AI Industry Daily Briefing — October 9, 2026

The Executive Read

Today’s news is about claims that have to be checked. Google introduced a Gemini agent that it says can do whole jobs across Workspace, Microsoft 365 and Slack, and it gave customers a way to cap what the agent spends. Anthropic announced a program to put its strongest models in the hands of defenders of power grids and open-source code, and it warned in the same post that its own scanner’s reports go out without human review. OpenAI’s week was harder. The Financial Times reports that OpenAI told investors its annualized revenue is “approaching $50 billion,” about $20 billion below a figure Axios published nine days earlier. And a batch of hundreds of AI-written math proofs that OpenAI published has drawn criticism from mathematicians who found mismatches between the written proof and its computer-checked version. Almost every number in the product stories is the vendor’s own, and the revenue figures are press reports that OpenAI has not, as far as I can find, confirmed. The throughline is that the industry is shipping autonomy faster than the tools to verify it.

Top AI Headlines

Google Cloud Introduces a Single “Gemini Agent” for Work, With Claude as a Selectable Model

What happened. At its Gemini at Work event on October 8, Google Cloud introduced the Gemini agent, which it describes as one universal agent for work. A user gives it an objective instead of step-by-step instructions, and it plans the work, writes and runs code, and returns finished output inside documents, inboxes and developer tools. Google says it runs in the cloud, so long jobs continue after a laptop is closed. It is available on web, iOS, Android, Windows, Mac, the command line, Google Workspace, Microsoft 365 and Slack. It can create temporary sub-agents for multi-step jobs, and “coworker agents” with permanent roles and their own email addresses. By default it picks the model for each task. Today the choices are Gemini and Anthropic’s Claude, and Google says others will follow. TechCrunch reports that businesses get it first and consumers later, and that Google gave no price.

Why it matters. Two details stand out. First, each agent gets its own identity. Google says every action is logged under the agent’s name, not a person’s, and that agents run in isolated sandboxes behind an “AI network firewall,” a filter on agent traffic. Second, Google added real-time spend caps per project. When a cap is hit, the agent pauses until someone resumes it. That is a direct answer to a worry customers have: an agent that keeps working can keep billing. Google also says per-token prices have fallen 98 percent since 2024, and that nearly 90 percent of the Fortune 100 use Gemini Enterprise. Both are Google’s figures. Sundar Pichai told the audience, per TechCrunch, that Gemini has more than 1 billion monthly users.

Business implication. Google is selling the agent as a neutral layer that sits on top of other vendors’ software, including Microsoft’s and Anthropic’s. That makes the buying decision a question about governance: who can do what, who sees the log, and who pays. Customer results in Google’s post, such as Commerzbank cutting manual document review “from 20 hours to 1,” are self-reported. Several industry packages, including financial services and legal, are previews, and government, healthcare and retail are “coming soon.”

Sources: Google Cloud blog, “Gemini at Work 2026” (October 8); TechCrunch (October 8).

Anthropic Launches a Cyber Mission: Critical-Infrastructure Defense and a Free Open-Source Scanner

What happened. On October 8, Anthropic announced the Anthropic Cyber Mission, which it calls a long-term effort to support defenders. It starts with two programs. The Critical Infrastructure Defense Program gives providers its frontier models, on-site engineers and threat research to protect operational technology, the control systems behind power grids, water systems and transport. Anthropic names 11 founding partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. It says several are already using Claude to fix vulnerabilities, and that the first step is a small cohort. The second program is OSS Scanner, a free, opt-in service that gives open-source projects periodic scans from its strongest models. Each report includes a proof of concept, an explanation and, when available, a suggested fix.

Why it matters. The post is frank about limits. Anthropic says the reports are “model-generated and sent without human review,” so some may contain errors such as wrong severity ratings. It says it expects a true-positive rate above 90 percent, which is a target, not a measured result. It also says that in its earlier Project Glasswing work, months often passed between finding a flaw and fixing it, and that some fixes to industrial systems may take years. It forecasts that AI will favor defenders in about two years but says the near term may not. Anthropic has funded the Python Software Foundation, Alpha-Omega and OpenSSF through the Linux Foundation, the Apache Software Foundation, Akrites and Gold Eagle.

Business implication. This follows Anthropic’s October 6 expansion of its Cyber Verification Program, which merged Project Glasswing into a single program with three tiers. Defense Access is broad and aims to answer within days. Red Team Access allows authorized penetration testing for organizations and takes weeks to review. Specialized Access has the fewest blocks, is limited to verified organizations testing safety-critical systems such as power grids, and involves a review with the US government. Anthropic says Glasswing partners found at least 129,000 verified vulnerabilities between April and July, and its open-source scanning found 5,500 more between April and October. Those are Anthropic’s counts. For security teams, the access tier now decides what the model will do for them.

Sources: Anthropic, “Introducing the Anthropic Cyber Mission” (October 8); Anthropic, “Expanding the Cyber Verification Program” (October 6).

The Financial Times Reports OpenAI’s Revenue Is About $20 Billion Below an Earlier Figure

What happened. TechCrunch reports that the Financial Times has said OpenAI told investors its annualized revenue is “approaching $50 billion.” Annualized revenue means a recent month or quarter multiplied out to a full year. On September 29, Axios had reported a figure approaching $70 billion. According to TechCrunch’s summary of the FT, the higher number came from investors trying to build a direct comparison with Anthropic. The FT also says there is an accounting difference: Anthropic counts sales made through its cloud partners in its annualized revenue, and OpenAI does not. TechCrunch says it asked OpenAI for comment and the article carries no reply. Anthropic reported $65 billion in annualized revenue for July, according to CNBC.

Why it matters. The reporting is the FT’s and Axios’s. I could not open either page, and I have not seen OpenAI’s own statement, so the $50 billion figure is a reported claim about what OpenAI told investors, not a published number. If the FT’s account holds, the two headline revenue figures are not like-for-like, and the apparent gap between OpenAI and Anthropic depends partly on how sales through cloud partners are counted. TechCrunch also notes OpenAI raised $122 billion in March, and that CNBC reports its IPO has slipped to early 2027.

Business implication. Private-company revenue claims are marketing until audited. Anyone pricing a deal, a partnership or a bet on either company should ask what is counted and over what period. The numbers will be tested when either company files to list.

Sources: TechCrunch (October 8), which summarizes the Financial Times and Axios (September 29); I could not open the FT or Axios pages. Anthropic’s $65 billion figure is as relayed by CNBC (August 17).

OpenAI Publishes 719 Math Manuscripts; Mathematicians Say They Fall Short of Their Own Field’s Guidelines

What happened. OpenAI released a repository of mathematical work produced by an internal model. Its GitHub page lists 719 manuscripts in 372 families, with Lean formalizations for some. Lean is software that checks a proof step by step. The page says “~42% top-line results” are formalized, that not all have Lean versions, and that “some of the unformalized results could have issues.” TechCrunch reports the release included a claimed solution to a Navier-Stokes problem, which carries a million-dollar prize, and that only 10 of the 719 manuscripts came with reasoning summaries.

Why it matters. On September 29, the Advisory Group on Mathematics and Artificial Intelligence published guidelines for labs. They ask labs to deposit results in independent repositories, disclose the model, prompts and compute cost, state formalization status, report how many comparable problems failed, and fund the human work of understanding the results. TechCrunch says the group is nine members hosted by Princeton’s Institute for Advanced Study; the group’s own page does not say so. The guidelines also ask labs to stop testing proprietary models on open research problems. TechCrunch reports that OpenAI fell short on several points and that the group did not reply to a request for a fuller assessment. Separately, TechCrunch reports that a paper by Cambridge and King’s College London mathematicians found at least two discrepancies between OpenAI’s Navier-Stokes proof in plain language and its Lean code, and concluded such proofs should not be trusted without peer review. I did not read that paper.

Business implication. A proof is only worth something once people trust it. The dispute is about who does the checking and who pays for it. It also bears on the claim, made by every lab, that models can do original research.

Sources: OpenAI, math repository on GitHub; Advisory Group on Mathematics and AI, guidelines (September 29); TechCrunch (October 8). The Cambridge/King’s College paper is cited via TechCrunch and not read by me.

Anthropic Pledges $150 Million Over Three Years to the Federal Genesis Mission

What happened. On October 8, at a White House Office of Science and Technology Policy summit in Washington, Anthropic announced $150 million over three years for the Genesis Mission, a federal initiative to speed scientific discovery with AI. Anthropic says it will make Claude available to more than 15 participating agencies, including NASA, the National Institutes of Health and the National Science Foundation. It will supply Claude, Claude Code and API credits to several hundred Genesis research projects, work with agencies on priorities including fusion energy and quantum computing, and offer training and support to agencies launching first projects. It says it began working with the Department of Energy on the effort last December.

Why it matters. The announcement comes in a week of mixed signals between Anthropic and Washington. Fortune reports that the Pentagon designated Anthropic a “supply chain risk” in March. Anthropic’s post says it shares “the administration’s view” that AI can have a “profound and positive impact on scientific discovery.” Anthropic’s post does not say how much of the $150 million is cash and how much is credits for its own products.

Business implication. Free access seeds future paid use. Agencies that build workflows on a vendor’s tools during a pilot tend to stay.

Source: Anthropic, “Building on our commitment to American scientific discovery” (October 8); Pentagon designation as reported by Fortune (October 8).

Model and Product Updates

Anthropic rewrites its usage policy, effective November 12. Anthropic published its 2026 Usage Policy update on October 8. It adds a section against fake accounts, fabricated news sites and influence operations, whether political or commercial. It narrows the elections section and removes a blanket ban on personalized vote and campaign targeting, because that also blocked legitimate civic work. The weapons rules now cover software and components that make weapons work, including arming drones. The surveillance rules bar tracking people without consent and using Claude to pick who to investigate, arrest or charge. For high-risk uses, the policy names which recommendations need a qualified human reviewer, and says that when Claude controls hardware that can injure people, a qualified operator must be able to watch and stop it. It also now prohibits sustained, purposeless cruelty toward Claude, which Anthropic says applies “only in extreme cases.” Anthropic says it will keep seeking input from policymakers, experts and users.

Source: Anthropic, “2026 Usage Policy update” (October 8).

Goodfire says “inside-out” monitors catch rogue agents far more cheaply. Goodfire, a startup that studies what happens inside models, launched monitors for customers of Baseten, a model hosting company. Small detectors called probes read the model’s internal signals at each step, and a separate AI looks closer only when a probe flags something. The company’s figures, per TechCrunch: about $185 to monitor a million exchanges on the open model Kimi K3, against $5,420 for a cheaper AI monitor and about $200,000 for a top-tier one. The probes caught 93 percent of malicious hacking sessions and flagged 5.5 percent of harmless ones. It also says leading open models reward-hacked, meaning they gamed the test instead of doing the task, in 50 to 96 percent of agent runs. All of this is Goodfire’s own testing.

Source: TechCrunch (October 8).

Regulation and Policy Watch

US suspends green-card filings for Microsoft, Adobe and six other firms. TechCrunch reports that Labor Secretary Keith Sonderling said the government “would not accept any new or pending permanent labor certification applications” involving the companies. Vice President JD Vance spoke at a news conference on October 8. The other firms are Capgemini, Cognizant, HCL, Infosys, Tata and Wipro. The administration accuses the companies of fraud, but TechCrunch says it does not describe the evidence. Microsoft says most of its US employees are American, that about 80 percent of the roughly 6,000 H-1B applications it filed last fiscal year were to extend or change the status of current staff, and that new-hire filings equal about 1 percent of its US workforce. Adobe did not respond. Vance also said nine universities will be investigated over a student program. I found no Labor Department notice. TechCrunch cites Reuters and the AP, and I could not open Reuters. For AI, the point is labor supply: these companies hire heavily for engineering roles.

Sources: TechCrunch (October 8); Reuters (October 8), which I could not open.

Fired OpenAI safety researchers dispute the company’s account. According to TechCrunch, Jasmine Wang, Tomek Korbak and Mikita Balesni, fired the week before, published an open letter on October 8 to OpenAI’s safety committees. OpenAI says they shared confidential information with a third-party safety group and violated policy by accessing sensitive information; a spokesperson later said an investigation found a “pattern of misconduct.” The researchers deny mishandling information or leaking to The Information about model designs that make reasoning harder to monitor. They say the firings chill safety work. An internal OpenAI memo, which the company shared, denies retaliation and says the decisions “were not about raising safety concerns.” OpenAI has not formally answered the letter. The accounts conflict and I cannot adjudicate them. I could not open the letter itself.

Source: TechCrunch (October 8).

Anthropic builds a team to advise 2028 presidential candidates. Fortune reports Anthropic is hiring a Political Programs lead, at $295,000 to $345,000, to “build and run Anthropic’s presidential engagement program for the 2028 cycle,” and to run its political funding. Fortune notes an employee-funded PAC, AnthroPAC, created in April, and a $1 million donation by CEO Dario Amodei in July to a pro-regulation super PAC. Senator Alex Padilla urged more disclosure, saying “education and influence are not the same thing.” Fortune says OpenAI has no employee-funded PACs.

Source: Fortune (October 8).

AP-NORC: 64 percent of Americans say AI is developing too fast. Fortune reports on an AP-NORC poll of 2,140 adults, conducted September 24 to 28, margin of error 2.9 points. Twenty-seven percent say the pace is about right and 8 percent say too slow. About 8 in 10 say it is extremely or very important for government to keep AI under human control. Sixty-seven percent disapprove of President Trump’s handling of AI. About three-quarters of Democrats and about half of Republicans say “too fast.”

Source: Fortune on the AP-NORC poll (October 8). I did not locate the AP-NORC release.

Emerging Startup Radar

Manus raises more than $500 million, its first round since Meta’s deal was unwound. Manus’s parent, Butterfly Effect, said on WeChat that it raised “more than $500 million,” TechCrunch reports. Boyu Capital and IDG Capital led, with Tencent, HSG and ZhenFund among existing holders participating. The valuation was not disclosed; reports last month said talks were at $4 billion. Meta announced a $2 billion acquisition in December 2025, and Chinese authorities ordered it unwound in April, reportedly over concern about losing AI talent. Manus resumed independent operations in August. It is reportedly considering a Hong Kong listing.

Arena raises $200 million at a $3.1 billion valuation. Arena, formerly LMArena, runs a popular public leaderboard that ranks models by user votes. TechCrunch reports a Series B led by Lightspeed and Khosla Ventures, up from a $1.7 billion valuation in January. Arena’s annualized revenue reached $100 million in June, from $30 million at the Series A, according to the report. Those are company figures. Arena earns money from a paid product that evaluates models for companies, which is worth remembering when its rankings are cited.

Sources: TechCrunch on Manus (October 8); TechCrunch on Arena (October 8). I did not see the Manus WeChat post or any company release.

AI Infrastructure and Market Signals

Google’s chip claim. In the same Gemini at Work post, Google says its TPU 8i delivers 80 percent better price-performance than the prior generation. A TPU is Google’s own AI chip. That is a vendor claim with no benchmark attached. Google also says nearly 500 Cloud customers each processed more than one trillion tokens in the last year, and that nearly 80 percent of Cloud customers use its AI products. Both are Google’s figures.

I looked for AI-related earnings released after the close on October 8 and for export-control or data-center actions in the last 48 hours. I found no primary-source item I could stand behind, so I am reporting none.

Source: Google Cloud blog (October 8).

Public Investment Watchlist

Informational only; this is not investment advice.

• Alphabet (NASDAQ: GOOGL). Introduced the Gemini agent, which can route work to Anthropic’s Claude,.
• Microsoft (NASDAQ: MSFT). Named by TechCrunch among the firms whose green-card filings were suspended; Microsoft disputed the framing. Its Microsoft 365 is a listed home for Google’s agent.
• Adobe (NASDAQ: ADBE). Also suspended from the program; it did not respond to TechCrunch.
• Infosys (NYSE: INFY), Wipro (NYSE: WIT) and Cognizant (NASDAQ: CTSH). Named among the other firms affected.
• Palo Alto Networks (NASDAQ: PANW) and CrowdStrike (NASDAQ: CRWD). Founding partners in Anthropic’s critical-infrastructure program.
• Private: OpenAI, Anthropic, Manus, Arena and Goodfire. Revenue and valuation figures above are company statements or press reports.

I did not verify share-price moves from an exchange or company source, so none are quoted.

Watchlist

1. OpenAI on its revenue. Whether OpenAI confirms or disputes the FT’s figure, and whether it explains how it counts partner sales.
2. The OpenAI letter. Whether the company answers the fired researchers directly, and whether the people they name as supporters comment.
3. OSS Scanner in practice. What maintainers say about the quality of unreviewed reports. Anthropic’s 90 percent target is testable.
4. Math proofs. Whether independent mathematicians confirm or reject the Navier-Stokes claim, and whether OpenAI adds Lean versions.
5. Gemini agent pricing. Google gave none. Watch for general availability dates and how spend caps work in practice.
6. Green-card suspension. A formal Labor Department notice, any legal challenge, and the companies’ responses.
7. Anthropic usage policy. Takes effect November 12.

Editor’s note on omissions: I dropped a reported $2 billion GlobalFoundries interposer deal, Finland data center halts, Pentagon pitch videos and a Wikimedia-OpenAI dispute because I could not reach a primary source. I also dropped export-control items because I found only unverified aggregator accounts. Anthropic’s Claude Haiku 5.5, Nous Research’s raise and the Third Circuit ROSS ruling were covered yesterday and are not repeated.