway changed their main model in a single month (Section 06).
2 · Pick one process where AI output is piling up, and redesign the step it waits on. Owner: the executive who owns that process. Our read: ask what the waiting step is for, and whether it can change shape rather than just get more staff. Anthropic's reviewers now check what each change is meant to do instead of reading every line. Stakes: last week's study of 718 companies found more code, slower review and no measurable rise in finished work (The Synthesis).
3 · List the AI features that switched on inside your work tools this month. Owner: your chief information officer. Our read: check each against your data and records rules. Stakes: Gmail's new AI Inbox arrived with "no admin control" (The Signal).
This week's question: as AI spreads into more of the work, who is checking it, and who is paying for it?
Trust Scrutiny of AI firms grew in three forms: lawsuits, investigations and a White House charter. None adds a new rule yet. Primary: state attorney general filings. Reported: NBC, Axios, The Next Web.
Alleged in court. Eight states sued OpenAI on October 14 under their consumer-protection laws, alleging it misled people about ChatGPT's safety, including for children. Arkansas asks a court to stop "further advanced AI testing" until safeguards exist. OpenAI says it has built "industry-leading age-specific protections."
Under investigation. The FTC sent OpenAI and Anthropic demands for documents, Axios reported. The EU's AI Office sent requests about security and outside testing.
New rules: none yet. The White House AI task force's charter says labs are "ultimately responsible" for their models' risks, but defines no incident, reporting deadline or penalty.
Our read: the states aren't waiting for AI rules. They are using consumer-protection laws that already apply. (NBC DFW · Iowa AG · Arkansas AG · Axios · Euronews · European Commission · The Next Web)
Horizon Lenders and investors started putting a price on AI risk. Reported: Reuters, Axios, CNN. After Anthropic's big quarter, some investors discussed valuing it at about $1.5 trillion in its planned stock sale, below the roughly $2 trillion expected, Reuters reported. Nothing is agreed. CoreWeave, which rents out AI computing power, fell more than 15% on October 14, CNN reported, and its bond yield hit a record high, Axios reported. South Korea's main stock index fell 7.6% on October 15, led by memory-chip makers. Our read: investors now want evidence that AI revenue will cover the bill. ( Reuters, via WMBD · Axios · CNN))
Work AI is switching on inside the tools people already use, sometimes without an off switch for IT. Primary: Google. Reported: Android Authority. Gmail's new AI Inbox, which pulls to-dos and summaries out of email, began reaching paid Workspace accounts on October 12. Google's page says: "There is no admin control for this feature." Gmail Live, which answers spoken questions about your inbox, is on by default but follows existing IT settings. Our read: the vendor now sets the default. A company's AI policy has to cover what arrives switched on. ( Google: AI Inbox · Android Authority: Gmail Live)
Deploy Big banks blocked outside AI agents from logging in to customer accounts. Reported: Fortune, American Banker, FT. JPMorgan's Jamie Dimon said the bank is "not going to allow just any outside agent" inside. "You don't want rogue agents." Schwab said it blocks agents that use clients' passwords, and points clients to its own assistant, Charley. The FT reported Citi is restricting agents too. Both JPMorgan and Schwab still allow some approved routes; what they shut is the side door. Our read: agents that act wherever you're signed in, like ChatGPT's new feature, now meet businesses that decide which agents get in. Any company with customer logins will need that rule. ( Fortune · American Banker · FT)
Proof A grocer plans corporate job cuts, and a software vendor went after the money companies spend on people. Reported: AP, The Register. Kroger plans to cut 3,000 corporate jobs over three years. Its memo cites technology, not AI; AP tied the move to Kroger's AI plans. Workday told investors its agents could reach "beyond the software budgets into the labor and services spend," billed by use. Our read: Workday wants a share of what companies spend on staff and outside services, not just software. The question for buyers is who keeps the savings. ( AP, via ABC News · The Register)
Below the Cut
Cheap open models won most of the volume. Most of the money still went elsewhere. Primary: Vercel. On Vercel's AI Gateway, models anyone can download handled 71.5% of the work in September, measured in tokens, but took just under a quarter of estimated spending. Anthropic's Claude Opus 5.5 alone drew more than one dollar in ten. Our read: cheap open models now carry most of the volume, while closed models still take most of the money. ( Vercel · The Register)
| End of skim · deep read begins |
| 04The Margin-Proof Tracker |
| |
One row added, none moved. N = 20. Bank of America's test came due and did not clear the bar. On its October 14 call, its chief executive gave AI benefit figures for this year and next, but no method and no cost. Its filing that day counts AI users and uses but attaches no dollar figure to AI. The row holds at 2. UnitedHealth is new at 1: it counts clinician time saved but puts no money on AI. Duolingo is still the only Stage 3 row. |
This tracks public AI value claims and the evidence behind them. A low rung is not a verdict on whether a claim is true. The evidence ladder: 0 · Narrative (a story, no numbers) · 1 · Operational (activity counted, no money attached) · 2 · Financially linked (a money figure tied to AI, mixed with other causes or too narrow to be the whole cost) · 3 · P&L-attributed (a margin change credited to AI in a filed document) · 4 · Sustained (Stage 3 held four quarters).
| Company |
Evidence |
Grade |
Next test |
| Bank of America (test came due, not cleared) |
EvidenceOn the October 14 call, the chief executive said about 130 AI projects are in place. He put their benefit at about $250 million to $300 million this year and "$700 million next year," and said benefits should run two to three times the expense over time. He gave no method or cost figure, and did not say whether this year's amount is already booked. Nothing links these figures to the $800 million benefit on $400 million cost the bank gave on September 14. The filing counts "more than 200,000 teammates using AI tools" and over 300 approved uses, with no dollar figure for AI. Call, Oct 14 · Filing, Oct 14 |
Grade2, unchanged |
Next testNext quarterly filing: does it give the AI benefit a cost and a method, say whether it is realized, and tie it to expense? |
| UnitedHealth (new) |
EvidenceOn the October 13 call, the chief executive said AI chart summaries, now used with more than 40% of its employed clinicians, are cutting chart review time by 33%, and appeal cycles are 26% shorter. No one on the call put a dollar figure on AI. The filed release credits a lower operating cost ratio, 12.8% from 13.5%, to "realized operating efficiencies," and lists "operational agentification" among the investments, not the savings. Disclosure: the author works at UnitedHealth Group; this grade uses only public filings and calls. Call, Oct 13 · Release, Oct 13 |
Grade1, new |
Next testNext filing: does it put a dollar figure, or a cost-ratio effect, on AI? |
Next tests. Microsoft reports Wednesday, October 28 (Microsoft). Bank of America's next test is a filing that prices and explains its AI figure. Duolingo's Q3 is the next step toward Stage 4.
When the work outgrows the system that manages it
Last week we looked at a study of 718 companies. After they adopted AI coding agents, code sent for review rose 23% per worker, each review took 49% longer, and finished work did not measurably rise (Issue 017). The work had piled up at the step after the code was written.
This week Anthropic described a related strain at a different step. Its problem was less the people reviewing code than the system that tracks and merges it. In May it merged about 2,500 code changes a day. By September it was merging more than 30,000, about twelve times as many. Git, the standard tool developers use to track and merge changes, could no longer keep up. Anthropic says Git's design, built for people working offline on full copies of the code, was "actively undermining" its ability to grow once most of the work came from always-online AI agents. So it replaced Git for its main code base. Primary source. (Anthropic, Oct 13)
What changed was the process, not just the volume.
A small team. Seven engineers built the new system, called Scion, and a new code-review tool, and moved the main code base over in five months.
AI built the tools that manage AI's output. Claude helped write both. When Claude rewrote an early prototype of one part of the system, the new version used 20 times less memory and ran 4 times faster. Every engineer on the project would now rate Claude "better at writing code than them," Anthropic writes.
People did the judging. Anthropic writes that the engineers' role "was to tell Claude 'no.'" They steered it away from over-engineering. Its adversarial review is effective, it says, but "can also be myopic." Engineers now review the intent of each change rather than reading every line.
The result, by Anthropic's count. Scion uses 200 times less processing power and memory than the fleet of Git copies it replaced, and automated testing gained 25% in processing capacity.
What we do not know. Anthropic gives no defect or rollback rate, so there is no way to tell whether quality held at twelve times the volume. Every number is Anthropic's, about its own work. And Anthropic says Scion "is by no means a suitable replacement for git for general use." It makes sense only at very large scale. GitHub, Microsoft's code-hosting service, said in April that "agentic development workflows have accelerated sharply," so the pressure is not unique to one lab. (GitHub, Apr 28)
A different choice, outside software. Britain's Home Office uses an AI tool built on OpenAI's GPT-4o to write summaries of asylum interviews. After it was rolled out in April, 45 of the 203 summaries that got an initial expert check failed it, or 22%. An earlier, separate pilot evaluation rejected 9%, so the two rates are not a clean before-and-after. The figures come from a report by the Open Rights Group, a digital-rights campaign, based on freedom-of-information requests. The group estimates experts checked only 0.3% of summaries from April to July. Applicants are not specifically told when AI was used and cannot correct the summaries. The government says people still make the decisions; the Home Office did not answer The Register's questions. Reported. (The Register, Oct 13) Unlike Anthropic, which redesigned its systems for faster production, the Home Office rolled out AI summaries with very little expert checking. The two cases are not the same kind of work. They show two different choices about how to redesign work once AI changes how it gets produced.
Analysis: Anthropic did not make yesterday's process faster. It asked which parts of that process were built around limits that no longer hold. Git assumes each developer works offline on a full copy of the code; Anthropic's agents are always online. A merge queue that locks the whole code base for each change "doesn't make much sense at agent scale." More of the testing now happens after code is merged. Reviewers now check what a change is meant to do instead of reading every line. Each of those habits made sense when people wrote code at human speed. Most companies are nowhere near Anthropic's scale. But Britain's asylum tool shows the risk of changing how work gets produced without equal care for how it gets checked. So the question for any process now carrying AI's output is: which of its steps exist because of a limit that is gone? The engineers did not become less important. Their job moved toward design, judgment and knowing when to say no.
| 06Where the Minds Disagree |
| |
Which score should a buyer trust: the maker's setup, or a shared one?
ARC-AGI-3 is a reasoning test of small puzzle games, run by the nonprofit ARC Prize. In its shared setup, every model gets the same plain interface and keeps only the notes it chooses to write down. In the provider setup, the model keeps its own private working state between turns, as it does in its maker's products. That wrapper of memory, tools and instructions is called a harness. At the same low setting, Argon scored 97.5% in the provider setup and 49.8% in the shared one, which finished only 42 of 55 games. OpenAI's GPT-6 Astra showed a similar gap in September. (ARC Prize · ARC Prize, Sep 3)
The case for the maker's setup. You buy the system, not the bare model. If you will run Argon through Google's own tools, Google's setup is the one most likely to predict your results. Keeping private working state between turns is real engineering, and part of what you pay for.
The case for the shared setup. It is the only fair way to compare models. And a company that runs a model inside its own product uses its own setup, not the maker's. Teams are switching often: on Vercel's gateway, 48% of paying teams with steady use changed their most-used model in September. (Vercel)
Our read: use the shared score to compare models, but trust neither score to predict your own results. The provider score shows what Google's own setup achieved. The shared setup gives buyers a more consistent comparison, though this run was cut short. If the model will sit inside your own tools, your own test should decide. If you buy Google's complete product, test that product. And Vercel shows how often teams change models, not whether their results got better or worse. For a buyer, that missing number is the one that matters.
|
Dated tests, each with the source for its date.
Gemini 4 Argon's price doubles · Tuesday, December 1. The introductory price ends November 30. The next test: whether Google extends it, and whether the cheaper "Flex" tier holds. ( Google pricing)
Anthropic's stock sale · November, Reuters reported. The next test: the price range, and whether it sits nearer the $1.5 trillion some investors discussed or the $2 trillion expected before. ( Reuters, via WMBD)
China's proposed limits on exporting AI technology · comments close Sunday, November 15. China's commerce and science ministries listed 40 technologies, 23 of them in AI, including large models and agents. Companies must already register listed technologies before export. The next test: the final list, and whether it covers the open models many companies now use. ( China Ministry of Commerce)
The White House AI task force's 120-day window ends · about mid-February, by our count. Semafor reported the charter gives the task force 120 days. The next test: whether its incident-reporting protocols come with a definition of an incident and a deadline. ( Semafor)
Accenture reportedly capped what it will pay for each employee's Claude use at $29 a month. Some heavy users had been spending far more. Only 404 Media has reported it, and Accenture has not confirmed the cap. 404 Media, Oct 15
Ramp's AI Index: 55.9% of the businesses on its card platform paid for AI in September, and their spending on AI usage was flat for a second month as prices fell. The data covers Ramp's own customers. Ramp, Oct 14
No corrections this issue. We re-checked 43 claims from Issue 017 against their sources. One update: Issue 017 reported break-ins at seven or more South Korean financial firms, which CrowdStrike linked to an AI hacking tool. Korea's Financial Services Commission said on October 14 that 16 firms were hit, with about 100,000 cases of leaked data (a count of cases, not people). It is still investigating whether AI was involved, the Star reported. (The Star, Oct 14)
How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis. Written and edited by Mario Suarez · Independent analysis. Mario works at UnitedHealth Group; this newsletter is his personal analysis from public sources and does not speak for UnitedHealth Group. Anthropic and OpenAI models assisted in creating this newsletter.
|