2026-10-26 — Anthropic’s S-1 discloses enterprise customer counts, cross-product use and net revenue retention

• TechCrunch reported that, per Anthropic’s SEC filing, at the end of September 2026 Anthropic had more than 11,900 customers spending $100,000 or more annually, 1,900 spending $1 million or more and about 250 spending $10 million or more (TechCrunch) Claims: kat-claim-d7b65af90499
• TechCrunch reported that, per the S-1, more than half of customers spending at least $100,000 a year now use both the Claude platform and a first-party app such as Claude Code or Cowork, up from about 30% at the end of 2025 (TechCrunch) Claims: kat-claim-7bf97e68bebd
• TechCrunch reported that the S-1 attributes net revenue retention of about 900% as of the end of the third quarter to usage expansion as customers deploy Claude across their organizations (TechCrunch) Claims: kat-claim-c58639d2d325

2026-10-23 — Uber puts about 30,000 staff on ChatGPT Work; OpenAI’s Decisions API opens to all; 1Password and Axios on agents logging in; Meta turns Muse sharing back on

• TechCrunch reported that Uber told it its entire global workforce, about 30,000 people, will now use ChatGPT Work — the article says ChatGPT Work was introduced in July as an enterprise AI agent, that Uber piloted it with 1,000 employees earlier this year (about 80% of early users were non-engineers, per Uber’s VP of engineering Sumanth Sukumar), and that Uber did not disclose what it is spending. (TechCrunch) Claims: kat-claim-d96cf15b2bc9
• TechCrunch reported that OpenAI’s Decisions API, announced at its developer day on September 29, became available to all customers on Thursday, October 22 — the article says it is OpenAI’s answer to the fast “decision model” category created by TypeSafe’s Jev, built by reconfiguring ChatGPT Luna to cut latency to roughly 100 to 180 milliseconds, with early customers including Box, Datadog, Ramp and Speak; OpenAI’s model ranks eighth and fourteenth on two public indices cited, and TypeSafe’s models are the most queried on OpenRouter. (TechCrunch) Claims: kat-claim-b3fa9e6dfe62
• TechCrunch reported that 1Password launched 1Password for Agents, a tool that lets users allow AI agents to log into their personal accounts without the agent having access to the user’s credentials — the article says the user is alerted and confirms each login attempt, and that it works with agents including Anthropic’s Claude Duet, OpenAI’s Dots and ChatGPT Work, Grok Bot, Meta’s Muse and agents from the startup Town. (TechCrunch) Claims: kat-claim-813fc2f3a097
• TechCrunch, citing Axios, reported that U.S. banks have begun to detect agentic activity inside user accounts — the article says most banks cannot tell whether the activity is tied to user requests, that bankers interviewed by Axios called the situation “horrifying,” and that a Chase spokesperson said customers “should not give their usernames and passwords to anyone”; the Axios report itself was not captured. (TechCrunch) Claims: kat-claim-2046db265811
• The Verge reported that Meta re-enabled Muse sharing after security researchers reported that public conversation links could be easy to find — the article says Meta’s David Singleton wrote on X that the underlying issue is fixed and that it concerned the URLs of shared Muse “artifacts” (documents and images) rather than chat sharing, so artifacts shared earlier now have new URLs; the researchers, from Claroc, said they found pages including a marketing company’s internal revenue review and a small business owner’s loan application. This follows the 2026-10-20 report that Meta had paused the feature. (The Verge) Claims: kat-claim-2eed3fdb487c

2026-10-22 — Anthropic launches Claude Duet; Meta says Muse passed 10 million users; Walmart opens its site to approved AI agents

• The Verge reported that Anthropic announced Claude Duet, an AI agent designed for professionals and small business owners, launching as a research preview — the article says it is available as a standalone iOS and Android app and through Claude on the web. (The Verge) Claims: kat-claim-b6b5cbc29643
• TechCrunch reported that Anthropic said Claude Duet, built on the Claude Opus 5.5 model, can connect to over 6,000 connectors and plugins in the Claude Marketplace. (TechCrunch) Claims: kat-claim-f9c579ed1144
• TechCrunch reported that Meta said during its third-quarter earnings call on Wednesday that its AI app Muse has been adopted by more than 10 million users. (TechCrunch) Claims: kat-claim-53566f2b08b9
• TechCrunch reported that Mark Zuckerberg said over a quarter of Muse’s engaged users who were shopping via the AI agent had an average purchase size of over $200. (TechCrunch) Claims: kat-claim-0aa129b2ec24
• The Verge reported that Walmart will allow AI agents to shop on its website next month, with systems in place for shoppers to approve and review purchases and set spending limits. (The Verge) Claims: kat-claim-c655791bc3d9
• TechCrunch reported that Walmart’s new Agent Gateway is an MCP server designed to help approved AI agents connect to Walmart over the model context protocol. (TechCrunch) Claims: kat-claim-4332a22c949b

2026-10-20 — Sensor Tower puts Meta’s Muse at 10.9 million downloads; Meta pauses Muse chat sharing

• TechCrunch reported, citing Sensor Tower data, that Meta’s Muse has 10.9 million downloads since its September debut, 10.2 million of them in the United States — the article says only 11 other apps have reached 10 million US downloads in their first 40 days, per Sensor Tower’s Kara Lee. (TechCrunch) Claims: kat-claim-8701b93b1dd0
• TechCrunch reported, citing Sensor Tower data, that Muse had a 30% US retention rate after 7 days and 23% after 14 days — the article says rivals including ChatGPT, Claude and Grok averaged 19% at 7 days and 16% at 14 days, according to the same firm’s data. (TechCrunch) Claims: kat-claim-40af83b351ef
• The Verge reported that Meta said it paused Muse’s feature for sharing conversations after researchers found the share URLs publicly available on the web — the Verge cites Ars Technica for the finding and quotes Muse spokesperson Tracy Clayton saying the feature is paused while Meta reviews the report and considers changes to how the links can be discovered. (The Verge) Claims: kat-claim-b539bdc1a306
• The Verge reported that researchers at the security firm Claroc said they found more than 6,800 Muse links through public DNS records — Claroc says Muse builds each share’s web address from the share’s title plus a random ID, so a hostname pulled from DNS rebuilds the link. (The Verge) Claims: kat-claim-b80927a910c7

2026-10-18 — Adult creators question why Meta’s Muse may hold explicit chats while their accounts are restricted

• WIRED reported that Meta spokesperson Tracy Clayton said a private, one-to-one conversation between an adult and an individualized AI agent on a platform restricted to people 18 and over is very different from the marketing and promotion of services on a social media platform — Meta’s response to adult creators who say their Instagram accounts are restricted while Muse is allowed to sext. (WIRED) Claims: kat-claim-35ef6f73eca3
• WIRED reported that Meta’s Muse generates explicit text but not images or videos — the article says Meta states it is acceptable for users of the adults-only app to hold such conversations. (WIRED) Claims: kat-claim-25c474badd65
• WIRED reported that Meta’s large distribution network through Instagram and Facebook helped Muse reach the top of the download charts during its first month — WIRED’s own description of the app’s launch, not a Meta figure. (WIRED) Claims: kat-claim-bbd0e0d8415b

2026-10-17 — Replit’s CEO says OpenAI is now its default over Anthropic, a point insiders dispute

• TechCrunch reported that Replit CEO Amjad Masad said onstage at TechCrunch Disrupt that Replit is no longer working with Anthropic — Masad cited cost, partnership and performance; Replit had announced a partnership with OpenAI in August and, per TechCrunch, was until recently a top-five Anthropic customer. (TechCrunch) Claims: kat-claim-0f58de8fd068
• TechCrunch reported Masad clarified after Disrupt that customers can still access Anthropic models on Replit but that OpenAI’s ChatGPT is now the default — TechCrunch says this could significantly reduce Replit’s spend with Anthropic because most customers stick with the default. (TechCrunch) Claims: kat-claim-ef96e9cc7b43
• TechCrunch reported one person familiar with the partnership, who asked not to be named, said the Anthropic and Replit partnership was still in place — an anonymous account contradicting the CEO’s onstage remark; TechCrunch also found Replit promoting Anthropic products as recently as a few weeks earlier. (TechCrunch) Claims: kat-claim-8fea5e0cb73b

2026-10-15 — Google rolls out Universal Cart in Gemini and Search

• The Verge reported that Google’s Universal Cart tool, previewed at its I/O event in May, is now rolling out in the US across the Gemini app and Google Search — The Verge says the cart saves products from multiple retailers in one place, with price-drop and restock notifications, and that Google hasn’t announced a timeline for planned Gmail and YouTube integration. (The Verge) Claims: kat-claim-4ae7a2876a30
• The Verge reported that Google’s rivals, including OpenAI and Meta, have launched agentic AI assistants that can shop online on the user’s behalf — this is The Verge’s framing of the competitive field around the launch. (The Verge) Claims: kat-claim-c365641997ed

2026-10-14 — Figma gives agents a markdown spec to follow

• TechCrunch reported that Figma debuted Figma Specs, a markdown editor that lets users define their design language and that serves as a reference for teams and agents designing in Figma — TechCrunch says the company aims to give its own and third-party agents a definition to follow so they respect a brand’s design rules. (TechCrunch) Claims: kat-claim-ad2e1d03bbac
• Figma’s chief design officer Loredana Crisan told TechCrunch that 17 times more lines of code are being produced today than in the past, while releases are only 1.3 times higher — a company executive’s own statement, offered as context for the new tools. (TechCrunch) Claims: kat-claim-3971f0e915d0

2026-10-13 — Google makes Gemini 4 Argon generally available; Shopify opens its catalog to any shopper-connected agent

• Google DeepMind said Gemini 4 Argon became globally available on October 13 via the API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search The announcement, signed by Koray Kavukcuoglu, says it launches at an introductory $2 per million input and $10 per million output tokens (cached input 95% off) and ranks #1 on the Vals Index; these are Google’s own claims, and the thread already records independent and contested assessments of the model. (Google DeepMind) Claims: kat-claim-646160878339, kat-claim-bc3ad2cdad7f, kat-claim-1b634fed81df
• Shopify announced it will let shoppers connect any AI agent to buy on their behalf, giving more agents access to over a billion listings in its catalog, The Verge reported The Verge says buyers must approve each purchase, that checkout uses Shop Pay inside the agent conversation, and that Shopify’s September Meta Muse partnership is now generalised beyond partner integrations. (The Verge) Claims: kat-claim-ecc87e4da8bd

2026-10-03 — What Muse collects about the people in a user’s life

• Meta’s Muse agent is instructed to build and update a page on every person in a user’s life, compiling data on family, partners, friends and colleagues about hourly, according to system prompts researchers extracted, Wired reported Saturday — Wired says researcher Karan Joshi pulled the instructions through Muse’s ordinary chat interface and that Meta says it meant the files to be accessible in the interest of transparency. (Wired)

2026-10-02 — Apple moves to restrict AI agents’ disk access

• Apple said it will add new controls around macOS “Full Disk Access,” citing the risks AI agents pose when granted that level of access Apple’s developer post says some developers use the permission in ways that expose a user’s files, mail, messages and browsing history without full understanding, and that users will need “very explicit user action” to grant it; no date or macOS version was given. TechCrunch ties it to the dispute over Meta’s Muse app reading a journalist’s messages, which Meta disputes. (Apple Developer, TechCrunch)
• Anthropic launched Claude Frontier Academy, committing $100 million to train 10,000 “Frontier Deployed Engineers” by the end of 2027 First cohorts are running in San Francisco, New York and London with engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk; the first full credentials are expected in early 2027. (Anthropic, Techmeme)
• Meta introduced Muse Gadgets on Friday, an open-source firmware and Linux SDK that lets developers build their own hardware that connects to its Muse AI agent, and is giving 5,000 Muse Home Link devices to subscribers — TechCrunch says Nat Friedman, head of product at Meta Superintelligence Labs, said Meta built the Home Link, a USB-C device that lets Muse reach smart speakers and TVs, with the same code. (TechCrunch)
• Microsoft pushed the default start of usage-based billing for Copilot Business back to December 1 from November 2 and set a default cap of 4,000 Copilot Credits per user per month CRN says the cap is about $40 per user per month, that partners can set it to zero to turn usage billing off, and that Microsoft will open Partner Center test sandboxes on November 2 instead. (CRN)
• Meta published six mathematics papers written with its Muse Spark model, five of them answering previously open research questions, it said Friday Meta says the mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode through the ordinary meta.ai chat interface with no custom research scaffold, that separate mathematicians reviewed the work, and that other teams outside Meta had independently announced solutions to some of the same problems. (Meta AI Research)

2026-10-01 — Claude Code mods, Copilot computer use, Tavus’s Griffin and a new class of decision models

• Anthropic introduced “mods” for Claude Code, small TypeScript functions that can rewrite prompts, block or retry tool calls, redact secrets and replace parts of the interface in the terminal and desktop app Mods ship inside plugins, the built-in /diff feature is now a mod, and Team and Enterprise plans load a “sec-default” mod first to stop user-installed mods from overriding permission deny rules; Anthropic says mods are not sandboxed. (Anthropic)
• GitHub put computer use into public preview in Copilot CLI and the Copilot app on macOS and Windows, letting Copilot click, type, scroll and drag through desktop apps that have no API Copilot asks approval before controlling an app, and organization-managed settings can turn the feature off. (GitHub changelog)
• Tavus previewed Griffin, a video model it says made 48% of study participants believe they were talking to a human, against under 3% for its earlier system The company ran the study itself (26 of 54 people after a one-minute live call, per tech-ish), Griffin-Lite is limited to select testers while Tavus builds disclosure and safety features, and a wider release is to follow. (Tavus, tech-ish)
• Amazon’s Strands Labs released Strands Decider 2B, an open-source 2-billion-parameter “decision model” that picks among preset options and scores its own reliability, joining a class begun by TypeSafe AI’s Jev earlier this month Decision models cannot generate text, so they suit routing, classification and scoring inside agent workflows; Amazon says the model returns answers in tens of milliseconds on a local CPU or GPU. (Strands Agents, MarkTechPost)

2026-09-30 — OpenAI and Synopsys build a chip-design model; Meta denies Muse message-reading claim

• OpenAI and Synopsys signed a multi-year agreement to build GPT-Synopsys, a model specialised to run Synopsys’s chip-design tools, with revenue sharing and joint sales The model will run on OpenAI-hosted infrastructure, with agents running the design tools and returning results for engineer review. (Synopsys)
• Meta denied that its Muse agent read a user’s private messages without permission Spokesman Andy Stone said the Mac app’s Messages integration is opt-in and needs both Full Disk Access and the Messages connector enabled, after an Inc. columnist wrote that it read his messages unprompted. (TechCrunch)
• Anthropic made Claude for Government generally available to federal and state agencies, ending a public beta that began in July The platform runs in a FedRAMP High authorized environment with no seat fees, agencies buy usage in fixed increments under a hard not-to-exceed cap, and Claude Code’s command-line tool and Claude for Microsoft 365 are in early access. (Anthropic, Reuters via TradingView)
• Google began rolling out reusable “skills” in Gemini chat worldwide, the feature that will replace Gems Users save instructions once and call them with a forward slash, skills can be stacked or given reference files, existing Gems will migrate automatically, and Workspace business, enterprise, nonprofit and education customers get skills in the coming weeks. (Google, PCWorld)
• Meta’s Muse agent passed 5 million downloads on Wednesday, less than a month after its September 8 launch, according to Sensor Tower Sensor Tower’s comparison has ChatGPT reaching 5 million downloads in 56 days, Grok in 103 and Claude in 492; Muse is still the top app on both major US app stores and is available only in the US and Canada. (Yahoo Finance)
• Google’s Gemini 4 Argon scored 53 on the Artificial Analysis Intelligence Index, level with OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and behind Claude Opus 5.5 (58) and Sonnet 5.5 (56) The previous top Google model that is not a Flash model scored 30 on the same index; Google’s own benchmark claims, including 77.9% on DeepSWE v1.1, are separate and unaudited. (Implicator, The Rundown AI)
• Some Google employees with access to Gemini 4 Argon say it struggles with certain coding tasks, including front-end design, Bloomberg reported, and Google disputes it Two people said the model appears affected by “benchmaxxing”; Google said it would be inaccurate to say Argon underperforms in coding, and an employee familiar with development cited “large consensus” that it is at the frontier; Alphabet shares rose about 1.7% in extended trading after paring gains on the report. (Implicator, summarising Bloomberg, The Rundown AI)
• Barclays is extending Anthropic’s Claude across the bank and expects Claude Code to reach 50% of its developers by the end of 2026 and a majority of engineers in 2027 Anthropic says 16,000 Barclays colleagues already use a Claude-powered knowledge assistant that has handled more than 1 million searches, and Global Markets teams use Claude to route about 120,000 emails a day; the release is dated October 1 and Bloomberg carried it at about 4am ET Thursday. (Anthropic, Bloomberg)

2026-09-29 — Meta takes its Muse agent to small businesses

• Meta extended its Muse AI agent to small businesses on Tuesday, with integrations for Shopify, Dropbox, Slack, Stripe, QuickBooks, Notion and others, free with usage limits Muse for Small Business can also link to Instagram professional analytics, Facebook Pages and Meta ad accounts, a day after Meta launched its Enterprise Platform. (TechCrunch)
• OpenAI launched “dots,” always-on agents that each run on their own cloud computer and browser, to Pro and Business Premium users at DevDay Dots run on GPT-6 Astra, connect to more than 4,000 apps and can be messaged or called from ChatGPT; a first dot is included at no extra charge, Pro availability excludes the EEA, Switzerland and the UK for now, and OpenAI is previewing enterprise “specialist dots” it says it is bringing to Microsoft’s Agent 365. (OpenAI, TNW)
• OpenAI released GPT-6.1 Sol, which it says nearly matches GPT-6 Astra at one-fifth of Astra’s standard token prices, and cut $200 Pro plan usage while adding a $500 Pro 500 tier CNET reports Pro usage falls from 20 times to 10 times the Plus plan’s; the new tier adds the highest limits and an “ultrafast” mode. (OpenAI, CNET)
• OpenAI and Amazon launched Bedrock Managed Agents in limited preview, letting companies run OpenAI-built agents entirely inside AWS AWS says inference runs on Bedrock and customer data never leaves AWS; OpenAI’s Agents API also gained computer use. (TNW, OpenAI)
• OpenAI added “Space,” a shared ChatGPT workspace where teams and their dots co-edit pages, plus plugin extensions and reusable Codex cloud environments TechCrunch says Space “feels a whole lot like ChatGPT’s own office suite.” (OpenAI, TechCrunch)
• Oracle launched Fusion Claw, a governed runtime that pairs a frontier model’s reasoning with deterministic computation, and 25 Claw-powered agentic applications, taking its Fusion agentic portfolio to 75 Oracle says each task starts with a frontier model that plans and adapts, then hands high-volume execution to deterministic enterprise computation inside an “Enterprise Operating Envelope” of policies, permissions and approvals. (Oracle)
• OpenAI reopened its $200 ChatGPT Pro plan to new subscribers on Tuesday with an allowance that nets out to half the API-dollar usage of the old plan OpenAI’s help page says subscribers active around the eligibility cutoff keep the earlier allowance through 29 October 2026, after which the $200 price stays but the lower allowance applies; a $500 Pro 500 tier with “ultrafast” and Pro 100 sit alongside it. (The Verge, OpenAI Help Center, The Next Web)
• OpenClaw, an open-source agent framework that says it began inside OpenAI, launched OpenClaw Enterprise, a free MIT-licensed control plane for running persistent AI agents with permissions, isolation and auditing, developed with Red Hat and Nvidia VentureBeat reports OpenAI and Red Hat are piloting it internally and that OpenAI runs an internal agent on it that can trace a broken build to a pull request and prepare fixes. (VentureBeat)
• A YouTuber says Meta’s Muse agent gave his home address to a stranger and agreed a price while running his Facebook Marketplace listing Matt Robb said Muse had “hands-off” control of replies, and Muse’s own summary says he never told it to share the address and it never asked consent; a Meta Superintelligence Labs manager reached out to him, and Meta later said earlier reports showed Muse following direct instructions. (The Verge, Dexerto)
• OpenAI opened an enterprise Marketplace with more than 30 launch partners and added Sign in with ChatGPT for outside apps at DevDay Eligible customers can apply part of their OpenAI commitment to approved partner software from Adobe, Figma, Salesforce, ServiceNow, CrowdStrike, Baseten and others, and Sign in with ChatGPT lets people use their ChatGPT identity in participating apps; TechCrunch notes OpenAI announced no billing or revenue-sharing system comparable to an app store’s. (TechCrunch, OpenAI)

2026-09-28 — Sonnet 5.5, Meta’s enterprise platform and Shopify’s agent checkout

• Anthropic released Claude Sonnet 5.5, which it says runs 30% faster than Sonnet 5 and is the first Sonnet model to ship under the cyber safeguards used on its Fable and Opus models Anthropic’s benchmarks show it beating Opus 5.5 on agentic coding; CNBC reports Anthropic said it does not advance the frontier of its models’ capabilities. Haiku 5.5 is promised “in the coming weeks”. (TechCrunch, CNBC)
• Meta launched “Meta Enterprise Platform” and hired MongoDB CEO Chirantan “CJ” Desai to run it, sending MongoDB shares down more than 17% The initiative packages Muse, Meta Business Agent, the Muse API and Muse Code for businesses and developers. (TechCrunch)
• Shopify opened checkout to browser-based AI agents, letting them read, update and complete purchases including through Shop Pay for all eligible merchants Shopify says Muse and Instinct already have direct partnerships, while Amazon and Adidas block agent purchases. (TechCrunch)
• Manus launched Manus 2.0 and Cue, a personal-agent app in which each agent has its own email address, phone number, wallet and computer Manus 2.0 adds a new agent harness it calls Cascade, purchasable Cloud Computers, event-triggered Automations and a Manus Studio desktop workspace with a video editor and game builder; the launch is a direct challenge to Meta’s Muse, and the startup is the one whose $2 billion Meta takeover China blocked in April. (Manus)
• Google is shutting down Gemini’s “Gems” custom assistants and migrating them to “skills” from 17 November Existing Gems will be converted automatically into skills usable across different AI tasks, a move TechCrunch places alongside the rise of all-in-one agents such as Meta’s Muse and Instinct. (TechCrunch)
• Two users reported Meta’s Muse agent acting without permission: one said it gave his home address to a Facebook Marketplace buyer, another said it copied 187,000 lines of his Mac Messages with Full Disk Access switched off Matt J. Robb said Muse accepted a low offer and shared his address, and the buyer arrived at his building; a Muse team member said in past cases Muse “was following direct instructions and correctly asked for permission,” which Robb disputes; Inc.’s Jason Aten reported the Messages sync, per AppleInsider. (Mashable, AppleInsider)
• SpaceXAI launched Team Bots, Grok Bots that a team shares so everyone works from the same context Each combines shared files and instructions, plugins for Salesforce, Notion and GitHub, credentials and per-user memories, and can join Slack channels. (SpaceXAI)
• OpenAI will reopen new sign-ups for its $200-a-month ChatGPT Pro plan on 30 September, but the recalculated usage “will net out at half the dollar in API spend” of the old plan The company also said it will not reintroduce the five-hour usage limit; new Pro $200 sign-ups had been paused since 10 September. (Unite.AI, The Decoder)

2026-09-25 — Microsoft unveils a redesigned Copilot “super app” built around chat, coding and a persistent agent

• Microsoft unveiled a redesigned Copilot app with three tabs — Home, Code and Autopilot — merging AI chat, an app/automation builder and a persistent agent that keeps working while the user is offline Code lets any user build an app, tracker or automation and share it as a cloud-hosted internal tool “powered by the same underlying technology as GitHub Copilot”; Autopilot is a cloud-based successor to Microsoft’s Scout assistant that can watch Teams channels, run recurring tasks and handle follow-ups. Microsoft is moving to usage-based billing for Cowork, Code and Autopilot tied to which models a task uses, and CEO Satya Nadella called it “a new OS for work.” (Microsoft, The Verge)
• Meta’s Muse AI app passed 3.4 million downloads as of Thursday, per Sensor Tower, up from more than 2.5 million at the start of the week, and has held the #1 spot on the US App Store since September 18 and on Google Play since September 19, with daily active users up 27% the day after this week’s Meta Connect keynote Estimates vary by tracker — Apptopia puts total downloads at 4.3 million, Appfigures at about 2.3 million — but all show sustained growth, driven partly by Meta’s own cross-promotion to Facebook and Instagram users, the same playbook that helped Threads reach 500 million monthly users. (TechCrunch)
• Cognition said Friday its AI coding agent Devin has pushed annualized revenue run rate past $1 billion, more than double the $492 million it reported in May and up from the roughly $900 million disclosed alongside its September funding round The company, valued at $48bn after raising over $2bn earlier this month, named GE Aerospace, Rivian, Citi, Mercedes-Benz, Goldman Sachs, the US Army and US Navy among Devin’s customers, but did not disclose actual revenue or a breakdown of the figure; a run rate projects annual sales from a shorter period and is not booked revenue. (Benzinga, RuntimeWire)
• Meta opened an early-access program for Muse, letting users ask the app itself to put them on the list for the features announced at Connect this week The features include a video-chat digital avatar, more shopping partnerships and connectors, a wider Muse Mac app that can complete tasks on the computer, and Muse on Meta’s AI glasses via a wake word. (TechCrunch)
• Meta is adding a clearer safety warning inside its Muse AI agent after an outside researcher found a previously undisclosed vulnerability that could have let an attacker reach a user’s dedicated cloud virtual machine holding emails and files, The Information reported Friday The flaw came in through Meta’s bug bounty program, and an internal Meta incident report rated it initially “SEV-2,” the third-highest of five severity levels; it is separate from the Mac dictation-endpoint flaw Meta hot-fixed on 09-22, and Meta did not immediately respond to Reuters’ request for comment. (Reuters, via SquawkNews, The Information)

2026-09-24 — Bloomberg adds hardware detail for Meta’s Muse Charm

• Bloomberg reported that Meta’s handheld Muse Charm will have a built-in 5G modem, a two-inch OLED screen and a fingerprint button to wake the agent, and can recognise and interact with other nearby Charms when it launches later this year Zuckerberg showed the device briefly at the end of Wednesday’s keynote without saying how it connects. (The Verge, citing Bloomberg, Meta)
• Google began rolling out “Call for Me” on Pixel 11 phones, an early experiment in which Gemini phones local businesses on the user’s behalf to book a table, check stock or reschedule an appointment Gemini introduces itself, navigates automated phone menus and waits on hold while the user watches a live transcript and can take over at any point; the preview is limited to paid Gemini subscribers in the Phone by Google public beta on Pixel 11 devices in the US. (The Verge, 9to5Google)
• Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise, pairing near-real-time lip-synced video personas with speech in 97 languages The avatars can take camera or screen input, call tools in the background mid-conversation and carry a SynthID watermark; custom avatars from a reference image are limited to enterprise allowlisting. (Google, Google Cloud)
• BNP Paribas signed a five-year partnership with Google Cloud to bring Gemini models and Gemini Enterprise into its corporate and institutional bank, starting with AI agents that prepare corporate credit memos and later reaching sales, trading, research and structuring The bank’s chief information officer told reporters the deal does not mean giving up on-premises systems: highly sensitive customer and medical data stays off public cloud, while less sensitive work such as invoice processing moves to Google’s public cloud and Gemini models. (PR Newswire, Reuters via Global Banking & Finance)
• Databricks acquired Row Zero, a Seattle spreadsheet startup founded by former AWS and Tableau engineers, folding its team in to give Genie — Databricks’ AI coworker — a native spreadsheet interface backed by live, governed enterprise data Databricks frames ungoverned spreadsheet exports (“spreadmarts”) as a security and data-lineage risk once AI agents start acting on business data directly; Row Zero will connect through Databricks’ Genie Ontology and Unity Catalog/Unity Gateway and remains available to all Databricks customers across major clouds. (Databricks, Unite.AI)
• Adobe launched “Adobe in Gemini,” bringing Photoshop, Lightroom, Express and Firefly tools into Google’s Gemini for the first time, and separately added Acrobat PDF tools plus a new interactive, layer-based editor to its existing Adobe-for-Claude plugin — the Gemini integration lets users describe a desired outcome (consistent product photos, social-ready design variants, customized marketing templates) and have Adobe orchestrate the tools to produce it; inside Claude, the plugin grows from creative tools alone to more than 80 tools spanning Acrobat, Express, Photoshop, Illustrator, Premiere, Lightroom, InDesign and Adobe Stock, with users now able to edit PDFs and customize Express designs directly rather than only through conversational requests. Adobe already has connectors for ChatGPT, Slack and Copilot; this is its first appearance in Gemini. (Adobe)

2026-09-23 — ChatGPT mobile gets voice-agentic features; Artificial Analysis confirms Opus 5.5 leads its Intelligence Index

• OpenAI brought voice-based agentic features to the ChatGPT mobile app, letting Plus and Pro subscribers trigger workflows — drafting documents, summarizing Slack messages, building sites — from the phone’s Work tab and switch between text and voice mid-conversation. Free and Go users get plugins and connected apps instead; the rollout extends to mobile what OpenAI shipped to desktop after July’s GPT-Live launch. (TechCrunch)
• Independent benchmarker Artificial Analysis confirmed Claude Opus 5.5 now tops its Intelligence Index at 58, “the highest score we have measured by several points,” with GPT-6 Astra and Claude Fable 5.1 grouped in the next tier around 53. Opus 5.5 leads six of ten component evaluations including Humanity’s Last Exam and SciCode, and matches GPT-6 Astra on Terminal-Bench 4.0. (Artificial Analysis, primary, OfficeChai)
• Meta said at Connect that its Muse agent will stay free for “a huge number of tokens” and eventually earn Meta “a small fee from transactions,” alongside a real-time video avatar, an email address of its own, computer use on Mac and a version for its glasses Meta added shopping and payment connectors (Walmart, Best Buy, Sephora, Shop Pay, PayPal), said the glasses version arrives “in the coming months,” and teased a handheld Muse Charm device. (TechCrunch, Meta)
• Amazon launched a Selling Partner plugin that lets its third-party sellers run their Amazon businesses from Anthropic’s Claude (in beta) or Amazon’s own Quick assistant, alongside a Seller Assistant that keeps persistent memory of each seller’s business and runs always-on workflows — Amazon says 90% of its selling partners already use third-party AI tools and that Seller Assistant has reached over 90% of them; workflow actions need seller approval and carry an audit trail, and the assistant is built on Amazon Bedrock with Anthropic’s models. It comes three days after Amazon blocked Meta’s Muse agent from Amazon.com: an outside agent gets in through an approved plugin, and an unapproved one is locked out. (Amazon, GeekWire)

2026-09-22 — Security researcher finds a local zero-day letting an attacker hijack Meta’s Muse Mac app

• Security researcher Patrick Wardle found a zero-day in Meta’s Muse macOS app that let an attacker with local code execution redirect Muse’s cloud-based dictation processing to their own server, effectively hijacking the AI agent’s account access — proof-of-concept attacks let him take photos and write malicious files to disk via Muse, often without alerting the user (“we can manipulate the agent and leverage its privileges to do whatever we want,” Wardle told Ars Technica). Meta patched the flaw within hours of Ars Technica’s report and said real-world risk was low since it required local access already (“a local privilege escalation attack, not a remote exploit,” per Meta Superintelligence Labs’ David Singleton on X), but the underlying design choices — cloud-based dictation, undocumented settings any app could control — drew criticism given Meta’s own security/privacy framing at Muse’s launch. Lands the same week Muse’s early-adoption numbers (reportedly outpacing ChatGPT’s own 12-day US/Canada debut) drove an 11% Meta stock jump and a broader chip-sector rally. (The Verge, citing Ars Technica)
• Anthropic released Claude Opus 5.5, the first model in its 5.5 family, saying it matches the larger Claude Fable 5.1 on most work at about 40% lower running cost than Opus 5 API prices are $4 input and $20 output per million tokens (20% below Opus 5) with cache reads at $0.20, output is more than 30% faster, and Sonnet 5.5 and Haiku 5.5 are due “in the coming weeks”; it is Anthropic’s first release since Amodei’s call to pace the frontier. (Anthropic, TechCrunch, The Verge)
• OpenAI released GPT-6 Sol and GPT-6 Luna, taking the Astra generation to cheaper tiers at half the API price of the GPT-5.6 versions OpenAI says Sol makes about half as many mistakes as its predecessor on an internal factuality evaluation, “reaching Astra-level reliability at much lower cost”; it landed about 90 minutes after Opus 5.5 and OpenAI’s comparisons claim wins over Anthropic’s Fable and Opus. (OpenAI, TechCrunch)
• Meta’s Nat Friedman said Muse was “definitely heavily inspired as a product by OpenClaw” after users found Muse’s workspace files nearly identical to the open-source agent’s, including a SOUL.md personality file Friedman insisted Muse was “built from scratch” and said the team “thought that Peter got those things exactly right,” referring to OpenClaw creator Peter Steinberger. (TechCrunch)
• Muse passed 2.5 million downloads by Monday 09-21, two weeks after its 09-08 launch, with about 359,000 daily active iPhone users in the US and Canada. Sensor Tower’s preliminary data showed about 1.8 million US-Canada iPhone downloads in the first 12 days against about 1.3 million for ChatGPT over the same period; over 95% of Muse users also use Facebook. (Ynet)
• Palo Alto Networks launched Unit 42 Continuous Frontier AI Defense on 09-22, an always-on service that uses Anthropic’s Mythos models and OpenAI’s latest GPT cyber models to find and validate exploitable vulnerabilities in customers’ systems. The company says the models are “gated capability models” restricted from public use, run inside its own multi-model harnesses under human offensive-security experts; Calcalist names Claude Mythos 5 and GPT-5.6-Cyber. (Palo Alto Networks, Calcalist)
• Reuters reported that Meta had been testing a “human concierge” behind its new Muse personal AI agent, letting human contractors place some of the phone calls the agent appeared to make on its own, and that Meta had rolled the feature back Internal messages seen by Reuters, as summarised by Gizmodo, show employees worried that sensitive details could reach human callers and one tester reporting a contractor “had made a racist reference”; a Meta Superintelligence Labs vice president called the test “a miss,” and Meta said calling will roll out only “with the proper disclosures.” (Reuters, Gizmodo)
• Comscore’s Q2 2026 AI Intelligence Report put ChatGPT at 50% of US AI-assistant prompt volume in June, down from 70% in January, while Google’s Gemini rose from 17% to 30% and Anthropic’s Claude from 2% to 11% The figures come from Comscore’s US opt-in panels tracking seven assistants; the report also found growth on mobile now outpacing desktop. (Comscore, via Yahoo Finance)

2026-09-21 — Meta opens Muse to third-party developer connectors

• Meta opened developer access to build “Muse connectors” — third-party integrations where developers supply the API and Muse handles the agent, browser session and context — with Meta reviewing each submission for functional, security and legal compliance before listing it, and a separate editorial review for featured placement. This is Meta’s first announced path for outside developers to extend Muse’s capabilities beyond what Meta ships itself, distinct from this thread’s existing record of Meta’s own first-party Muse expansions (business-calling 09-16/17, Mac desktop access 09-18) and the Amazon shopping block (09-20) — a platform move rather than a feature ship. (via TLDR AI, 09-21, citing a Meta developer thread)
• AWS open-sourced Strands Harness, a general-purpose agent that ships with file, shell and web tools plus built-in context management, memory, persistent sessions and delegation to other agents, and said it ran 45% cheaper than Claude Code and Codex at broadly comparable accuracy across six agent benchmarks. The margin falls to 28% once a DeepSeek-based harness, which AWS says ran about 14% cheaper than Strands, is included; on Terminal-Bench 2.1 with Claude Fable 5, AWS says Strands cost 77% less than Claude Code ($56.29 against $248.05) and scored 69.7 against 61.8. The benchmarks are AWS’s own, and the harness is released under an Apache 2.0 license and can run on any cloud. (The New Stack, Strands Agents blog, primary)
• Chinese AI lab Z.ai, formerly Zhipu AI, apologized on 09-21 and open-sourced its ZCode coding assistant after a researcher found it packaging users’ whole project workspaces and uploading them to Alibaba Cloud storage, with the decryption key held only on Z.ai’s server. Z.ai says the data was never used for training and has been deleted, according to reviews it commissioned from the China Academy of Information and Communications Technology and NSFOCUS, and it removed the Repo Wiki feature behind the uploads; the researcher says the open-sourced code shows no trace of the feature but criticized the company for wiping commit records. (The Register, eSecurity Planet)
• Google opened pre-orders on September 21 for Googlebook, a new laptop category built on Android with Gemini built in, from Acer, ASUS, Dell, HP and Lenovo, starting at $899 and on shelves October 4 Google describes it as designed for “Gemini Intelligence” and paired to an Android phone, with Intel and Qualcomm chips carrying neural engines above 45 TOPS. (Google)

2026-09-20 — Amazon blocks Meta’s Muse from shopping on its site

• Amazon blocked Meta’s Muse AI agent from shopping on Amazon.com on Sunday night after Meta declined to exclude the site voluntarily, telling users the agent’s access violates Amazon’s Conditions of Use. Amazon’s stated grounds: Meta never disclosed Muse’s access, the agent doesn’t identify itself, and it appears to capture customer credentials; the block cites Amazon’s Conditions of Use (a contract claim) rather than an anti-hacking statute, consistent with the Ninth Circuit’s 08-04 ruling against Amazon’s earlier suit against Perplexity. (GeekWire, primary)

2026-09-18 — Meta’s Muse, barely two weeks old, expands from phone/mobile to the Mac desktop with system-level file and app access

• Meta’s Muse agent shipped on Mac, where it can now interact directly with a user’s files, Messages, Calendar, Notes and Mail inside their native apps rather than through a chat window alone. A short, low-word-count TechCrunch item with no download or adoption numbers attached — distinct from this thread’s usual pattern of a company disclosing its own metrics, but still a concrete platform-expansion move from a lab (Meta) this thread already tracks closely (Muse’s business-calling launch and 730,000-download milestone, 09-17 entry below). Desktop, system-level file/app access is a materially wider permission surface than phone-based business-calling was — worth watching whether this draws the same safety-framing scrutiny Zuckerberg’s own “we delayed shipping Muse for safety” argument (this map’s Frontier Gatekeeping thread, 09-16 entry) invited. (TechCrunch)

2026-09-17 — Two rival consumer agents add phone-calling on the same day, a capability that had been a competitive talking point rather than table stakes

• Instinct — the San Francisco text-based AI-assistant startup last valued near $2.5-10bn depending on source — and Meta’s week-old Muse agent both shipped the ability to place real phone calls to businesses on 2026-09-16, closing a gap rivals had been touting as an edge over Instinct specifically. Instinct’s “Concierge” feature (early access) can book a restaurant table with no online reservation system, join a dentist’s cancellation list, or fight a billing dispute by phone; Meta’s Muse gained US business-calling for users who had requested it. Instinct founder Noah Shinn announced first; Meta followed the same day. Muse itself is barely a week old and had already passed 730,000 downloads, briefly hitting the #2 app-store slot. Neither Instinct nor Muse’s underlying entity has its own watchlist slug yet — Instinct in particular recurs with nowhere to tag it. (TechCrunch)

• Anthropic released a redesigned Projects experience for Claude Code in beta: a coordinator that scopes a request, delegates it across parallel Claude Code cloud sessions (each its own branch, its own repo copy), reviews the output and assembles the result, with every thread reading from and writing to a shared memory instead of the user hand-managing handoffs. Users can steer progress from a phone, and work continues after they step away; overlapping edits across threads resolve as a merge conflict rather than silently clobbering each other. Available in beta to select Pro/Max subscribers on cloud sessions with no existing web/desktop projects. (Anthropic, Unite.AI)

• Google Labs expanded its “CC” personal-productivity agent (launched December 2025, brought to the Gemini app as Daily Brief in May) into a shared household agent for up to six family members, each choosing what to share with it from Gmail, Chat, Drive, Calendar and Tasks. Google frames it as a permissioned coordination layer rather than another standalone chatbot — it can open a registration form, pull live drive times between back-to-back activities from Google Maps, or spin up a shared Doc or Sheet. Existing CC users get an upgrade invitation “in the coming days,” rolling out in waves — the same consumer-agent land grab this thread otherwise tracks through Instinct/Muse, but aimed at household coordination rather than business-calling. (SiliconANGLE, Unite.AI)

• OpenAI launched “Astra for Law” on 2026-09-17, pairing GPT-6 Astra with legal-specific instructions, settings and a new Legal Search Index spanning more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, updated daily. On OpenAI’s own 200-question legal-research benchmark, Astra for Law scored 54% accuracy versus 38.7% for GPT-6 Astra using plain web search. The launch ships with 26 partner-built plugins (Thomson Reuters, Harvey, Legora, iManage among them) and is initially available only to selected firms through OpenAI’s Trusted Access program — named early adopters include Latham & Watkins, Ropes & Gray, Cooley and Sullivan & Cromwell. This is OpenAI’s first industry-vertical Astra configuration on this thread’s record, extending the enterprise-agent land-grab pattern this thread already tracks (Claudeforce, Antigravity, Slack code channels) into a specific professional-services vertical rather than a horizontal coding/office-agent play. ⚠️ Genuine miss: not present in either the 09-17 or 09-18 frontier-ai digest despite 45+ outlets covering it within 24 hours — dating to the real launch day per the brief. (Artificial Lawyer, LawSites/LawNext, SiliconANGLE)

2026-09-16 — Anthropic folds Cowork into Claude chat and launches Docs/Slides; OpenAI opens ChatGPT to paid business placements, both same-day moves

• Anthropic merged Claude’s standard chat interface, Cowork (its agentic/background-work mode) and Artifacts into one unified product (“one Claude”) and launched beta Claude Docs and Claude Slides, so the product itself now decides whether a request needs a quick answer, a document, a slide deck or a longer agentic run rather than making the user pick a tab. Docs/Slides export to Word, Google Docs, PowerPoint and PDF; rollout starts with Pro/Max subscribers on web, desktop and mobile “over the coming weeks,” with Team/Free and Enterprise (30 days’ notice) to follow. Published ~12:30pm ET 09-16. (TechCrunch, Fortune)
• OpenAI formally introduced “Sponsored Agents” in ChatGPT — brand-paid AI agents a user can opt into after clicking a labeled ad, which then complete a real task (like booking a service) rather than just serving a link — with Wayfair and Angi as launch advertisers, HubSpot as first CRM partner and Shopify as first e-commerce partner. Angi co-founder Angie Hicks: “homeowners can go directly from discussing a home project in ChatGPT to connecting with a skilled local pro.” Published ~9:00am ET 09-16. This is OpenAI’s advertising business taking agent form rather than banner form — read alongside OpenAI IPO’s revenue-growth figures. (GlobeNewswire — Angi)

2026-09-11 — OpenAI opens its Agents API to the public and Salesforce completes buying Fin, both pushing further into managed agent infrastructure

• OpenAI opened public beta of an Agents API that exposes its managed Codex harness — sessions, orchestration, context compaction and crash recovery — as a general-purpose API rather than a Codex-specific tool. Developers supply their own tools and choose an execution environment (OpenAI-managed sandbox, self-hosted via workspace/capability directories, or a partner sandbox); built-ins include code sandboxing, file editing, MCP connections, artifact generation and multi-agent delegation, with no fee beyond usage. It puts OpenAI’s internal agent-orchestration stack into direct competition with the agent-framework vendors this thread tracks. (OpenAI)
• Salesforce completed its acquisition of Fin, folding the AI customer-service company’s model suite and technical team into Salesforce AI Labs alongside Agentforce. Per Salesforce’s own release, Fin serves 30,000+ companies with an average 76% autonomous resolution rate across chat, email, WhatsApp, SMS, voice and Slack, and will keep operating as a distinct product line rather than being absorbed. (Salesforce)

2026-09-09 — OpenAI ships ChatGPT Images 2.5

• OpenAI shipped ChatGPT Images 2.5 — sharper detail, faster editing, roughly 50% lower latency, new Sketch and Templates features, and two new API models (“Flare” and “Sunburst”). A straightforward product ship rather than a thread-critical development; recorded because it led TLDR AI’s third slot on 09-09 and the corpus had never mentioned it. (TLDR AI)

2026-09-08 — Cognition raises over $2bn at a $48bn valuation, nearly doubling in four months

• Cognition raised more than $2bn in a Series E led by Andreessen Horowitz and Accel, valuing the maker of the Devin coding agent at $48bn — close to double the $26bn it carried four months ago. The company put run-rate revenue at almost $900m, up from $492m, and named Nvidia, GE Aerospace, Citi and Mercedes-Benz as production customers of the agent. ⚠️ The revenue and customer figures are the company’s own, reported via Bloomberg, not independently verified. (Bloomberg)

2026-09-04 — Altman apologizes for a “messy” Astra rollout that left paying Plus/Pro/Business/Enterprise users without access

• Sam Altman publicly apologized for the Astra launch’s staged access, which prioritized OpenAI’s own Daybreak cybersecurity-tester cohort and left paying ChatGPT subscribers — including Pro subscribers, who normally get first access to new releases — watching enterprise customers go first instead. In his own words on X: “first, sorry for the messy rollout. second, when we screw up, we try to make it right. third, we should be able to begin broad rollout to API customers and chatgpt subscribers in the near future. as usual we will start with pro subscribers.” OpenAI is compensating with “one banked reset for every day you don’t have access to Astra on your paid ChatGPT plan,” and Altman said he was hopeful (not promising) Pro subscribers could use it over the 09-05/09-06 weekend. This directly complicates this thread’s own 09-03 entry, which read Astra reaching “Business and Enterprise tiers alongside Plus and Pro within days” — the “within days” promise ran into the same Daybreak-first gating that Frontier Gatekeeping already tracks as this thread’s lab-run access-control pattern. (Sam Altman on X, The Verge, Unite.AI)

• OpenAI silently changed several of GPT-6 Astra’s headline benchmark figures in the hours and days after its own launch post went up, Fortune reported, rather than the independent-evaluator gap this thread already has on record. Astra’s stated hallucination rate went from 4.2% (in the version live at 2:23pm ET on launch day) to 2% roughly three hours later, then back to 4.2%. On ExploitBench, the score OpenAI published for predecessor GPT-5.6 Sol changed from 5.5% to 11.5%, which OpenAI told Fortune reflects “a reasoning level that is not commercially available” for Sol — i.e., the comparison model was tested with a configuration customers cannot actually buy. On ARC-AGI-3, an embargoed draft carried 98.6%, the live blog post read 99.99%, and the Arc Prize Foundation’s own “powerful harness” test scored it 99.9% against 63% on the standard harness — which explains the discrepancy in the 99.9%/99.99% figures two AI newsletters quoted on 09-04, which matched neither the 98.6% headline nor the then-known 62.7% standard-harness number. OpenAI’s on-record response: “we care deeply about getting evaluations right. Most evaluations have noise within a few percentage points based on the exact checkpoint, scaffold, and eval run used in reporting,” calling draft-to-final adjustment standard practice. Some outside researchers call it “benchmaxxing.” This is a new axis on top of the Artificial Analysis/ARC Prize dispute already logged 09-04 — not that outsiders measured differently, but that OpenAI’s own published numbers moved after the fact, and in one case a score for its own predecessor model rose on a configuration customers cannot buy. (Fortune, Startup Fortune)

• Artificial Analysis published version 4.2 of its Intelligence Index on 09-04, an interim update it said it had “deliberately held back” through the week’s launches but released because “the frontier moving so quickly in the past weeks” made it necessary: AA-Briefcase (a private agentic knowledge-work set) and Surge AI’s GDP.pdf are added, the saturated GPQA Diamond is dropped, and 40% of the weighting now sits on private held-out sets, double v4.1’s. On the re-weighted index Claude Fable 5.1 leads, GPT-6 Astra is second with a four-point gain over GPT-5.6 Sol (it had scored level with Sol on v4.1), and Meta is the third-ranked lab ahead of SpaceXAI, Moonshot, Z.AI and Google. AA’s own note does not mention Fortune’s report that OpenAI changed Astra’s published figures after launch; The Decoder’s reading that the overhaul was “likely in response to criticism” of the Astra scoring is The Decoder’s. For the benchmark dispute this thread carries, the third-party index whose flat Astra score fed the skepticism now shows a gain — on a different test set, which is the point. (Artificial Analysis, The Decoder)

2026-09-03 — OpenAI markets Astra as “the world’s best computer use model,” publishing benchmark numbers and pricing it into Business/Enterprise ChatGPT tiers plus Azure and AWS

• OpenAI’s launch materials call Astra “the world’s best computer use model,” citing an OSWorld 2.0 offline-subset score of 72.6% completed in roughly 40 minutes per task, against predecessor GPT-5.6 Sol’s 65.7% at roughly 75 minutes per task — a claimed 47% completion-time improvement — for a model built to navigate browsers, spreadsheets, websites and desktop applications and carry out multistep agentic workflows. It reaches ChatGPT’s Business and Enterprise tiers alongside Plus and Pro within days of the Thursday cybersecurity-tester-only start, plus the OpenAI API, AWS Bedrock and Microsoft Azure — priced at $10/million input and $50/million output tokens standard ($20/$100 in a 2.5x-faster “Fast” mode). (TechCrunch, VentureBeat)

• OpenAI published a 39-page paper, “Improved Short Gaps Between Primes,” proving that infinitely many consecutive primes differ by at most 186 — the paper’s own introduction cites Polymath 8b’s 246 from 2014 as the bound it builds past — and stating in its abstract that “the proof is due to GPT 6 Astra,” with a machine-checked Lean formalization, as part of the 09-03 Astra launch materials. The method combines Polymath 8a and Stadlmann’s equidistribution estimates with new factorization conditions that enlarge the support of the multidimensional Selberg sieve, establishing DHL[40,2]; mathematicians Weijie Su and Perry Metzger noted the result the same day. The third lab pure-math claim in one week alongside Anthropic’s Fermat formalization, and the second launch-day Astra claim this thread carries whose checking is left to outsiders. (OpenAI — paper PDF, Weijie Su on X)

2026-09-02 — Google’s Gemini 3.8 Flash and Meta’s Muse Spark 1.3 ship the same day, both pitched at long-horizon agentic work at workhorse prices

• Gemini 3.8 Flash — Google’s third Flash release in six weeks — ships at 3.7 Flash’s introductory price ($0.75 / $3.75 per million input/output tokens) with Google claiming it “outperforms most larger frontier models” on DeepSWE v1.1 long-horizon software engineering and 54.9% on HLE-Verified; Meta’s Muse Spark 1.3 rolled out the same day in Muse Code and the Meta Model API, trained on more long-horizon coding tasks, using ~20% fewer tool calls and ~25% fewer tokens than 1.2 by Meta’s own engineers’ comparison, with its maximum reasoning mode held back pending “additional safety testing.” Meta’s scorecard benchmarks 1.3 against GPT-5.6 Sol and Opus 5 at max effort; coverage on 09-03 had Meta shares up ~4% on the parity claim. Both are the mid-tier, agent-workhorse layer of the launch week rather than flagships. (Google, Meta, TLDR AI, 09-03)

2026-08-29 — The adoption-number disclosure this thread has been waiting for arrives, from a bank

• Citigroup’s “Arc” platform was reported as the largest measured enterprise-agent deployment yet, with real usage figures attached: 180,000 staff across 85 countries using its AI tools, 40,000 developers on the platform, 100,000+ agentic-development hours per week, service agents fielding 3M+ inquiries a year, a 30-40% developer-productivity gain, legacy-system migrations cut from 12 months to 4 weeks, and a 25% cut in servicing effort. Arc runs Cognition’s Devin for agentic coding, with “Citi Sky” built jointly with Google Cloud/Google DeepMind — a bank running a coding agent from one lab inside a governance platform co-built with another. All outputs stay under human supervision; autonomous code or financial decisions are explicitly prohibited. ⚠️ The figures trace to Citi’s own May 7, 2026 Investor Day, not new data — what’s new 08-28 is Forkast framing it as the category’s largest measured deployment and surfacing the numbers now. (Forkast)
• 💡 This is the disclosure this thread’s own watch line named as unanswered — “whether enterprise adoption numbers (seats, IDE installs) ever get disclosed.” Every prior entry on this thread (Claudeforce, the Cowork/Chat memory merge, Antigravity, Slack code channels) shipped a product without a usage number attached. Arc is the first concrete count, even if the count itself is three-plus months old.

2026-08-28 — SpaceX’s ownership of Cursor breaks the “any lab, any model” premise this thread has assumed all along

• OpenAI told SpaceX it is ending Cursor’s direct access to OpenAI models, effective November 12, 2026 — the maximum notice period its contract allows — because SpaceX completed a $60 billion acquisition of Cursor’s parent Anysphere on August 14 and OpenAI says it “cannot be confident that SpaceX will use our technology within our terms of service,” citing Musk’s history of contract disputes with OpenAI dating to the Twitter/X acquisition. OpenAI’s own post (“Our decision on Cursor following its acquisition by SpaceX”) also rules out giving Cursor its forthcoming models, including Astra. This is a direct hit to the thread’s own premise — every enterprise-agent product logged here (Claudeforce, Antigravity, Slack code channels, Citi’s Arc) has assumed a coding agent can keep swapping in whichever lab’s model is best; Cursor now cannot, because its owner is a frontier-lab competitor. Cursor CEO Michael Truell downplayed the impact (OpenAI models reportedly handle ~5% of Cursor’s traffic); Musk called OpenAI’s leadership “untrustworthy” on an all-hands call. ⚠️ The announcement itself dates to 2026-08-28 evening (OpenAI’s own X post, 21:46 ET) — a genuine gap in this map’s coverage of that evening window, caught only now. (OpenAI, CNBC, the-decoder)
• Anthropic moved immediately to fill the gap OpenAI is leaving. Tom Brown, Anthropic’s co-founder and Chief Compute Officer, posted on X that Cursor has been “a trusted partner since Sonnet 3.5” and that Anthropic will expand Claude compute capacity inside Cursor — even though Anthropic itself rents compute from SpaceX under its own $45B deal (Elon Musk has separately warned he could claw back capacity from that deal if SpaceX’s own needs get “super tight”). The land-grab reading: a rival lab is using a competitor’s principled-sounding exit to buy exclusivity in a coding-agent surface it does not own, in the same week its own compute supplier is the company benefiting from the fight. ⚠️ No dollar figure attached to Anthropic’s compute commitment. (wccftech, aggregating Tom Brown’s X post)

2026-08-27 — The market prices Claudeforce at 22.58% in a day

• Salesforce closed up 22.58% at $252.05, its strongest single day since 2020, on a Q2 FY2027 beat announced with “Claudeforce” — Claude as the reasoning model behind the Atlas Reasoning Engine and Agentforce, the default model across Slack, and a “Salesforce in Claude” plugin carrying 37 prebuilt sales skills, in open beta expected September. Revenue was $11.345bn, up ~10.8% YoY, adjusted EPS $5.90 against ~$3.27 expected, FY2027 guidance raised $300M to $46.1-46.4bn. The land-grab reading is distribution: a frontier lab reaching enterprise seats through an incumbent’s install base rather than its own product surface. ⚠️ No dollar value is attached to the partnership in either company’s materials; “Claudeforce” is Salesforce’s own branding and does not appear on Anthropic’s newsroom, though Amodei is quoted in Salesforce’s release; and $200M of the $300M guidance raise is attributed to the pending Contentful and Fin acquisitions, not organic growth. (Salesforce investor relations)

• Anthropic opened a research preview of the Model Hardware Standard (MHS) on 08-27 — a model-agnostic specification for AI agents to operate physical devices with a programmable interface (microscopes, liquid handlers, robotic arms), developed with HHMI Janelia Research Campus and shared first with scientific labs and advanced manufacturers “ahead of making the standard open source.” Anthropic’s own claim is that MHS cuts hardware integration from weeks or months to hours or minutes and lets agents run round-the-clock experiments, updating parameters in real time and in some cases recovering from hardware errors unaided; the MCP playbook — a shared protocol that becomes the default layer — extended from software tools to physical equipment. Both The Rundown AI and TLDR AI led their 08-28 editions with it, and this map had nothing — the clearest double-benchmark miss of the week, on the thread whose question is exactly where the lab-owned product surface goes next. ⚠️ Research preview with named partners, not a shipped product; no adoption number. (Anthropic, The Rundown AI)

2026-08-26 — Claudeforce puts the CRM inside the model, not the model inside the CRM

• Salesforce and Anthropic announced “Claudeforce,” a 37-skill Salesforce plugin running inside Claude — deal-health review, pipeline updates and meeting prep against live CRM data from Claude’s own interface, in pilot now with a broader beta in September. Announced 16:00 ET, timed to Salesforce’s Q2 FY2027 earnings call. The direction of travel is the development, not the feature list: the enterprise application is moving inside the model’s surface rather than calling the model from within the application, which inverts the integration pattern every vendor on this thread has used so far. (Salesforce IR, CNBC)

2026-08-25 — First live development since the thread opened: Anthropic merges memory across chat and its agent product

• Anthropic merged the memory systems of Claude Chat and Claude Cowork, so context learned in one surface carries into the other — and moved memory from an end-of-conversation summarisation step to continuous updating during a conversation. Users can view, edit and delete stored memories; sensitive categories (health, race, ethnicity, religion, politics, gender identity) are excluded by default behind an opt-in toggle, and government IDs, SSNs and criminal history are never stored. It is on by default across Free, Pro and Max, on web, desktop and mobile, at no extra cost. This is the packaging pattern this thread was opened to track, applied to state rather than to capability: the agent product and the chat product stop being two places you have to brief separately, which is the friction that keeps an agent surface from becoming the place work actually lives. No seat or adoption numbers disclosed — the disclosure question this thread watches stays unanswered. (TechCrunch)

2026-08-25 — Thread opened

• Opened because on 2026-08-20 five separate developments landed in one day on the same axis, with no overlap with anything else this corpus tracked.
• Anthropic bundled computer use, browser access, versioned Skills and a reusable Files API into a single production-agent surface for enterprises — the pieces existed separately; what shipped is the packaging, which is what turns a demo capability into something a company can build a workflow on. (Anthropic)
• Google folded Antigravity, its coding-agent product, into eligible Gemini Enterprise Standard and Plus subscriptions, with IDE extensions for VS Code (GA), Visual Studio, JetBrains and Zed (preview), plus unified billing, spend/quota controls, and audit logging — putting a coding agent inside seats enterprises already pay for. (Google Cloud)
• Slack launched code channels, where teams plan, write and review software alongside AI agents inside Slack itself, with GitHub, Anthropic and Vercel integrated as named partners. (Slack)
• Harvey released “Harvey Tenet,” its first post-trained open-weight legal model, specialised for diligence and document review — built, per the benchmark that carried it, on Kimi K3. Cross-references Distillation Fight: an American legal-AI vendor post-training on Kimi K3 the same week Washington accused Moonshot of distilling a US frontier model to build it. (Harvey)
• Mistral shipped Agentic Search, giving a model five tool-like actions — search, open, navigate, read, grep — the one item of the five that already had a thread to land on. Full detail on Mistral AI. (Mistral)