{
  "version": "https://jsonfeed.org/version/1.1",
  "title": "Usewright AI News",
  "home_page_url": "https://usewright.app/news/",
  "feed_url": "https://usewright.app/news/feed.json",
  "description": "Source-based AI announcements and practical editorial context. Event dates are separate from publication dates.",
  "items": [
    {
      "id": "https://usewright.app/news/gemini-4-argon/",
      "url": "https://usewright.app/news/gemini-4-argon/",
      "title": "Gemini 4 Argon rolls out on paid consumer and developer tiers",
      "summary": "Google announced subscriber and paid-API access, a unified token limit and introductory API pricing scheduled to end November 30, 2026.",
      "content_text": "Google announced subscriber and paid-API access, a unified token limit and introductory API pricing scheduled to end November 30, 2026.\n\nWhat happened\n\nGoogle announced Gemini 4 Argon on October 13, 2026. The rollout covers the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers, the Gemini API paid tier, and Antigravity for Ultra subscribers.\n\nGoogle also lists enterprise routes and a researcher access program subject to eligibility, capacity and selection. These distinct routes should not be described as unconditional or free access for everyone. Account entitlements were not tested for this article.\n\nWhat to compare\n\nThe announcement specifies a one-million-token unified input-output limit through the API and points developers to a continuation protocol. Its introductory API pricing is scheduled to end on November 30, 2026, after which higher published rates apply.\n\nGoogle’s performance comparisons and robotics example are vendor-reported results. A demonstration does not establish how the model will perform on your tasks.\n\nWhat to do with this\n\nEditorial suggestion: Compare models on a small, repeatable workload with known checks. Record accuracy, latency, total usage and applicable access limits. Calculate ongoing costs using the rate that will apply after any introductory period, rather than choosing a plan from a launch headline alone.\n\nSources\n\nGoogle: Introducing Gemini 4 Argon: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-13",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/eu-ai-office-loss-of-control/",
      "url": "https://usewright.app/news/eu-ai-office-loss-of-control/",
      "title": "EU AI Office follows up on reported model-control incidents",
      "summary": "The Commission describes incident reports and security questions, without announcing an infringement finding or sanction in this update.",
      "content_text": "The Commission describes incident reports and security questions, without announcing an infringement finding or sanction in this update.\n\nWhat happened\n\nIn an October 12, 2026 announcement, updated October 13, the European Commission described follow-up to reported loss-of-control incidents involving AI models. The AI Scientific Panel met on October 9 to discuss the incidents and related technical developments.\n\nThe Commission says the AI Office had requested and received serious-incident reports and sent questions about security practices. Further information requests combining the panel’s and AI Office’s questions were planned.\n\nWhat the update does not establish\n\nThis account does not name the companies or models, explain the incidents in detail, or announce an infringement finding or sanction. The term loss of control should not be treated as proof that a system escaped oversight permanently.\n\nThe Commission’s AI Act guidance distinguishes duties for model providers, including additional systemic-risk duties, from duties for deployers of AI systems. A headline about one group does not establish the obligations of every organization using AI.\n\nWhat to do with this\n\nEditorial suggestion: Keep a factual incident log recording the system, observed behavior, effects and response. Follow the Commission’s specific update for confirmed developments. Ask your compliance owner which reporting duties apply to your role and use case; this brief is not legal advice.\n\nSources\n\nEuropean Commission: Special meeting of the AI Scientific Panel: https://digital-strategy.ec.europa.eu/en/news/special-meeting-ai-scientific-panel\n\nEuropean Commission: Navigating the AI Act: https://digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-12",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/meta-muse-wonder/",
      "url": "https://usewright.app/news/meta-muse-wonder/",
      "title": "Meta introduces Muse Wonder and subscription-only Wonder Full",
      "summary": "Muse Wonder is available through Meta’s developer tools in preview, while the consumer Muse rollout is still forthcoming.",
      "content_text": "Muse Wonder is available through Meta’s developer tools in preview, while the consumer Muse rollout is still forthcoming.\n\nWhat happened\n\nMeta introduced Muse Wonder and Muse Wonder Full on October 12, 2026. It describes Active Memory as retaining relevant context and Active Focus as sustained work toward a goal. Performance and safety comparisons in the launch posts are Meta’s reported claims, not Usewright test results.\n\nMuse Wonder is available through Meta Model API and Muse Code. The API documentation marks it as preview. Wonder Full is available only through Muse Code subscriptions on Power Usage and Maximum Usage plans.\n\nPricing and rollout\n\nMeta’s launch post lists Muse Wonder API rates in USD per million tokens: $4 input, $0.20 cached input and $20 output. These are usage rates, not an all-inclusive task price or the price of a Wonder Full subscription.\n\nMeta says Wonder will reach consumer Muse experiences later, beginning with subscribers. The announcement does not establish that every consumer account already has access.\n\nWhat to do with this\n\nEditorial suggestion: Evaluate one reproducible coding issue in an isolated project with explicit acceptance tests and a budget. Check changes before merging them. Keep the task brief and failure evidence so another model can be compared on the same work.\n\nSources\n\nMeta AI Research: Muse Wonder announcement: https://research.meta.ai/blog/muse-wonder\n\nMeta: Build with Muse Wonder and Muse Wonder Full: https://dev.meta.ai/resources/blog/build-with-muse-wonder\n\nMeta: API documentation and preview status: https://dev.meta.ai/docs/overview\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-12",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/anthropic-usage-policy-update/",
      "url": "https://usewright.app/news/anthropic-usage-policy-update/",
      "title": "Anthropic clarifies high-risk rules and adds hardware safeguards",
      "summary": "The policy effective November 12, 2026 keeps existing high-risk review and disclosure duties while adding controls for autonomous physical actions.",
      "content_text": "The policy effective November 12, 2026 keeps existing high-risk review and disclosure duties while adding controls for autonomous physical actions.\n\nWhat happened\n\nAnthropic announced its updated Usage Policy on October 8, 2026, with an effective date of November 12, 2026. This brief focuses on its high-risk and physical-action provisions; it is not a complete account of the policy refresh.\n\nAnthropic explicitly says the existing high-risk requirements have not changed. Covered uses still require qualified human review, including authority to change recommendations, and notice to the affected individual that AI was used. Revised language clarifies which recommendations fall within scope.\n\nWhat is new for physical actions\n\nFor models connected to hardware that acts autonomously and might cause injury, the update adds safeguards: a qualified operator must be able to observe and stop the equipment, and the equipment must hold a safe state if Claude disconnects. The linked Usage Policy identifies the November 12 version.\n\nWhat to do with this\n\nEditorial suggestion: List the clauses relevant to your deployment, the responsible reviewer and evidence that each control works. Separate existing obligations from genuinely new controls before planning changes. These are Anthropic’s product-use rules; this summary does not determine your legal obligations or replace qualified advice.\n\nSources\n\nAnthropic: 2026 Usage Policy update: https://www.anthropic.com/news/2026-usage-policy-update\n\nAnthropic: Usage Policy effective November 12, 2026: https://www.anthropic.com/legal/aup\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-08",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/claude-haiku-5-5/",
      "url": "https://usewright.app/news/claude-haiku-5-5/",
      "title": "Claude Haiku 5.5 launches alongside claimable subscription API credits",
      "summary": "Anthropic targets narrow, high-volume work and adds monthly Claude Platform credits for eligible paid Max and Team subscribers.",
      "content_text": "Anthropic targets narrow, high-volume work and adds monthly Claude Platform credits for eligible paid Max and Team subscribers.\n\nWhat happened\n\nAnthropic released Claude Haiku 5.5 on October 7, 2026. It reports an average running-cost reduction of about 75% versus Haiku 4.5, not a uniform per-token discount. Its fastest-model claim applies at standard speeds; Opus Fast Mode is faster. Computer-use and browser-use SDK support is in beta.\n\nCredits need to be claimed\n\nThe official help page lists monthly USD credits of $100 for Max 5x and $200 for Max 20x. Team credits are pooled: $20 per Standard seat and $100 per Premium seat, capped at $500 monthly for the team. These are included paid-plan benefits, not cash or universal free access.\n\nEligible subscriptions must be active for seven days. Claiming requires linking one Claude Console organization and reviewing the credit terms. Team claims require a Primary Owner or Owner; the Console role must be Owner, Admin or Billing. Unused credits expire each billing cycle. They cover eligible Claude Platform usage, not interactive Claude Code or third-party cloud billing.\n\nWhat to do with this\n\nEditorial suggestion: Follow the linked claim instructions and check the receiving organization carefully. For an initial evaluation, use a small, redacted classification or summary task with known answers and a spending limit. Review errors before automating anything consequential.\n\nSources\n\nAnthropic: Introducing Claude Haiku 5.5: https://www.anthropic.com/claude-haiku-5-5\n\nClaude Help Center: Monthly API credits and claim instructions: https://support.claude.com/en/articles/17154008-monthly-api-credits-for-max-and-team-plans\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-07",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/gpt-6-intelligent-ui-for-everyone/",
      "url": "https://usewright.app/news/gpt-6-intelligent-ui-for-everyone/",
      "title": "GPT-6 Intelligent UI begins a staged ChatGPT rollout",
      "summary": "Interactive responses are expanding across ChatGPT tiers, with workspace controls and a separate limitation for the Pro reasoning option.",
      "content_text": "Interactive responses are expanding across ChatGPT tiers, with workspace controls and a separate limitation for the Pro reasoning option.\n\nWhat happened\n\nOpenAI announced GPT-6 with Intelligent UI on October 7, 2026. It describes responses that combine text with charts, forms, buttons and other interactive elements. Answers can begin while the model continues thinking or using tools.\n\nThe announced Chat-tab rollout began with Plus, Pro, Business and Enterprise on October 7, expanding to Free and Go from October 8. Enterprise availability depends on administrator settings. A rollout announcement does not verify access in any particular account.\n\nWhich modes are covered\n\nOpenAI’s release notes identify GPT-6 Sol for the listed paid tiers and GPT-6 Luna for Free and Go. Intelligent UI spans Instant through Extra High reasoning. The Pro reasoning option uses GPT-6 Astra and does not support Intelligent UI. This release does not change the models powering Work or Codex.\n\nWhat to do with this\n\nEditorial suggestion: Supply a small dataset and request a chart with the underlying values shown. Check labels, units and axis ranges before sharing it. Treat a polished interactive answer as a presentation to inspect, rather than evidence that the numbers or assumptions are correct.\n\nSources\n\nOpenAI: GPT-6 and Intelligent UI announcement: https://openai.com/index/gpt-6-for-everyone/\n\nOpenAI Help Center: October 7 ChatGPT release notes: https://help.openai.com/en/articles/6825453-chatgpt-release-notes\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-07",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/synthid-detector-for-everyone/",
      "url": "https://usewright.app/news/synthid-detector-for-everyone/",
      "title": "Google expands public access to SynthID Detector",
      "summary": "The English-language tool checks supported image, video and audio files for participating providers’ SynthID watermarks.",
      "content_text": "The English-language tool checks supported image, video and audio files for participating providers’ SynthID watermarks.\n\nWhat happened\n\nGoogle announced wider SynthID Detector access on October 7, 2026, describing the tool as available globally in English. Its announcement links directly to the official detector.\n\nThe tool checks images, video and audio for SynthID watermarks from Google and participating partners, including OpenAI, NVIDIA and Kakao. Google listed Apple support as forthcoming. DeepMind explains that SynthID embeds an imperceptible signal in supported generated media.\n\nHow to read a result\n\nThis is a check for a particular provenance signal, not a universal AI detector or a verdict that a scene is true. No detected watermark does not prove human authorship. The announcement’s scope should not be stretched to every file from every AI tool.\n\nThe public detector page did not expose usable account or quota information in this source check. No sign-in, upload or functional detector test was performed.\n\nWhat to do with this\n\nEditorial suggestion: Check provenance alongside the original uploader, date and independent reporting. Preserve the file and result if they matter to your work. Do not upload private or confidential media without permission, and avoid making accusations from a detector result alone.\n\nSources\n\nGoogle: SynthID Detector public-access announcement: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/\n\nGoogle DeepMind: How SynthID works: https://deepmind.google/models/synthid/\n\nGoogle: Official SynthID Detector: https://synthid.com/\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-07",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/windows-hybrid-intelligence/",
      "url": "https://usewright.app/news/windows-hybrid-intelligence/",
      "title": "Microsoft outlines local-and-cloud AI updates for Windows",
      "summary": "The October 7 announcements mix available agent containment, planned Copilot features, an experimental routing preview and upcoming hardware.",
      "content_text": "The October 7 announcements mix available agent containment, planned Copilot features, an experimental routing preview and upcoming hardware.\n\nWhat happened\n\nOn October 7, 2026, Microsoft announced general availability of Microsoft Execution Containers on Windows 11. Microsoft says MXC lets organizations restrict the files and networks an agent can access, with enforcement at runtime.\n\nOther announcements have different timelines. Local-and-cloud HydraFusion routing for GitHub Copilot is scheduled for experimental preview later in October. Copilot hybrid-intelligence features are expected on Copilot+ PCs in the coming months, with permission required for local context.\n\nHardware and limits\n\nMicrosoft says the announced RTX Spark laptops begin shipping October 16, 2026. Dev boxes and mini desktops follow later; the October date does not cover the whole hardware range.\n\nLocal processing can change where work happens, but lower costs, privacy improvements and offline operation depend on the application, routing, telemetry, hardware and network requirements. They are not automatic benefits of a hybrid design.\n\nWhat to do with this\n\nEditorial suggestion: Before permitting an agent to change files, test on copies and verify a restore path. Check which model and services each task actually uses. For hardware decisions, wait for relevant independent measurements rather than treating vendor demonstrations as your expected performance.\n\nSources\n\nMicrosoft: Building Windows for hybrid intelligence: https://blogs.windows.com/windowsexperience/2026/10/07/building-windows-for-hybrid-intelligence/\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-07",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/mistral-large-4-preview/",
      "url": "https://usewright.app/news/mistral-large-4-preview/",
      "title": "Mistral previews Large 4; downloadable weights are still planned",
      "summary": "Mistral Large 4 entered public API preview on October 6, with weights promised by the end of October 2026.",
      "content_text": "Mistral Large 4 entered public API preview on October 6, with weights promised by the end of October 2026.\n\nWhat happened\n\nMistral announced Large 4 on October 6, 2026. The multimodal model is available as an API preview through Mistral Studio. Its documentation specifies 1.05 trillion total parameters and 52 billion active parameters.\n\nThe company says downloadable weights will follow by the end of October 2026. That is a release plan, not confirmation that weights are already available. Its announcement already cites independent evaluations alongside company tests; benchmark results should be read with their methods and conditions.\n\nWhat remains conditional\n\nMistral connects the release to its €3 billion Series D. Its separate funding announcement gives a post-money valuation above €21 billion and names Samsung Electronics as the lead investor.\n\nOpen weights may offer additional deployment choices, but practical self-hosting will depend on the released license, infrastructure and operating requirements. They do not by themselves establish regulatory compliance.\n\nWhat to do with this\n\nEditorial suggestion: Compare the preview with your current model using the same non-sensitive tasks and success criteria. Record failures, latency and total cost. Keep any self-hosting decision separate until the actual weights and license can be reviewed.\n\nSources\n\nMistral: Introducing Mistral Large 4: https://mistral.ai/news/mistral-large-4/\n\nMistral: Large 4 preview documentation: https://docs.mistral.ai/models/mistral-large-4\n\nMistral: Series D funding announcement: https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-06",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/openai-math-results/",
      "url": "https://usewright.app/news/openai-math-results/",
      "title": "OpenAI releases mathematical manuscripts and proof artifacts",
      "summary": "The public repository includes results from an unreleased model, with supporting material and varying levels of formal verification.",
      "content_text": "The public repository includes results from an unreleased model, with supporting material and varying levels of formal verification.\n\nWhat happened\n\nOn October 6, 2026, OpenAI announced mathematical results produced by an internal frontier model and released a public GitHub repository. The company says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study about disclosure practices.\n\nThe release includes reasoning summaries, attempt statistics, compute estimates, revision records and citation instructions. OpenAI describes the average compute per result as roughly equivalent to three hours of ChatGPT Pro thinking. That comparison does not establish access to the unreleased model through a consumer account.\n\nWhat the repository establishes\n\nThe repository contains manuscripts and Lean proof artifacts. Its README says verification stages vary, some results lack formalizations, and unformalized work may contain issues. Making a result public is not the same as independent mathematical validation.\n\nWhat to do with this\n\nEditorial suggestion: Choose one claim and trace it to the exact manuscript version, assumptions and available proof artifacts. Look for expert scrutiny and revisions. Apply the same habit to everyday AI answers: separate the claim, the evidence supplied and what still needs checking.\n\nSources\n\nOpenAI: Sharing AI progress in mathematics: https://openai.com/index/sharing-ai-progress-in-mathematics/\n\nOpenAI: Mathematical manuscripts and proof artifacts: https://github.com/openai/math\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-06",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/openai-eu-text-watermarking/",
      "url": "https://usewright.app/news/openai-eu-text-watermarking/",
      "title": "OpenAI announces eligible EU text watermarking and API opt-in",
      "summary": "Watermarking is rolling out regionally, while the initial text detector is restricted to approved researchers and expert organizations.",
      "content_text": "Watermarking is rolling out regionally, while the initial text detector is restricted to approved researchers and expert organizations.\n\nWhat happened\n\nOn October 5, 2026, OpenAI announced a coming-weeks rollout of invisible watermarks for eligible ChatGPT and Codex text across plans in the EU. It is not a worldwide default. API customers worldwide can opt in for selected models; API watermarking remains off by default.\n\nInitial access to OpenAI’s text detector requires approval for researchers or expert organizations. OpenAI warns that editing and short passages can undermine detection, and that false positives and missed watermarks are possible.\n\nThe policy context\n\nThe European Commission’s guidance distinguishes providers’ machine-readable marking duties from deployers’ disclosure duties for specified uses. Technical feasibility and exceptions matter. Article 50 transparency provisions apply from August 2, 2026; that does not mean every piece of AI-assisted text requires identical treatment.\n\nWhat to do with this\n\nEditorial suggestion: Record which tool produced a passage and how it was edited. Do not treat a detector result as proof of authorship, accuracy or misconduct. For a workplace deployment, ask your compliance owner to check the applicable use case and disclosure rules; this brief is not legal advice.\n\nSources\n\nOpenAI: Our approach to EU text provenance rules: https://openai.com/index/eu-text-provenance/\n\nEuropean Commission: Navigating the AI Act: https://digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-05",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/anthropic-claude-frontier-academy/",
      "url": "https://usewright.app/news/anthropic-claude-frontier-academy/",
      "title": "Anthropic commits $100 million to Claude Frontier Academy",
      "summary": "The nomination-based program aims to train 10,000 Frontier Deployed Engineers by the end of 2027, with two assessment stages.",
      "content_text": "The nomination-based program aims to train 10,000 Frontier Deployed Engineers by the end of 2027, with two assessment stages.\n\nWhat happened\n\nAnthropic announced Claude Frontier Academy on October 2, 2026, backed by a $100 million commitment. Its goal is to train 10,000 Frontier Deployed Engineers by the end of 2027. That figure is a target, not a count of graduates.\n\nThe company says initial cohorts are running in San Francisco, New York and London. Participating organizations nominate engineers with a specific Claude project to lead.\n\nHow the credential works\n\nA multi-day, in-person course ends with a practical assessment. Passing earns the Claude Resident Engineer badge and entry to a 12-week residency at the engineer’s organization. A second assessment is required for the Claude Frontier Deployed Engineer badge. Organizations can ask their Anthropic account team or Partner Account Manager about eligibility.\n\nWhat to do with this\n\nEditorial suggestion: Use the staged assessment model when planning your own learning. Pick a concrete deployment problem, identify prerequisites, and define evidence that the solution works. Keep a review step after real-world use. A vendor credential alone should not replace checking someone’s relevant work and judgment.\n\nSources\n\nAnthropic: Claude Frontier Academy announcement: https://www.anthropic.com/news/claude-frontier-academy\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-10-02",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/doordash-text-to-order-agent/",
      "url": "https://usewright.app/news/doordash-text-to-order-agent/",
      "title": "DoorDash opens a U.S. iOS beta waitlist for text ordering",
      "summary": "DoorDash says its text-ordering assistant can suggest carts and handle checkout after confirmation, with access beginning through a limited beta.",
      "content_text": "DoorDash says its text-ordering assistant can suggest carts and handle checkout after confirmation, with access beginning through a limited beta.\n\nWhat happened\n\nOn September 30, 2026, DoorDash announced an ordering assistant in Apple Messages. Its official announcement says people can request a dish, ask for local suggestions or reorder familiar meals. The assistant searches menus, proposes a cart and can send restaurant food photos.\n\nDoorDash also describes support for group quantities and dietary preferences. The customer confirms the order within the text conversation before checkout. These are announced capabilities, rather than results from a Usewright ordering test.\n\nWho can use it\n\nThe launch is a beta waitlist for U.S. consumers on iOS. DoorDash links to its official text-ordering signup page; joining a waitlist does not establish immediate access. The signup page also describes recurring automated messages, so review that consent before submitting a phone number.\n\nWhat to do with this\n\nEditorial suggestion: If invited, start with a simple order. Check every item, quantity, substitution, total and delivery address before confirming. Treat allergy requirements as something to verify directly with the restaurant; an assistant remembering preferences is not an allergy-safety guarantee.\n\nSources\n\nDoorDash: Text-ordering announcement: https://about.doordash.com/en-us/news/text-doordash\n\nDoorDash: Dash Forward 2026 and beta eligibility: https://about.doordash.com/en-us/news/dash-forward-2026\n\nDoorDash: Official text-ordering waitlist: https://www.doordash.com/text\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-09-30",
        "source_checked_at": "2026-10-14"
      }
    },
    {
      "id": "https://usewright.app/news/openai-dots-always-on-agents/",
      "url": "https://usewright.app/news/openai-dots-always-on-agents/",
      "title": "OpenAI begins rolling out dots, agents with a cloud computer",
      "summary": "OpenAI announced persistent agents for eligible paid plans, with read-only background research, permission controls and a separate deeper-work allowance.",
      "content_text": "OpenAI announced persistent agents for eligible paid plans, with read-only background research, permission controls and a separate deeper-work allowance.\n\nWhat happened\n\nOn September 29, 2026, OpenAI introduced dots. Its announcement describes GPT-6 Astra-powered agents with their own cloud computer and browser, connections to more than 4,000 apps, and conversations across ChatGPT, Slack and Teams.\n\nOpenAI says background proactive research is read-only. Other actions are checked against permissions and safety requirements; some sensitive actions remain for the user. Your computer stays separate unless you connect it.\n\nAccess and limits\n\nThe launch is rolling out to Pro and Business Premium in eligible markets. Enterprise, including Edu and Healthcare, has an administrator-enabled beta. Initial setup uses the desktop app or a desktop browser; mobile messaging follows.\n\nThe first dot is included in eligible Pro and Business Premium plans. Deeper work has an allowance, temporarily extended for the first month after launch. Conversation usage and tasks started in Work or Codex are counted differently; those tasks retain their usual limits.\n\nWhat to do with this\n\nEditorial suggestion: Start with one low-stakes research task. Specify the output, source requirements and approval boundaries. Review the result before expanding access, and use a delegation brief to make the same expectations easy to reuse.\n\nSources\n\nOpenAI: Introducing dots: https://openai.com/index/introducing-dots/\n\nSource-based summary; product performance and account access have not been independently tested.",
      "_usewright": {
        "event_date": "2026-09-29",
        "source_checked_at": "2026-10-14"
      }
    }
  ]
}