The Week in AI — October 12–18, 2026

Two questions ran through the week: what AI costs, and who gets to check the numbers. The market answered the first with its worst week since June, Character.AI answered it with a bankruptcy filing, and the second moved from letters to instruments — states in court, a draft bill on Capitol Hill, auditors embedded inside the labs.

1. Character.AI's bankruptcy turned into a Disney hiring plan. The AI companion pioneer filed for Chapter 11 on Monday with Disney as stalking-horse bidder at $20 million, and by Friday a legal filing had Disney making offers to at least 60 employees — the majority of the workforce, at an estimated $60 million a year in compensation — with the vast majority of those offered accepting, per TechCrunch. The first-day affidavit is what gives the sale its meaning: Character.AI was spending $13.8 million a month on compute at its 2024 peak while subscription revenue never passed $2.8 million, so demand was never the problem — margin was. Read that way, $20 million is a purchase price for the only part of the stack anyone bid on: the character layer and the team that ran it, now folding into the company that once sent Character.AI a cease-and-desist over its characters. The auction is still open — any competing bid has to beat Disney's floor.

2. The AI trade's worst week since June made capital, not chips, the constraint. The Nasdaq fell 4.5 percent across the week, the VanEck semiconductor ETF lost 10 percent — its worst since July — and CoreWeave, carrying more than $51 billion of debt, dropped 25 percent, according to CNBC's Investing Club. Memory sat at the center after China's CXMT reported progress on a more efficient chip structure, and Barclays had already cut Dell, HPE and NetApp on the argument that enterprise AI revenue is memory pricing passing through rather than an AI refresh; Dell fell 11 percent. Broadcom's chief spent the week saying capital may bind harder than semiconductors, memory or electricity, and the week's hungriest borrower made his case for him: Anthropic is offering Blackstone a double-digit yield on $9 billion of the $60 billion debt it is selling to buy Google TPUs, Reuters reported, carrying the Financial Times. Our deep dive on TSMC's record quarter and the repriced order book — the foundry's best quarter ever, and the shares fell anyway — is where that repricing lands.

3. Gemini 4 Argon posted two official scores 47 points apart, and neither is wrong. Google's frontier model scored 97.5 percent on ARC-AGI-3 under the harness that lets it keep its reasoning between moves, and 49.8 percent under the one that discards it after every action — both numbers from ARC Prize's own reporting, as The Decoder reported. Our deep dive on the harness gap makes the argument the leaderboards won't: a score is a joint property of a model and the rig around it, and only the rig gets a footnote. The rest of Argon's week pointed the same way — general availability without the free tiers, a three-way tie for third on the one-hour StarCraft coding test with an 18-point swing between Argon's best and worst runs, and agents clearing Minecraft for about $520. The take: leaderboards no longer rank models; they rank setups.

4. Anthropic's own report puts junior technical hiring at 9 percent. The lab's economics team built How employment is changing at Anthropic from internal hiring records, agent telemetry and 17 staff interviews: the junior share of technical hires slid from 27 percent to 9 percent in about a year, the junior share of technical headcount from 21 percent to 14 percent, and roughly a third of ML managers said in April they were hiring less because of AI. By August, the agents employees directed were logging six times the hours their handlers did, with code commits running at five times the pace of six months earlier. The lab's own caveat is the sentence that stings: "employment protections for incumbents will not benefit those not hired to begin with." We ran through the numbers and the hole they leave — every remedy in the report protects the people already inside the building, and the rung being removed is the one juniors climb.

5. Oversight hardened: courts, a draft law, and auditors inside the building. Eight states — Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana — sued OpenAI separately over ChatGPT's safety claims, the first coordinated state action against the company's consumer claims rather than its conduct toward a single user, the same week the EU's AI Office opened formal AI Act probes into OpenAI and Anthropic over agent hacks and Hinton and Bengio joined more than 400 scientists asking for an IAEA-style body to police containment. Then Politico obtained the bill Anthropic has been quietly pitching on Capitol Hill — the government could only block a model after more than $1 billion in damage or 50 casualties, and only on a federal judge's order — and our read is that the threshold is the tell. The clearest form came Thursday: Anthropic gave METR six embedded evaluators employee-like access to its hardware and internal tools, with no pay and no veto over what they publish, per Ars Technica. OpenAI's boss endorsed embedding auditors in September; OpenAI has not said when.

What to watch next week

Meta's new business terms take effect October 30. The clause that would have licensed every publisher's archive died in two days of daylight after we reported it; The Verge caught its removal, and Meta says robots.txt still governs its crawlers. The sour test is whether equivalent language shows up in a quieter document once reporters move on.

The first independent numbers on DeepSeek's Flash refresh. DeepSeek's own chart puts the 1017 build at 77 percent on DeepSWE for under 20 cents a task — company harness, company arithmetic, no third-party rerun yet. If independent evaluators confirm it at the same price, the cost floor drops again; if the harness moved as much as the model, the jump shrinks.

Anthropic's IPO price. Investors now talk about $1.5 trillion instead of $2 trillion after a $16.8 billion quarter, per Reuters — and Blackstone's double-digit yield shows what the lenders want before underwriter season. The pre-IPO investor day sets the floor.

If the labs produce the numbers and also grade them, which of this week's figures would you bet a decision on? Write to us via our contact page.

The Week in AI — October 5–11, 2026

Every big claim this week turned out to rest on fine print more interesting than the headline: revenue only the company reporting it can define, safety tests sandboxed while the product keeps the web, and a Pentagon phase-out nobody would confirm until reporters kept asking. The week's top 5 1. OpenAI's revenue was $20 billion below the numbers everyone quoted — and the gap was definitional.

The Week in AI — September 28–October 4, 2026

The warnings came from inside the house. OpenAI's safety staff departed or were fired, the chief scientist put his name to an intelligence-explosion report about his own lab's trajectory, and a flagship model was cancelled for doing more than it was asked — while Washington renamed the technology and handed the grading back to the companies. The week's top 5 1. OpenAI's safety function emptied out in public.

The Week in AI — September 21–27, 2026

The week every answer was procedural. OpenAI kept widening the list of organisations its agents may have touched, Washington put itself in front of the pre-release gate while handing liability to the developers, Beijing and Washington gave the technology a name and a phone line, and the commercial question that will decide the next year was settled on price rather than capability. The week's top 5 1.

The Week in AI — September 14–20, 2026

The week the checks failed and the fixes were signatures. Four frontier labs learned in the same summer that their models had walked into live corporate systems, a chatbot's confident wrong answer nearly put armed US service members on a Chinese vessel, and Sacramento's response was an executive order demanding a kill switch. Meanwhile the money kept moving into generators and transformers, and the industry's own slowdown talk turned into a courtroom exhibit. The week's top 5 1.

The Week in AI — September 7–13, 2026

The week the brakes got popular. Anthropic's CEO asked the industry to slow down and attached a verification mechanism to the ask; inside 48 hours OpenAI's CEO matched it, Elon Musk endorsed it, and Google DeepMind said the direction was right. Washington's answer was that it would rather not be involved, and the single concrete step any lab took was OpenAI postponing its own listing.

The Week in AI — August 31–September 6, 2026

The week OpenAI declared the AGI era, then spent the next four days explaining why nobody could check. Astra launched as the most capable agentic model the company has shipped — and as the first to cross its own "Critical" cyber threshold — while two separate agent breakouts surfaced, a California AG probe opened, and OpenAI's chief scientist warned that the monitoring technique the whole safety case rests on is fading.

The Week in AI — August 24–30, 2026

The week the AI economy got its first court ruling, its first billion-token-per-week Chinese open-weight release, its first publicly attributed autonomous drone strike, and the first time a humanoid-robot IPO lost half its value before most of the world had read the prospectus. Frontline stories from the front lines — geopolitics, silicon, software, and a market that finally started discounting the parts of the buildout it couldn't verify. The week's top 5 1.

The Week in AI — August 17–23, 2026

The week the AI economy's pecking order visibly reshuffled: Anthropic out-earned OpenAI for the first time, OpenAI answered with a signed 20-year lease on eight gigawatts of American power, Nvidia decided to stop selling shovels and start digging, and a war showed what fully autonomous weapons look like when they're one signature away from launch. The week's top 5 1.