Executive Summary
I now read atlas browser features explained as a past-tense query, because OpenAI retired ChatGPT Atlas on 9 August 2026 after less than a year, yet the browser left behind a surprisingly durable blueprint for AI-assisted browsing. Its most important features were not a single magic button. They were the combination of a page-aware ChatGPT sidebar, browser memories, Agent Mode, smarter search and tab controls, profile and extension support, and a deliberately separated native interface over Chromium.
That distinction matters for anyone landing on an older Atlas review. The product OpenAI launched on macOS in October 2025 is no longer a browser you should install, buy, or standardise on. OpenAI moved the direction of travel into ChatGPT Work, its desktop browser, Computer Use, the Chrome sidebar, plugins, and longer-running agent workflows. A current guide therefore has to do two jobs at once: explain what Atlas actually did and distinguish those historical capabilities from the successor features available today.
I traced the product through OpenAI’s launch material, versioned release notes, OWL engineering note, agent-security documentation, shutdown guidance, and the July 2026 ChatGPT Work launch. Because Atlas had already reached its shutdown date when this article was prepared, I did not pretend to run fresh post-shutdown hands-on tests. Where a behaviour cannot be reproduced now, I label it as historical documentation rather than presenting memory or inference as live evidence.
What ChatGPT Atlas Was, and Why It Still Matters
ChatGPT Atlas was OpenAI’s attempt to make ChatGPT the organising layer of a full web browser rather than a destination users visited in a tab. At launch, the browser combined familiar Chromium-era expectations with an assistant that could see the page a user was viewing, answer questions about it, rewrite text in context, retain optional browsing-related memories, and, on eligible paid plans, take actions through Agent Mode. That package made Atlas more ambitious than a normal browser extension because the assistant, browsing surface, history, profiles, and agent controls were designed as one product.
The timeline is essential. Atlas launched for macOS on 21 October 2025 and OpenAI later added tab groups, vertical tabs, multiple profiles, extension imports, richer developer tools, cross-window tab search, proactive toolbar suggestions, AI-assisted tab organisation, and more persistent Agent Mode behaviour. In July 2026, OpenAI announced that the standalone browser would be sunset and that its browser-based agentic lessons would move into ChatGPT. Our earlier coverage of the Atlas shutdown and Work transition is therefore the right contextual companion to a feature explainer because it shows why the feature list can no longer be separated from the migration story.
The product still matters because many design choices survived the product name. ChatGPT Work now includes a built-in browser on desktop, can operate across websites and apps, supports Computer Use, can connect to workplace systems through plugins, and can run Scheduled Tasks. OpenAI explicitly says its updated Chrome sidebar and new browser capabilities build on what it learned from Atlas users. In other words, Atlas became less important as a destination but more important as a prototype for an agentic interaction model that escaped the browser shell.
James Sun, an OpenAI product staff member associated with Atlas, described users as people who had ‘took a leap of faith on a new browser’ when he discussed the sunset on X in July 2026. The line captures the product’s place in the market: Atlas was an experiment in changing browser behaviour, not merely a cosmetic ChatGPT integration.
Atlas Browser Features Explained: The Core Stack
The easiest way to understand Atlas is to follow a normal research-to-action workflow. A user could open a new tab, search the web, switch between result types, ask the ChatGPT sidebar to interpret the current page, use inline writing help in an editable field, carry selected context forward with browser memories, and then hand a bounded sequence of actions to Agent Mode. Later releases reduced friction around the same loop by adding smarter tab discovery, tab groups, multi-profile support, extension imports, and proactive suggestions.
Atlas Browser Features Explained in One Workflow
This is why a simple list of ‘AI browser features’ understates the design. The sidebar provided in-page reasoning. Browser memories supplied optional continuity. Agent Mode added execution. Tab and profile features managed the expanding state created by those interactions. The native shell kept the experience feeling like a Mac application even though web content still depended on Chromium. If you compare the design with our Atlas versus Perplexity Comet guide, the useful distinction is that Atlas placed more emphasis on delegated action while Comet has traditionally kept research and source context closer to the foreground.
The stack also contained ordinary browser capabilities that mattered precisely because Atlas was trying to replace a default browser. OpenAI added support for importing passwords, bookmarks, extensions, history, and site data from other browsers, plus downloads, developer tools, profiles, passkeys, and flexible tab layouts. A browser agent is not very useful if every specialised site, identity workflow, and extension forces the user back into Chrome or Safari.
The hidden limitation was lifecycle risk. Browser adoption carries switching costs that are different from trying a new chatbot. Users import identity state, bookmarks, extensions, and habits. OpenAI’s shutdown guidance warned that bookmarks, tabs, and history might not automatically transfer elsewhere and explicitly reminded users that cookies could contain sensitive information. That is a product-management lesson as much as a technical one: an AI browser can be experimental at the model layer, but it is expected to be conservative at the continuity layer.
| Feature | What Atlas Did | Important Limit or Status |
| New Tab and Search | Combined browsing with ChatGPT-oriented search flows and result-type switching. | Historical. Atlas stopped working in August 2026. |
| ChatGPT Sidebar | Summarised, compared and discussed the page in view without requiring copy-paste. | Successor concept now appears in ChatGPT browser workflows and Chrome sidebar. |
| Inline Writing | Helped rewrite or generate text inside editable fields. | Dependent on page context and user review. |
| Browser Memories | Could remember selected browsing context when enabled to improve follow-up assistance. | Optional, privacy-sensitive, and separate from ordinary ChatGPT conversation history. |
| Agent Mode | Navigated, clicked and typed through multi-step web tasks for eligible plans. | Could not execute code, download files, install extensions or access saved passwords. |
| Tab Intelligence | Added cross-window tab search, AI tab organisation and proactive suggestions. | Useful state management, but increased reliance on browser context. |
| Profiles and Extensions | Supported multiple profiles and migration of common Chromium extensions. | Compatibility was broader than a closed browser, but platform remained macOS. |
| Developer Tools | Added device emulation and conventional Chromium-derived inspection capabilities. | Primarily useful to technical users; not an agent API. |
Search, Sidebar, and Inline Writing
Atlas made the page itself a conversational object. Instead of opening a site, copying a paragraph, moving to ChatGPT, pasting it, asking a question, and then returning to the site, the sidebar kept the reasoning layer adjacent to the material being examined. OpenAI positioned it for tasks such as summarising, comparing, and analysing pages. That sounds modest beside Agent Mode, but it was arguably the feature most likely to become habitual because it removed repeated context transfer without asking the user to surrender control of the page.
The new-tab experience also treated search as a choice between navigation and answer generation rather than one fixed results page. This reflected a broader change in web behaviour in which a user may want a ranked set of links, a synthesised answer, or a direct action depending on the query. Our explainer on how modern answer engines work provides useful background for this shift: the interface competition is no longer only about which search index returns a page fastest, but about when a system should retrieve, synthesise, cite, or act.
Inline writing completed the low-friction loop. In editable areas, users could ask ChatGPT to rewrite or generate text without moving to a separate workspace. That made Atlas useful for email drafts, form responses, CMS fields, comments, and short business writing. The feature was productivity-oriented rather than technically exotic, but it demonstrated a key browser-agent principle: the assistant becomes more useful when it can operate at the point where information is consumed and where output must be entered.
There was also a governance consequence. A page-aware assistant can see context that a conventional chat does not receive unless the user deliberately pastes it. That reduces friction but enlarges the amount of information that may influence a model response. For sensitive work, the correct question is not only ‘can the sidebar understand this page?’ but also ‘should this page be in model context at all?’ Atlas offered privacy controls, but organisations still needed rules for confidential portals, regulated data, and pages containing untrusted instructions.
Browser Memories and Context-Aware Personalisation
Browser memories were one of Atlas’s most distinctive ideas because they tried to convert browsing history from a passive archive into optional, queryable context. When enabled, the browser could retain useful information from sites and use it to improve later assistance. The value proposition was continuity: a user returning to a project could ask questions that depended on earlier browsing without manually reconstructing every tab and detail.
That was more consequential than remembering a favourite site. By early 2026, Atlas could also use browser memories alongside user instructions to help organise tabs. Cross-window tab search and AI-assisted grouping suggested that OpenAI was treating browser state as a semantic workspace. Our coverage of the AI search war in 2026 is relevant here because the strategic contest is increasingly about which system owns durable context across search, browser, assistant, and operating environment, not just which model writes the best paragraph.
This is the first information-gain finding that older launch coverage often misses: Atlas was evolving from ‘ChatGPT beside a webpage’ into a state-management product. A browser with twenty tabs has a retrieval problem before it has an automation problem. If an assistant can understand what those tabs are for, remember prior browsing, group related work, and surface the right tab later, it reduces a cognitive cost that conventional browsers mostly leave to the user.
The trade-off is obvious. Context that is useful to an agent can also be sensitive. OpenAI separated browser memories from normal ChatGPT conversations and offered controls, but the underlying tension remains. The more a system remembers about where someone browsed and what they were doing, the more carefully consent, retention, account separation, and enterprise policy have to be designed. Multi-profile support helped by separating browser identities, but it did not turn memory into a risk-free feature.
Agent Mode: What It Could Do and Where It Stopped
Agent Mode was the feature that made Atlas meaningfully different from a browser with a chatbot sidebar. On Plus, Pro, and Business during its preview period, the agent could work through end-to-end browser tasks by navigating sites, clicking controls, and typing into pages. That enabled shopping research, booking research, form-heavy workflows, repeated website checks, and other jobs where the cost lies in moving through interfaces rather than generating prose.
OpenAI also set deliberate boundaries. The Atlas agent could not run code, download files, install extensions, access other applications or the local file system, read or write ChatGPT memories, or use saved passwords and autofill data. Agent-controlled pages were not added to normal browsing history. Those restrictions are important because they reveal a two-layer design: Atlas the browser could support downloads, extensions, passwords, profiles, and developer tools, while Atlas the agent was intentionally prevented from using several of those capabilities directly.
That separation is the second information-gain finding. A casual feature list can make it appear that ‘Atlas has extensions’ and ‘Atlas has an agent’, therefore the agent can use extensions. It could not. The same applies to saved passwords and file operations. The browser surface was broad; the autonomous permission envelope was narrower. Anyone designing agent workflows should copy that distinction: capability available to a human should not automatically become capability delegated to a model.
For a live alternative, our guide on how to browse the web with Perplexity Comet shows the competing research-first approach. The lesson is not that one agent is universally safer or stronger. It is that agentic browsing works best when permissions, task sensitivity, and human approval are explicit. Low-risk, reversible work is a better starting point than purchases, financial actions, account changes, or irreversible submissions.
| Capability Area | Browser Available to Human | Agent Mode Permission |
| Navigate and click pages | Yes | Yes, within task flow |
| Type into website fields | Yes | Yes, subject to safeguards |
| Downloads | Yes | No direct file downloads |
| Extensions | Supported and importable | Could not install extensions |
| Saved passwords / autofill | Browser could store or import them | Agent could not access them |
| Local file system | Normal browser interactions only | No direct access to other apps or file system |
| Code execution | Developer tools existed for users | Agent could not run code |
| ChatGPT memories | Separate ChatGPT feature | Agent could not read or write them |
| Browsing history | Normal browser history existed | Agent-controlled pages were excluded from normal history |
Tabs, Profiles, Extensions, and Developer Tools
By March 2026, Atlas had accumulated the unglamorous features a browser needs if it expects to become someone’s default. OpenAI added tab groups, vertical tabs, multi-select tab actions, cross-window tab search, tab renaming, multiple ChatGPT logins mapped to separate browser profiles, extension importing, site-data migration, downloads improvements, passkey support, and developer tools enhancements including device emulation. These details mattered because switching browsers is usually blocked by workflow compatibility long before it is blocked by AI quality.
The extension story also placed Atlas firmly in the Chromium ecosystem. Users could bring across familiar extensions rather than rebuild every specialised workflow from scratch. That made Atlas more practical for password managers, productivity add-ons, and niche web tools, even though Agent Mode itself was not allowed to install or freely operate through those extensions. If you need the broader market context, our Comet versus Chrome analysis shows why extension depth, cross-platform stability, and enterprise policy remain powerful advantages for mature browsers even when an AI-first browser offers a better research or automation experience.
Profiles were especially important for professional users. Multiple ChatGPT logins with separate browser profiles reduced the chance that personal and work browsing state would collapse into one memory or session context. This did not solve every data-separation question, but it reflected an understanding that an AI browser sees more semantic context than a conventional address bar. Account isolation therefore becomes part of model governance, not just a convenience for people with two Gmail accounts.
The developer-tool additions are also a useful reminder that Atlas was not simply an agent demo. Device emulation and familiar Chromium inspection tools signalled an attempt to serve developers and technical testers. There was no public Atlas API that turned browser internals into a general automation platform, however. For integrations, the more durable path is now the ChatGPT Work plugin model, Codex tooling, and standard web or application APIs rather than building against a discontinued browser surface.
OWL Architecture: Native Mac UI Over a Chromium Host
Atlas’s most technically interesting feature was invisible to most users. OpenAI did not simply fork Chromium and paint ChatGPT into the toolbar. Its engineering team created an architecture called OWL that separated a native client from a Chromium host. The OWL Client, which was the Atlas application users interacted with, used native macOS technologies including SwiftUI, AppKit, and Metal. The OWL Host ran Chromium and exposed browser capabilities to the client.
The two sides communicated through Mojo IPC, Chromium’s inter-process communication system, with custom bindings that let OpenAI work across Swift and TypeScript. OpenAI described API surfaces such as Session, Profile, WebView, WebContentRenderer, and LayerHost or LayerClient, plus services for bookmarks, downloads, extensions, and autofill. This architecture allowed the team to preserve Chromium compatibility while giving the visible product a native interface and keeping ChatGPT-related components away from direct Chromium internals.
That design connects to OpenAI’s later consolidation. Our reporting on the OpenAI superapp strategy describes the organisational move to bring ChatGPT, Codex, and Atlas closer together. It would be speculative to claim that OWL’s maintenance cost caused Atlas to shut down because OpenAI has not publicly said that. It is reasonable, however, to infer that a separate browser with a native shell, Chromium host, custom IPC bindings, release engineering, extension compatibility, security patches, and platform expansion is a materially larger product commitment than shipping browser capabilities inside an existing ChatGPT application.
This is the third information-gain finding: Atlas’s differentiation was architectural as well as model-driven. The product tried to make browser UI itself an AI-native client while still outsourcing web compatibility to Chromium. That approach gave OpenAI control over the interaction layer, but it also meant Atlas carried the obligations of a browser vendor. The successor strategy keeps more of the agentic experience while reducing the need for users to adopt a wholly separate browser identity.
| Technical Layer | Atlas Implementation | Why It Mattered |
| Web Engine | Chromium in the OWL Host | Preserved compatibility with modern websites and extensions. |
| Visible Application | OWL Client | Allowed OpenAI to design a native AI-first browser shell. |
| macOS UI | SwiftUI, AppKit and Metal | Made the interface native rather than a web-wrapped Chromium chrome. |
| IPC | Mojo with custom Swift and TypeScript bindings | Connected native UI and Chromium processes while keeping boundaries explicit. |
| Core APIs | Session, Profile, WebView, WebContentRenderer, LayerHost / LayerClient | Separated navigation, rendering, profile and composition responsibilities. |
| Browser Services | Bookmarks, downloads, extensions and autofill services | Supported migration and default-browser expectations. |
| Platform | macOS during public life | Windows, iOS and Android did not ship before the product sunset. |
Pricing, Plan Access, and Hidden Usage Limits
Atlas did not have a standalone subscription price. At launch, OpenAI made the macOS browser available across consumer plans including Free, Go, Plus, and Pro, with Business in beta and Enterprise or Edu access controlled by administrators. The important paid distinction was Agent Mode, which OpenAI initially made available in preview for Plus, Pro, and Business users. That meant the browser shell and the autonomous browser agent were not the same commercial product even when they appeared in one application.
By August 2026, the more useful pricing question is what it costs to access the successor workflows. ChatGPT Work is available through the ChatGPT ecosystem rather than as an Atlas licence. OpenAI says the desktop application exposes Chat, Work, and Codex on every plan, including Free, while web and mobile Work rolled out first to Pro, Enterprise, and Edu before expanding to Plus and Business. Usage is not a simple fixed number of browser actions: OpenAI says longer or more complex Work tasks can consume more of a plan’s included usage and that Work follows the same usage structure as Codex.
This is where older comparison articles can mislead. A monthly subscription price is only the entry cost. Effective capacity depends on the plan, the complexity of the task, workspace-level limits, individual overrides, and additional credits in managed environments. OpenAI also introduced Business Premium seats in August 2026 at a higher price than Standard seats, with five times more usage and different reset behaviour. Our coverage of the broader ChatGPT and Codex superapp shift helps explain why pricing is moving from ‘browser tier’ to ‘agent capacity across a unified work surface’.
For procurement, the safe approach is to separate confirmed list prices from variable usage. Consumer prices can change by market and taxes, Enterprise remains sales-led, and usage allowances can be updated without changing the headline subscription. The table below therefore labels plan prices as current published entry points rather than pretending they are a permanent all-in cost for browser automation. The figures were checked against OpenAI’s current ChatGPT pricing page, Pro tier guidance, Business billing guidance, Business Premium announcement, and ChatGPT Go launch.
| Plan | Published Entry Price | Atlas Historical Access | 2026 Work / Browser Context |
| Free | $0 | Atlas browser available at launch; Agent Mode not part of the paid-agent preview. | Desktop Chat, Work and Codex available, with limited included usage. |
| Go | US launch price $8/month; local pricing may vary | Atlas browser available at launch. | Lower-cost ChatGPT plan; verify local Work limits in account before relying on automation. |
| Plus | $20/month | Atlas browser plus Agent Mode preview. | Expanded ChatGPT Work access; complex tasks consume plan usage variably. |
| Pro | $100 or $200 usage tiers were published in 2026 | Atlas browser plus Agent Mode preview. | Higher Work capacity; exact effective throughput depends on task complexity. |
| Business Standard | $25/user monthly or $20/user/month annual | Business access was introduced in beta. | Workplace controls and shared governance; minimum-seat terms may apply. |
| Business Premium | $125/user monthly or $100/user/month annual | Not an Atlas-era tier. | Five times more usage than Standard and different reset behaviour, per OpenAI August 2026 announcement. |
| Enterprise / Edu | Custom or contracted pricing | Admin-enabled Atlas access. | Central governance, Compliance API, browser/network controls and spend management. |
Security and Privacy: The Powerful Boundary Problem
Atlas concentrated several sensitive browser roles in one place: authenticated sessions, page content, browsing history, optional browser memories, a model that interprets language, and an agent that can take actions. That combination made prompt injection a first-class security problem. OpenAI’s security team described prompt injection as a significant long-term challenge for browser agents because malicious instructions can be hidden inside the very pages, emails, or content the agent is asked to process.
OpenAI responded with automated red teaming and defence-in-depth rather than claiming deterministic safety. Its published example included a malicious instruction embedded in content that could steer an agent towards an unwanted high-impact action until mitigations caught the attack. The broader research literature points in the same direction. BrowseSafe examined indirect prompt injection in realistic browsing settings, while 2026 work such as StepJack shows that multi-step attack chains can materially increase attack success for some computer-use agents. A user should therefore treat ‘the agent can read it’ and ‘the agent should obey it’ as separate questions.
Security boundaries inside Atlas were designed to reduce blast radius. Agent Mode could not use saved passwords, run code, download files, install extensions, or reach the local file system. Users could also run the agent logged out of sites when authentication was not required. These controls did not eliminate indirect prompt injection, but they reduced the number of privileged capabilities a compromised browser task could immediately exploit.
A useful industry counterpoint comes from Aravind Srinivas, Perplexity co-founder and CEO, who argued in 2026 that ‘A traditional operating system takes instructions. An AI operating system takes objectives.’ The ambition is clear, but objectives create a larger interpretation surface than clicks. Ajit Varma, Head of Firefox, has similarly emphasised that users have different expectations of their browsers. The practical conclusion is plural rather than ideological: use agents where their autonomy creates measurable value, and retain conventional browser controls for sensitive transactions, identity changes, and high-stakes decisions.
What Replaced Atlas: ChatGPT Work, Chrome Sidebar, and Codex
OpenAI did not replace Atlas with another standalone browser. It moved browser-based work into a broader agent surface. ChatGPT Work can operate across web, mobile, and desktop, connect to applications through plugins, use a built-in browser on desktop, carry out Computer Use actions across local apps and the browser, and run Scheduled Tasks that revisit websites or connected systems. The Chrome extension is also being updated so ChatGPT can live in Chrome’s sidebar without forcing users to switch their default browser.
This successor model preserves the strongest Atlas idea, context plus action, while broadening the scope beyond websites. A browser agent could compare vendors and fill a form. A Work agent can combine that web research with Slack, Teams, Google Drive, SharePoint, email, calendars, CRMs, project trackers, local files, and document creation. That is not a one-to-one feature migration. It is a shift from ‘AI browser’ to ‘agent workspace that includes a browser’.
OpenAI’s own early customer examples show why the scope widened. Angela Ferrante, Head of Enterprise Marketing at Zapier, used Work in a system that surfaced ‘seven figures in potential sales’. Vaneet Seth, R&D Efficiency Manager at RingCentral, described a workflow that scaled support to ‘roughly 50 product managers’. Nathan Bolt, Head of Digital Products at Virgin Atlantic, reduced competitive analysis from ‘weeks of analysis to hours’. Will Daney, Go-to-Market Manager at NVIDIA, replaced preparation work consuming ‘about 40% of his pre-event time’. These examples are vendor-published case studies, not independent benchmarks, but they demonstrate the kind of multi-system outcome Atlas could not deliver by staying browser-only.
Greg Brockman, OpenAI co-founder and president, wrote that the Work cloud browser was ‘really cool to use’ because a user could monitor the agent and intervene in the live application. The important word is intervene. The product direction is not full invisibility. It is delegated execution with observability and takeover. That model is likely to matter more than whether the assistant lives in a dedicated browser window or inside a general-purpose work agent.
Who Should Use What in 2026?
No one should adopt ChatGPT Atlas as a production browser in August 2026 because the standalone product has been retired. The useful decision is which current surface best matches the workflow Atlas helped define. ChatGPT Work is the natural successor for users who want multi-step tasks that move between websites, files, apps, and finished deliverables. Chrome with the ChatGPT sidebar is the lower-switching-cost option for users who want page-aware assistance without abandoning a mature default browser. Perplexity Comet remains relevant for research-heavy browsing where source visibility and reading context are central.
The best fit also depends on risk. A marketing analyst collecting public competitor information can safely delegate more than an administrator changing financial permissions. A researcher working through unfamiliar sources may prefer a source-adjacent assistant before turning findings into an automated workflow. A regulated enterprise may value browser and network controls, compliance logs, approval gates, and identity separation more than raw agent speed. The category is therefore moving away from one ‘best AI browser’ and towards policy-defined combinations of research, browsing, and computer-use agents.
Kevin Jiang, a former Perplexity Comet team member who later started an AI browser company, told TechCrunch in July 2026 that the first wave of AI browsers ‘wasn’t a strong long-term proposition’ when the pitch was essentially changing the default search or adding chat around tabs. That judgement fits Atlas’s outcome. Browser-level AI is valuable, but a new browser must justify migration, compatibility, identity, management, and security costs that a sidebar or broader desktop agent can sometimes avoid.
My practical recommendation is to evaluate the workflow rather than the logo. Measure task completion, source traceability, human intervention points, error recovery, prompt-injection exposure, and the cost of maintaining account and browser state. Atlas is now most useful as a design case study. Its surviving ideas belong in a procurement checklist for current systems, not in a download guide for a product that has already reached end of life.
Our Editorial Verification Process
This explainer was verified on 11 August 2026 against OpenAI’s official Atlas launch page, Atlas release notes, Agent Mode documentation, OWL engineering article, security post, Atlas shutdown guidance, current ChatGPT Work launch material, and current OpenAI pricing announcements. I used the version history to distinguish launch-day capabilities from features that arrived in early 2026 rather than flattening the product into one timeless feature list.
Live post-shutdown Atlas reproduction was not possible because OpenAI scheduled the standalone browser to stop working on 9 August 2026. I therefore did not label any post-shutdown behaviour as hands-on testing. Historical behaviours are presented as documented product capabilities. Successor behaviours are separately verified against ChatGPT Work documentation. Security claims were cross-checked with OpenAI’s own prompt-injection work and independent agent-security research including BrowseSafe and StepJack.
For internal linking, the requested XML sitemap endpoints did not return parseable sitemap content through the available browsing layer. Rather than fabricate sitemap entries, I used the prompt’s fallback rule and selected eight live, indexed Perplexity AI Magazine pages that are directly relevant to Atlas, Comet, AI search, answer engines, and OpenAI’s superapp transition. Each internal URL appears once, with contextual anchor text, in a different eligible body section.
This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.
Conclusion
Atlas lasted for less than a year, but treating it as a failed browser would miss the more useful lesson. OpenAI proved that a browser could combine page-aware reasoning, optional semantic memory, delegated web actions, AI-assisted tab state, ordinary Chromium compatibility, and a native interface without reducing the product to a sidebar extension. It also discovered that users expect a browser to be durable infrastructure, not merely an experiment with a clever agent.
The surviving direction is broader than Atlas. ChatGPT Work now carries browser-based agentic work into files, apps, plugins, scheduled tasks, local computer use, and finished business outputs, while the Chrome sidebar offers a lower-friction path for users who want assistance inside an established browser. The unresolved questions are the hard ones: how much context an agent should retain, how browser agents should resist untrusted instructions, how organisations should meter and audit long-running work, and when a human must approve or take over an action.
For readers searching for Atlas browser features in 2026, the most accurate answer is therefore historical but not obsolete. Atlas is gone. Its feature model is becoming the operating grammar of the next generation of browser and desktop agents.
Frequently Asked Questions
What Was ChatGPT Atlas?
ChatGPT Atlas was OpenAI’s macOS web browser with ChatGPT built into the browsing experience. It combined a page-aware sidebar, inline writing help, optional browser memories, Agent Mode, Chromium compatibility, profiles, extensions, tab-management features, and a native macOS interface. OpenAI launched it in October 2025 and retired the standalone product in August 2026.
Is ChatGPT Atlas Still Available in 2026?
No. OpenAI said Atlas was scheduled to stop working on 9 August 2026. Its browser-based agentic direction has moved into ChatGPT Work, the ChatGPT desktop app, Computer Use, and an updated Chrome sidebar. Old Atlas setup guides should therefore be treated as historical documentation rather than current installation instructions.
What Were the Main Atlas Browser Features?
The main features were the ChatGPT sidebar, browser memories, Agent Mode, inline writing, AI-assisted search, tab groups, vertical tabs, cross-window tab search, profiles, extension and site-data import, downloads, passkeys, and developer tools. Later releases also added AI tab organisation, proactive suggestions, and more persistent agent behaviour.
Did ChatGPT Atlas Use Chromium?
Yes. OpenAI built Atlas around a Chromium host, but the visible application used its own OWL architecture. A native macOS client built with technologies including SwiftUI and AppKit communicated with the Chromium host through Mojo IPC. This let OpenAI keep web compatibility while controlling a distinct AI-first interface layer.
Could Atlas Agent Mode Use Passwords or Download Files?
No. OpenAI documented hard boundaries for Agent Mode. It could navigate, click, and type on websites, but it could not run code, download files, install extensions, access saved passwords or autofill, reach other applications or the local file system, or read and write ChatGPT memories.
Was ChatGPT Atlas Free?
The browser itself was available across several ChatGPT plans at launch, including Free. Agent Mode was a separate paid-plan capability during preview, initially available to Plus, Pro, and Business users. There was no standalone Atlas subscription. Current successor access depends on ChatGPT plan and Work usage rather than an Atlas licence.
What Replaced ChatGPT Atlas?
ChatGPT Work is the main successor direction. On desktop it includes a built-in browser and can use websites, local apps, files, plugins, Computer Use, and Scheduled Tasks. OpenAI is also updating its Chrome extension so ChatGPT can work from Chrome’s sidebar, reducing the need for a dedicated OpenAI browser.
How Did Atlas Differ From Perplexity Comet?
Atlas leaned more heavily towards delegated browser actions and tight integration with ChatGPT, while Perplexity Comet has been positioned more naturally around research, source context, and AI-assisted browsing. The comparison changed after Atlas was retired, so current buyers should compare Comet with ChatGPT Work or Chrome plus ChatGPT rather than treating Atlas as an active product choice.
References
OpenAI. (2025, October 21). Introducing ChatGPT Atlas.
OpenAI Help Center. (2026). ChatGPT Atlas release notes.
OpenAI Help Center. (2026). ChatGPT Atlas Agent Mode.
OpenAI. (2025, December 22). Continuously hardening ChatGPT Atlas against prompt injection attacks.
OpenAI Help Center. (2026, July). Evolving Atlas into ChatGPT for browser-based agentic work.
OpenAI. (2026, July 9). ChatGPT is now a partner for your most ambitious work.
