← Back to articlesAI news
Alexandre11 min read
AI News October 24, 2026: Anthropic's Claude Duet, Google's Agent Compute Engine, OpenAI on Google Cloud
Anthropic launches Claude Duet, Google unveils Agent Compute Engine, OpenAI lands on Google Cloud, and Muse passes 10 million users.
Alexandre11 min read
Credit: dpa, via TechCrunch
On October 21, Anthropic opened the preview of Claude Duet, an agent that works in the background for professionals and small businesses. The next day, Google Cloud introduced Agent Compute Engine, a service that gives each agent its own hardware-isolated micro virtual machine.
Agents run through the whole week. OpenAI puts its models on Google Cloud and is now on all three major cloud providers. Meta announces more than 10 million Muse users, with an average basket above $200 for purchases made by the agent. And according to the Wall Street Journal, Elon Musk ordered a 100 mph cap on Starshield, which cut the connection of Ukrainian drones on October 14.
Breakdown.
Click to enlarge
Google Cloud: OpenAI arrives, and Agent Compute Engine hosts the agents
OpenAI is making its models available on Google Cloud, through the Model Garden catalog and the Antigravity coding tool, in the coming weeks. With Azure and AWS, its models are now offered by the top three cloud providers, less than six months after the end of Microsoft's exclusivity.
The announcement came on Wednesday, October 21 (CNBC). The timeline is tight: OpenAI ended Microsoft's exclusivity over its intellectual property less than six months ago. Amazon announced it would host its models on April 28. Google was the last of the three missing.
For Google, the stakes show in its cloud numbers. Annualized growth of the segment went from about 63% in the first quarter to 82% in the second, driven in part by demand for AI models and in-house chips. Amazon and Microsoft posted 37% and 42% respectively over the same period, and all three report results next week.
Companies that already have a spending commitment with Google can draw on it without a new contract, which is the distribution lever being targeted. According to corporate card data from Ramp cited by CNBC, most enterprise AI spending goes to OpenAI and Anthropic, while Gemini is mostly integrated by some IT services.
OpenAI is also looking for investors at a $1.4 trillion valuation and is aiming for an IPO. Code agents that orchestrate several models in a single workflow have been covered in detail on the blog (architecture of AI code agents), and the multiplication of cloud access points goes in the direction of that kind of architecture.
Agent Compute Engine: one micro-VM per agent
On October 22, at the AGNTCon + MCPCon conference, Google Cloud introduced Agent Compute Engine, the service that already runs Gemini's agents. It will be available in preview in the coming months (Google Cloud). Each agent gets a microVM (a very lightweight virtual machine) isolated at the hardware level, with its own kernel, identity and Hyperdisk volume. Idle microVMs are paused by checkpoint, then resumed from a snapshot when the agent is waiting on a human or an external system.
For teams on Kubernetes, Google offers Agent Substrate, an open source runtime: in a test with 200,000 agents, it sustained 500 activations per second, with suspend and resume latency under 500 ms. It works with any agent framework, including Claude Code, OpenClaw and Hermes.
On the governance side, Agent Gateway, Agent Registry and Agent Identity (based on the SPIFFE standard) extend to Cloud Run, GKE and multicloud. Google also announces, in an upcoming preview, control of local agents such as Claude Code on managed workstations, with an operating-system-level sandbox.
At Meta, third-quarter results provide the first usage numbers for a consumer agent.
Click to enlarge
Meta: more than 10 million Muse users, with shopping in the lead
Mark Zuckerberg announced on October 21, during third-quarter earnings, that more than 10 million people in the United States and Canada set up a Muse agent within a month and a half. About 25% use it every day, and shopping is the number one use.
Meta reported revenue of $64.60 billion, against $63.37 billion expected, up 26% year over year. For the fourth quarter, the company forecasts between $73 and $76 billion (CNBC). The stock rose about 2.5% after the close.
The most discussed detail concerns purchases. More than a quarter of users who go through Muse to buy spend more than $200 per order on average. Zuckerberg cited Plaid, the service that connects payment apps to bank accounts: the company has "never seen" activity at this level on another service (TechCrunch).
The details of the lineup (Muse, Muse Code, Muse Glimmer, the gadgets and how the agents work) are covered in the dedicated article: Meta launches Muse, Muse Code and Muse Glimmer: inside its new AI lineup.
Sensor Tower had measured 10 million US downloads in 40 days, a threshold only 11 other apps have reached, according to TechCrunch.
An agent that is still expensive to run
Zuckerberg acknowledges that Muse consumes a lot of compute in the short term. The stated goal is to improve its cost efficiency by more than 10x within a year, so that the service becomes profitable once revenue scales. Meta plans to monetize Muse by keeping it free for most users, taking a percentage of the commerce it generates, and charging heavy users who want the most advanced models.
Meta also plans a "Muse network," where the agents of friends and family could collaborate, and aims for capacity to serve more than a billion people in the coming years.
Click to enlarge
The same day, October 21, Anthropic opened the preview of Claude Duet.
Click to enlarge
Claude Duet: Anthropic's agent targets professionals and small businesses
Anthropic opened a preview of Claude Duet on October 21, a proactive agent built on Claude Opus 5.5 for professionals and small businesses. It sends daily briefs, builds dashboards and works in the background in its own virtual machine, with human approval for sensitive actions.
Duet connects to more than 6,000 connectors and plugins from the Claude Marketplace. It can make online purchases through Stripe's Link wallet or Shopify's Shop Pay, but does not trade securities and does not move money between accounts (TechCrunch).
Anthropic added a classifier that checks every message to prevent leaks of sensitive information. By default, the agent refuses requests involving bank card numbers or identity documents.
Access is through a waitlist, on iOS, Android and the web, for Pro and Max subscribers. The company says Duet is less comfortable with personal errands, such as booking tickets, and that it "may misread a thread or miscalibrate what is urgent" (CNBC).
An "always-on" agent is an assistant that keeps working when the user closes the app, instead of waiting for a question. OpenAI with Dots, Meta with Muse and Anthropic with Duet now occupy the same ground, with different positioning (consumer for Meta, mixed for OpenAI, work for Anthropic). To understand how these agents rely on tools and memory, see how Claude Code works day to day.
Three days earlier, Anthropic had announced a commitment of a very different kind, on its capital.
Anthropic commits 5% of its shares to a public benefit fund
Anthropic announced on October 18 that a public benefit fund will receive 5% of its shares, with 2% more for each 50% increase in its value above an unpublished threshold.
Anthropic's fund will take the form of a limited liability company 100% owned by the company, not a 501(c)(3) nonprofit. It could finance large donations or approaches to address the effects of AI on employment. The company says it has already committed $200 million to study those effects. According to an August Pew Research report, 71% of American adults think AI will reduce the number of jobs over the next two years (CNBC).
The timing coincides with the IPO expected in November. Anthropic reported $16.8 billion in third-quarter revenue, as detailed in the October 16 AI news. Lise Buyer of Class V Group calls the announcement both necessary to prepare the IPO and "good for the image." Elon Musk replied on X that he "recommends giving the shares directly to the public." On October 21, Anthropic also appointed Harvard physicist Christopher Stubbs to its Long-Term Benefit Trust, the independent body that can appoint a majority of the board (CNBC).
Another matter has occupied Washington since October 14: a speed cap on a satellite network.
Click to enlarge
Starshield: a 100 mph cap that paralyzes Ukrainian drones
On October 14, Ukrainian officials accused SpaceX of limiting Starshield's speed without notice. Starshield is the government version of Starlink. Above 100 miles per hour, drones lost their link. The Wall Street Journal says on October 20 that Musk ordered the measure. Trump says he was not involved.
Starshield is a secure satellite network for governments, built on Starlink technology. The limit affects drones and other devices moving faster than 160 km/h, used by Ukrainian forces. A Ukrainian commander told the New York Times that the cap cost his unit more than a million dollars and delayed missions by three days (CNBC).
The US government side
According to the Wall Street Journal, citing American officials and people familiar with the matter, Musk imposed the cap and discussed it with Trump shortly after the change became public. One of its sources adds that Musk had spoken about it with several senior administration officials before applying it (Quartz).
Musk wrote on X that Starshield systems are "owned and controlled by the US government" and that the limits are "decided by the US military and, ultimately, by the president." On Wednesday, Trump replied: "I don't know who did it, but I was not involved" (CNBC). Volodymyr Zelensky said Monday that his army is still subject to speed restrictions "in different places along the front," with no explanation received from the American side.
Republican Senator John Cornyn believes Musk would not have acted without "at least the acquiescence of the White House." Democratic Representative Pramila Jayapal is calling for an investigation. A bipartisan group of lawmakers is asking that Musk be heard under oath.
Click to enlarge
A dependence that pushes Ukraine to look for other links
Starlink has more than 11,000 satellites in orbit, according to astronomer Jonathan McDowell, and operates in 174 countries according to Gwynne Shotwell, SpaceX's chief operating officer. No competing network offers that coverage today: Amazon's Leo network has 390 satellites deployed, Blue Origin's TeraWave targets its first satellites in the fourth quarter of 2027, and the European IRIS² program has no launch planned before 2029.
Eutelsat, which operates more than 600 satellites, gained 23% on the stock market on October 14 when the restrictions were announced. Consultant Christian Freiherr von der Ropp estimates that these solutions will remain far from Starlink in capacity and coverage, and that the weight of Eutelsat's terminals makes them poorly suited to the battlefield. Zelensky met this week with Norway's defense minister and industrial group Kongsberg to launch new satellite communication capabilities.
A single company controls the link of thousands of drones on a front, and its leadership can change the parameters with one message.
The governance of AI systems is the subject of another story this week, on the OpenAI side.
OpenAI: flaws found by outside evaluators
On Thursday, METR and Redwood Research reported on a pilot audit of OpenAI's internal monitoring system. It revealed previously unknown vulnerabilities, which the company says it has fixed.
From July 21 to August 11, one METR employee and one Redwood Research employee had access to OpenAI's codebase, documentation and some engineering teams, under escort and on devices that never left the premises. They found several "novel vulnerabilities," including a flaw that allowed agents to prevent monitoring of "hundreds" of recent actions. OpenAI says it has fixed them. A 61-page internal report was delivered to the company, and only a high-level summary is public (CNBC).
Sam Altman had promised in September "employee-comparable" access for independent evaluators, following a proposal from Dario Amodei. Donald Trump encouraged AI companies to work with an outside auditor in the voluntary agreement presented at the White House.
Click to enlarge
PHP2Shell: critical flaws discovered with the help of AI models
Two critical PHP flaws, dubbed PHP2Shell (CVE-2026-109618 and CVE-2026-109619), allow remote code execution on servers. Wiz, a Google subsidiary, and researchers working with OpenAI's models identified them with the help of AI models. Patches were published on October 21.
PHP is an open source language that runs a large share of the web. Not every site that uses it is necessarily vulnerable, but the scale of its use worries authorities. The NSA urged organizations on Wednesday to prepare for updates, recommending that affected systems be taken offline if an immediate fix is not possible. The FBI issued a similar warning, as did the authorities of Singapore, the United Kingdom, Australia, Canada and the European Union (CNBC).
Google says Wiz used AI models to find and analyze the flaws, and that Gemini was used to assess their reach, test possible attacks and verify the fixes (Google Cloud). OpenAI says researchers working with its models also contributed to the discovery, and that it coordinated a coalition of maintainers and security vendors for two months before disclosure (OpenAI).
Amazon detected, through its MadPot system, exploitation attempts before public disclosure, without knowing how the attackers obtained the information (AWS). Cloudflare deployed protection rules in its web application firewall pending the updates.
Other news in brief
Nvidia: neocloud demand is "off the charts," capital is short. At GTC Berlin on October 22, Jensen Huang told CNBC that demand is "off the charts" and that working capital is lacking. A neocloud is a compute provider specialized in AI, such as CoreWeave or Nebius, whose shares were down nearly 4% in premarket trading (CNBC).
Nvidia invests everywhere. The company took part in 116 funding rounds in 2026, against 12 in 2022, according to Dealroom, and took stakes in five British startups for $1.16 billion (CNBC).
Meta cuts its headcount by 12% in one quarter. The group has 66,000 employees, after summer layoffs that affected about 10% of teams, or about 8,000 positions (CNBC).
Tesla unveils the next-generation Roadster. Presented on October 15 at SpaceX's test site in McGregor (Texas), it is announced for 2028, with a claimed 0 to 100 mph in under a second, figures not independently verified (CNBC). Reservations start at $50,000.
Zuckerberg commits to staying in California for eight years. The clause comes with the creation of a new class of shares, as the Proposition 40 ballot measure, a wealth tax, could cost him several billion dollars according to Meta.
Reality Labs losses are set to fall. Meta's glasses and virtual reality division is expected to lose $15 billion this year, against $19.1 billion in 2025 (TechCrunch).
Conclusion: agents installed, infrastructure concentrated
Always-on agents are reaching the general public: 10 million Muse users in six weeks, Duet at Anthropic, Dots at OpenAI. OpenAI's models now run at all three major cloud providers, and the average basket for purchases made through Muse exceeds $200.
In parallel, a network change on SpaceX's side was enough to suspend Ukrainian drone missions, and European constellations are not expected before 2029. On governance, Anthropic sets aside 5% of its capital for a public benefit fund ahead of its IPO, while OpenAI fixes its flaws after a three-week external audit. On security, AI models contributed to the discovery of the PHP2Shell flaws in PHP, patched on October 21.
Alex
- Two critical PHP flaws (PHP2Shell) were discovered with the help of AI models from Google (Wiz) and OpenAI, and patched on October 21
- According to the WSJ, Musk ordered a 100 mph cap on Starshield that cut the link of Ukrainian drones, costing one unit more than a million dollars and delaying missions by three days
- OpenAI arrives on Google Cloud (Model Garden and Antigravity, in the coming weeks) and becomes available on the top three cloud providers
- Meta's Muse passes 10 million users in six weeks, 25% use it daily, and purchases exceed a $200 average basket for more than a quarter of buyers
- Claude Duet, Anthropic's agent based on Opus 5.5, opens in preview on a waitlist for Pro and Max subscribers, with more than 6,000 connectors
- Anthropic commits 5% of its shares to a public benefit fund ahead of its IPO, and METR and Redwood found vulnerabilities in OpenAI's internal monitoring
Related articles
AI News October 16, 2026: Anthropic books $16.8B in a quarter, Gemini 4 Argon at $2, Character.ai goes bankrupt
Anthropic reports $16.8B in Q3 ahead of its IPO, Gemini 4 Argon ships at $2, Meta teams up with Apple, Microsoft takes on OpenAI and Character.ai files Chapter 11.
AI News October 11, 2026: Mistral Large 4 goes open-weight, OpenAI publishes 722 math manuscripts, Haiku 5.5 at $0.10
OpenAI publishes 722 math manuscripts, Mistral launches Large 4 open-weight, Anthropic ships Haiku 5.5 at $0.10, and OpenAI's revenue figure sends the Nasdaq down.
AI News October 2, 2026: OpenAI launches Dots at DevDay, Gemini 4 Argon locked down, Sonnet 5.5 hits floor price
OpenAI launches Dots and GPT-6.1 Sol at DevDay, Google keeps Gemini 4 Argon locked down, Anthropic ships Sonnet 5.5. Same price. Here's the breakdown.Comments
Comments
Got a take on this article?
Create a free account in 10 seconds to comment, like, and get the next articles straight to your inbox.
Don't have an account yet?