Top 10 Claude Cowork Alternatives, Tested (August 2026)

---

title: "Top 10 Claude Cowork Alternatives, Tested (August 2026)" slug: "top-10-claude-cowork-ai-agent-alternatives-2026-guide" excerpt: "We run these agents on real work daily. The 10 Claude Cowork alternatives that actually held up in August 2026, from open-source desktops to cloud workforces." seoDescription: "We run these agents on real work daily. The 10 Claude Cowork alternatives that actually held up in August 2026, from open-source desktops to cloud workforces." keywords:

• "Claude Cowork alternatives"
• "open source Claude Cowork alternative"
• "free Claude Cowork alternative"
• "best Claude Cowork alternatives 2026"
• "ChatGPT Work vs Claude Cowork"
• "Eigent"
• "AI agent platforms 2026" tags:
• "Claude Cowork"
• "AI Agents"
• "Open Source AI"
• "AI Comparison"
• "2026 Guide" category: "AI Tools" author: "Yuma Heymans"

---

The tested August 2026 comparison of every serious alternative to Claude Cowork: verified pricing, the new 90% benchmark reality, the open-source desktop wave, and a decision framework built from running these agents on real work.

July 2026 was the most violent single month the agent market has ever had. In the space of twenty days, Anthropic took Claude Cowork to web and mobile with background sessions that keep working while your laptop is closed - TechCrunch, OpenAI shipped GPT-5.6 and ChatGPT Work on the same day, xAI released Grok 4.5, Anthropic followed with Claude Opus 5, and a Chinese automation company most Western readers had never heard of became the first to push past 90% on OSWorld, the benchmark for real computer use. Any comparison of Cowork alternatives written before that month, including the July 8 version of this very guide, is describing a market that no longer exists.

This is a full re-baseline, not a date bump. We have rewritten every product profile against what shipped through August 5, 2026, re-verified every price on the official page it comes from, extended the benchmark data to the new frontier, and added the category the market forced into the conversation this summer: open-source Cowork desktops like Eigent that replicate the product for free on your own hardware. Where the previous version of this guide was wrong or is now stale, we say so explicitly, because documenting the churn is half the value of a comparison in a market that repriced itself three times in eighteen months.

One framing note before the ranking. We operate agent platforms daily at O-mega, which means two things: we have first-hand operating experience with this category that pure aggregator listicles do not, and we have a product in the ranking. We handle that the honest way: O-mega is scored on the same criteria as everyone else, its trade-offs are stated as plainly as its strengths, and every claim about competitors links to a primary source you can check.

Contents

1. What Changed in July 2026: The Twenty Days That Reset the Market
2. Claude Cowork in August 2026: The New Baseline
3. ChatGPT Work and ChatGPT Agent (OpenAI)
4. Microsoft Copilot Cowork and the E7 Frontier Suite
5. Google: Gemini Spark, Workspace Studio, and Auto Browse
6. O-mega
7. OpenClaw
8. Eigent and the Open-Source Cowork Desktop Wave
9. Perplexity Comet
10. Manus
11. Amazon Nova Act and Alexa+
12. xAI Grok 4.5 and Grok Build
13. The Agentic Browser Shakeout
14. Benchmark Reality: The Road to 90%
15. Enterprise Consolidation: watsonx, Artemis, and ServiceNow
16. Verified Pricing in August 2026
17. Safety, Governance, and Concentration Risk
18. Decision Framework: Choosing in August 2026

The August 2026 Ranking at a Glance

Before the detailed profiles, here is the master comparison. Each alternative is scored 0-10 on four weighted criteria, and the table is sorted by the weighted final score, highest first. Capability (30%) measures how much real autonomous work the agent completes today, anchored to public benchmarks and shipped features. Value (25%) measures what you actually pay for that capability at entry level. Reach (20%) measures platform and channel availability: operating systems, browsers, mobile, messaging. Governance (25%) measures permission controls, isolation, auditability, and enterprise trust.

How to read the weights: Capability gets 30% because an agent that cannot finish tasks is worthless at any price. Value gets 25% because the spread between entry points is enormous ($0 for Comet, OpenClaw, and Eigent versus $300/month for full Grok 4.5 access). Reach gets 20% because agents only help where you actually work. Governance gets 25% because agent permissions, isolation, and audit trails are now procurement blockers rather than nice-to-haves. Enterprise control planes like IBM watsonx Orchestrate, Kore.ai Artemis, and ServiceNow (Moveworks) are covered in section 15 rather than ranked here, because they are governance layers you deploy on top of agents, not direct Cowork substitutes for an individual's daily work.

1. What Changed in July 2026: The Twenty Days That Reset the Market

Comparison guides in this category rot at the metadata layer first and the fact layer second, and July 2026 rotted the facts faster than any month on record. The structural reason is worth stating before the timeline: the three frontier labs now ship agent surfaces and frontier models as a single coupled release, so one model launch invalidates pricing, capability claims, and product boundaries all at once. When GPT-5.6 shipped, it did not just update a benchmarks table; it launched ChatGPT Work, an entirely new Cowork competitor, in the same announcement. When Opus 5 shipped, it repriced the economics of every Cowork session that runs on it. Model news and agent news are no longer separable stories.

The verified timeline of the month that forced this rewrite:

• July 8: Claude Cowork expands to web and mobile with background sessions, Max subscribers first - TechCrunch
• July 8: xAI releases Grok 4.5, its agent-focused flagship, at $2/$6 per million tokens - Fello AI
• July 9: OpenAI publicly releases GPT-5.6 (Sol, Terra, Luna) alongside ChatGPT Work - Wikipedia
• July 24: Anthropic releases Claude Opus 5 at $5/$25 per million tokens, the new default on Max - Anthropic
• July 27: Intelligence Indeed's Z-Agent hits 90.2% on OSWorld, the first agent past 90% - GlobeNewswire

Two more July items complete the picture in prose. Google shipped Gemini 3.6 Flash on July 21 as its new everyday-speed model - Google release notes, and OpenAI scheduled the shutdown of ChatGPT Atlas for August 9, 2026, folding its year-old agentic browser into the unified desktop app announced in March - Wikipedia. That second item deserves emphasis because it extends the consolidation pattern this guide documented in July: first the standalone agents died (Operator, August 2025; Mariner, May 2026), and now the standalone agent browsers are starting to die too, absorbed into unified desktop surfaces the same way the agents before them were absorbed into chat apps.

Why this matters for your buying decision: the half-life of an agent comparison is now measured in weeks, so the date on the article you are reading is a load-bearing fact, not a footnote. How to apply it: treat any page that still describes Cowork as desktop-only, ranks Atlas as a live product, or caps its benchmark narrative at the human baseline as pre-July, and assume its pricing is equally stale. Everything below is re-verified against the August 5 reality, and each profile states explicitly what changed since our last revision so you can calibrate how fast that particular corner of the market is moving.

2. Claude Cowork in August 2026: The New Baseline

You cannot evaluate alternatives to something you misunderstand, and the July version of Cowork already made most 2025 descriptions obsolete. The August version goes further: Cowork is now a cross-device agentic assistant, not a desktop app. Starting July 8, 2026, Cowork extended to web and mobile with background sessions, Max subscribers first, so a task started at your desk keeps running and reports to your phone "even if their laptop is closed," as Anthropic's example workflow puts it: schedule Monday's client prep for 6 a.m., and Cowork works through email threads and transcripts, builds the briefing doc, and leaves the follow-up email drafted but unsent - TechCrunch.

Access and pricing held steady while the surface expanded. Cowork is included in Claude Pro at $20/month ($17/month billed annually, $200 up front), Max from $100/month in 5x and 20x flavors, and Team at $20 per standard seat or $100 per premium seat billed annually - Claude pricing. The honest caveat remains: autonomous multi-step work burns usage limits far faster than chat, which is why Anthropic ran a no-charge promotion doubling the 5-hour Cowork limits for paid users as an adoption push - The New Stack. For a full product walkthrough, see our Claude Cowork autonomous desktop guide, and for the plan-by-plan economics our Cowork pricing and ecosystem analysis.

The model layer underneath changed in the most consequential way. On July 24, 2026, Anthropic released Claude Opus 5 at $5/$25 per million tokens (unchanged from Opus 4.8), making it "the new default model on Claude Max, and the strongest model on Claude Pro" - Anthropic. The claims that matter for agent work: on OSWorld 2.0, Anthropic says Opus 5 outperforms every other model at any given cost and surpasses Fable 5's best result at just over a third of the cost, and on CursorBench 3.2 it lands within 0.5% of Fable 5's peak at half the cost per task. Translated out of benchmark language: the model your $20 Cowork subscription runs on now delivers near-flagship agent capability at a fraction of the flagship's cost, which resets the value math for every alternative in this guide. Our Opus 5 vs 4.8 breakdown covers the generational comparison in detail, and the Fable 5 and Mythos 5 benchmarks profile the top of the family.

So why look at alternatives at all? The four honest reasons from our July analysis survive, with one modification. First, usage ceilings: heavy Cowork use still hits plan limits, and scaling to Max costs real money. Second, ecosystem gravity: an agent native to Microsoft 365 or Google Workspace touches your real documents with less friction than a general-purpose agent. Third, unattended scale: Cowork's background sessions narrow this gap but it remains one user's copilot, not a fleet of independent workers with their own accounts and browsers. Fourth, sovereignty, and here the market moved: the concern that a hosted frontier agent can be interrupted (proven by Fable 5's three-week export-control suspension in June) now has a much stronger answer than it did in July, because the open-source Cowork desktop category matured into real products (section 8). Each profile below tells you which of these gaps it actually closes, and which it does not.

3. ChatGPT Work and ChatGPT Agent (OpenAI)

OpenAI's answer to Cowork finally has a name, and it is not Atlas or agent mode: it is ChatGPT Work, released July 9, 2026 alongside GPT-5.6, whose Wikipedia entry states it plainly: GPT-5.6 "is set to be the operating agent of ChatGPT Work, a service that was released at the same time" - Wikipedia. ChatGPT Work is an agent mode that takes a brief and works in the background for minutes to hours, delivering finished artifacts rather than chat replies: spreadsheets with working formulas, presentation decks, formatted documents, and interactive web apps, staying on a single project for hours by working through steps one at a time - Fello AI. This is OpenAI's closest Cowork analog to date, and it anchors this entry the way agent mode alone no longer can.

The rollout details matter because they are aggressive. ChatGPT Work carries no separate price: it meters against existing plan allowances, is available on the desktop app on every plan including Free, and rolled to web and mobile for Pro and Enterprise first, then Plus and Business. Model access is tiered: free accounts run the mid-tier Terra, while paid tiers choose between Sol (flagship), Terra, and Luna (budget) - Fello AI. Note the fine print that heavy users hit first: Work shares its usage pool with Codex, so a long background build consumes the same allowance as your coding sessions. Our GPT-5.6 benchmark and pricing breakdown covers the model family, and our GPT-5.6 vs Opus 5 agent comparison puts the two July flagships head to head on agent workloads.

The architectural contrast with Cowork is sharper now than it was for any previous OpenAI product. Cowork's home advantage is your machine: local files, local folders, your real desktop. ChatGPT Work's home advantage is the finish line: it is optimized to return a completed deliverable from a cloud sandbox, which means it cannot reorganize your laptop but also cannot break it, and its outputs arrive ready to present rather than ready to review in a working folder. In practice the two products fail differently too. Cowork's characteristic failure is running out of usage mid-task; Work's characteristic failure is confident overreach, assembling a polished deliverable from a source it misread, which is why its approval gates and takeover mode remain load-bearing rather than decorative. The lineage runs back through ChatGPT Agent, which absorbed the shuttered Operator in August 2025; our Operator pricing analysis now reads as a historical document of how fast OpenAI repriced this capability downward.

One structural update closes this profile: ChatGPT Atlas shuts down on August 9, 2026, per its Wikipedia entry, following OpenAI's March announcement that Atlas, the ChatGPT desktop app, and Codex would merge into one application - Wikipedia. Atlas never shipped a Windows or mobile version in its ten months of life, and its retirement four days after this guide's revision date is the clearest confirmation of the consolidation thesis in section 1. Best for: anyone who wants the most capable deliverable-producing agent at mainstream prices, values the largest connector ecosystem and fastest iteration cadence in the industry, and does not need local file-system access. For the wider field around it, see our top ChatGPT Work alternatives ranking.

4. Microsoft Copilot Cowork and the E7 Frontier Suite

Copilot Cowork still carries the strangest lineage in this market: Microsoft built it by integrating "the technology behind Claude Cowork into Microsoft 365 Copilot," in its own words, announcing it March 9, 2026 and taking it to general availability on June 16, 2026 - Microsoft. The strategic reading we gave in July still holds and has only strengthened: Anthropic proved the scarce assets are model quality and distribution surface, not the agent harness, and Microsoft owns the distribution. Copilot Cowork grounds tasks in your emails, meetings, messages, files, and data across Word, Excel, PowerPoint, Outlook, Teams, and SharePoint, which means it touches your real work products without any file migration, precisely the friction that keeps desktop agents from spreading beyond early adopters. Our Copilot Cowork complete analysis dissects the architecture and the Anthropic deal in depth.

What changed since July is the packaging, and it changes the enterprise calculus. Microsoft now sells Microsoft 365 E7, the "Frontier Suite," at $99 per user per month: it bundles everything in E5 ($60) plus Microsoft 365 Copilot (a $30/user add-on when bought separately), Agent 365, and the Entra Suite, with "observability, security, and governance for agents" as the headline pitch - Microsoft. Read that bundle from first principles: Microsoft is pricing agent governance as the premium, not agent capability. The capability (Copilot Cowork) bills separately on consumption at $0.01 per Copilot Credit; the $99 E7 seat is what you pay to observe, secure, and govern what those agents do. Competitors now frame Copilot Cowork through E7 in enterprise deals, and so should your procurement analysis.

The consumption billing still deserves arithmetic before you commit a department to it, because it remains the dimension where Copilot Cowork and Claude Cowork genuinely diverge. At a cent per credit, light users cost almost nothing, which suits organizations where most staff run an agent twice a week. But an enthusiastic power user running long agentic sessions daily can quietly out-spend a flat Claude Max subscription, and nothing in a consumption model tells them to stop. The pilot pattern that works: set a spending alert per user, measure the credit burn of your five most common workflows for a month, and only then split your population between consumption billing and flat plans, given that the underlying agent technology is, unusually, the same on both sides of that decision.

Microsoft's other two fronts carry over from July intact. Windows 11's Copilot Actions and Agent Workspace run each agent under a separate non-admin Windows account in an isolated session, the most serious OS-level answer yet to "what exactly can the agent touch, and how do I prove it" - BleepingComputer. And Fara-7B, the MIT-licensed 7B computer-use model that runs on-device on Copilot+ PCs, still scores 73.5% on WebVoyager, ahead of much larger hosted models - Microsoft Research. Best for: Microsoft 365 organizations that want Cowork-class autonomy inside their existing compliance boundary, and security teams that will pay the E7 premium for OS-level isolation and agent observability. The trade-off is unchanged: consumption billing is easy to start and hard to forecast, a sharp contrast with Claude's flat plans.

5. Google: Gemini Spark, Workspace Studio, and Auto Browse

Google's agent story remains distributed across surfaces rather than concentrated in one product, and the July version of this guide underweighted the piece the search results now treat as a first-class Cowork alternative: Google Workspace Studio. Launched December 3, 2025 and rolled to scheduled-release domains by March 2026, Studio is "the place to create, manage, and share AI agents to automate work in Workspace, no coding required": you describe an agent in natural language, wire it to triggers like meeting transcripts landing or Sheet rows changing, and connect third-party apps including Asana, Jira, and Salesforce - Google Workspace Updates. Through 2026 it gained an "Ask a Gem" step, NotebookLM integration, list looping, and granular admin controls, which is the cadence of a product Google intends to keep. Studio is not a desktop agent; it is a no-code agent factory for the suite your company may already run on, included in mainstream Workspace Business and Enterprise editions rather than gated behind Ultra.

The premium piece is still Gemini Spark, the 24/7 personal agent for Google AI Ultra subscribers, now available in "select countries" for users 18 and up rather than the US-only footprint we reported in July - Google. Spark connects to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps (connectors off by default), checks with you before major actions, and works "even if your phone and laptop are turned off." One honest correction to our own July analysis: that always-on quality is no longer a unique differentiator, because Cowork's July 8 web and mobile expansion brought background sessions that survive a closed laptop to Anthropic's side of the fence. What remains genuinely structural is the attachment point: Spark belongs to your Google identity, not your hardware, so its strength is standing work inside Google surfaces rather than general computer use.

The model lineup and pricing moved in July. Gemini 3.6 Flash launched July 21, 2026 as the new everyday model, joining flagship Gemini 3.1 Pro and the video-focused Gemini Omni - Google release notes. Google AI Ultra comes in two tiers, 99.99 and 219.99 euros a month on the European storefront (the US tiers run $99.99 and roughly $200), with Spark and the Antigravity agent platform gated to Ultra - Google subscriptions. The mass-market piece remains Chrome Auto Browse, which gives AI Pro and Ultra subscribers autonomous task execution inside the world's dominant browser with human-approval pauses on sensitive actions like payments - TechCrunch. Our Gemini 3.1 Pro guide profiles the flagship in depth.

How the three pieces compose is the real evaluation question, because no single Google product matches Cowork's shape. Studio automates recurring Workspace workflows without code; Spark handles standing personal tasks attached to your account; Auto Browse executes one-off web tasks in the browser you already use. For a Workspace-native organization, that trio covers a surprising share of what a desktop agent does, at lower friction, and Studio's inclusion in ordinary Workspace editions makes it the cheapest serious agent capability most businesses already own. The honest limits: capability fragments across three products with three gating models, the best of it sits behind Ultra pricing, and nothing in the trio touches local files or non-Google desktop software. Best for: Google Workspace households and companies, and teams that want no-code agent automation on the suite they already pay for.

6. O-mega

Most tools in this guide are one agent attached to one person: your copilot, your browser, your machine. O-mega makes a different architectural bet, and the July market movement (Cowork's background sessions, ChatGPT Work's hours-long runs) is the rest of the market converging toward it: work is done by an AI workforce, a set of persistent agents with their own identities, their own browsers, and their own standing instructions, operating 24/7 in the cloud whether or not you are at your desk. You hire agents into roles, brief them like colleagues, and they execute: research, outreach, content pipelines, data work, monitoring, multi-step web operations.

The mechanics matter more than the framing. Each O-mega agent gets a dedicated cloud browser session with its own credentials, so an agent can log into the tools it needs and operate them the way a human contractor would, without borrowing your laptop or your logged-in identity. Agents run browser sessions, computer sessions, and internal work under one orchestration layer, hand tasks to each other through delegation, and escalate to you at defined human checkpoints rather than after the fact. Every session leaves a reviewable log, which is the governance piece individual desktop agents mostly lack: when an agent acted on your behalf at 3 a.m., you can see exactly what it did and why.

A concrete week makes the model tangible. A two-person e-commerce brand runs four O-mega agents: one monitors competitor pricing and stock across a dozen storefronts every morning and flags anomalies; one works the content pipeline, drafting and scheduling product copy against a standing brief; one handles supplier research, collecting quotes and building comparison sheets from the open web; one watches reviews and support inboxes overnight and escalates anything with legal or refund exposure to a human before replying. None of those four jobs needs the founders present, all four need to happen every day, and together they would otherwise be a part-time hire. That is the workload shape the platform is built for, and it is a different shape than "help me finish this document," which Cowork serves better.

Where does it sit against the August Cowork specifically? Cowork is stronger when the work is your local files and your judgment: refactoring a folder of documents, working alongside you in real time, and now picking that work up from your phone. O-mega is stronger when the work should not require you at all: recurring pipelines, always-on monitoring, parallel workstreams across many sites and accounts, the "unattended scale" gap from section 2 that background sessions narrow but do not close, because Cowork's sessions are still one person's tasks rather than a staffed set of roles. Pricing is a flat subscription rather than per-agent-hour metering, so an agent working through the night does not run a taxi meter the way consumption-billed platforms do. It is a web platform, so there is no local desktop install; the trade-off is the mirror image of Cowork's: O-mega's agents will not reorganize your laptop's file system.

Best for: founders, operators, and small teams that want delegation rather than assistance, and need work to continue while they sleep. It is also the natural graduation path when you notice your Cowork or ChatGPT Work usage is really three or four distinct standing jobs that deserve their own dedicated workers. For a broader map of workforce-style platforms, our top OpenClaw alternatives ranking compares the adjacent options through that lens.

7. OpenClaw

The open-source agent story of 2026 still has one name every developer knows: OpenClaw, the personal agent that its own site describes, quoting Y Combinator, as having gone "from a weekend project to the most-starred software repo on GitHub in under 5 months, with 346k+ stars" - OpenClaw. Creator Peter Steinberger joined OpenAI in February 2026, and the project now lives inside the OpenClaw Foundation, a non-profit with a full-time team, which keeps the code independent of any single vendor's roadmap. The star count and the foundation structure are the two facts that separate OpenClaw from every "open-source agent" that flared and died in 2025: it has both mass adoption and a governance answer.

OpenClaw's architectural bet is the inverse of every hosted product in this guide: the agent runs on your own hardware, with your own model API keys, and meets you in the messaging apps you already use, spanning 29 channels including WhatsApp, Telegram, Discord, Slack, Signal, and iMessage - OpenClaw. There is no subscription, no vendor-imposed usage ceiling, and no data leaving your infrastructure except the model calls you choose to make. You text your agent an instruction from your phone; it executes on the machine at home where it lives. For the sovereignty concern raised by June's export-control episode, this remains the strongest answer available: nobody can suspend a binary you run yourself.

The cost structure is honest but not free in practice. The software costs $0; the model calls behind a busy agent typically run $1-150/month depending on how hard you drive it, and our OpenClaw pricing breakdown works through real usage profiles. The real price is operational: you are the security team. A personal agent with shell access, messaging reach, and stored credentials is a serious attack surface, and hardening it (sandboxing, permission scoping, prompt-injection hygiene) is your job, not a vendor's. A managed middle path emerged for exactly this reason: Agent 37 hosts always-on OpenClaw instances in the cloud from $3.99/month for bring-your-own-keys hobby use up to $99.99/month team plans with a built-in browser - Agent 37, trading away the self-hosting sovereignty for someone else running the box.

Best for: technical users who want full ownership, unlimited customization, and a messaging-native agent, and who accept the operational burden that comes with self-hosting. For non-technical users and businesses, the calculus inverts: a hosted platform with a professional security team eliminates the entire category of self-hosting risk, which is exactly the trade our OpenClaw workforce guide walks through in depth.

8. Eigent and the Open-Source Cowork Desktop Wave

This section did not exist in our July revision, and its absence was our biggest blind spot, because half of what now ranks for "Claude Cowork alternatives" is open-source lists. A genuine product category formed here in the first half of 2026: open-source Cowork desktops, applications that replicate the Cowork experience (a local agent working your files, browser, and terminal) with your own model keys, for free. The category leader by adoption is Eigent, which describes itself as "the open source Cowork desktop application, empowering you to build, manage, and deploy a custom AI workforce," licensed Apache 2.0 with 14.7k GitHub stars and built on the CAMEL-AI multi-agent framework - GitHub.

What earns Eigent a ranked entry rather than a footnote is that it is not a single-agent clone. Its Workforce mode decomposes a task across a pool of specialized agents (Browser, Terminal, Multi-modal, and Document agents) that "divide work, collaborate in parallel, and execute complex multi-step workflows together," alongside a single-agent mode for direct tasks - Eigent. It is model-agnostic: point it at any provider's API or at local models, which makes it the natural fallback when a hosted frontier model is suspended, repriced, or deprecated. Self-hosting is free with your own keys; for users who want the convenience without the setup, Eigent also sells hosted credit plans at $19.90/month (Plus, 2,000 credits) and $99.99/month (Pro, 10,000 credits) - Eigent. The honest limits mirror OpenClaw's: security, updates, and cost control are yours, there is no mobile or messaging surface, and a community project's support story is a GitHub issue tracker, not an SLA.

Eigent is the strongest entry in a wider wave worth knowing by name. Kuse builds an open-source cowork desktop with a Rust-native core; OpenWork wraps Claude, OpenAI, Gemini, DeepSeek, and local models behind a bring-your-own-key desktop; AionUi does similar work for users who started on CLI agents and want a graphical surface. We deliberately keep the descriptions of these three short: the projects are young, their claims move weekly, and the durable point is the category, not any single repo. The category exists because Cowork proved the interaction pattern, and open-source proved (with OpenClaw, with Eigent) that the pattern is not proprietary. The moat questions for hosted vendors get harder from here.

The first-principles read on why this wave matters goes beyond price. A free clone of a $20/month product saves $240 a year, which is nothing to a business. What the open-source desktops actually sell is control: over which model runs your work, over where your files and credentials live, over whether a vendor's export-control suspension or pricing change can interrupt your operations. Section 17 argues that concentration risk is now a real line item; this category is its hedge. Best for: technical users and cost-sensitive teams who want the Cowork pattern under their own control, developers who want a multi-agent workforce they can extend, and any organization that needs a tested fallback path for the day its hosted agent vendor has a bad week.

9. Perplexity Comet

If the question is "what is the cheapest way to get a real agent working for me today," the answer in August 2026 is still Perplexity Comet, because the price is zero. Comet launched in July 2025, became free in October 2025, shipped Android on November 20, 2025 and iOS on March 18, 2026, making it the one agentic browser available free on all four major platforms - Wikipedia. With Atlas shutting down on August 9, Comet's position sharpens: it is now the last major standalone agentic browser standing, the only one that survived the consolidation wave as an independent product, and it did so by being free before its competitors were even cross-platform.

Comet's model of agency is browser-scoped. It navigates, researches, compares, fills forms, and manages tabs on your behalf, with the assistant living in a sidecar that can see and act on the page you are viewing, and its 2026 updates pushed it into producing documents and spreadsheets directly from research prompts. In practice, Comet pairs with a desktop or cloud agent rather than replacing one: Comet handles the exploratory phase (compare fifteen vendors, collect pricing pages, generate the first-pass comparison sheet), while the local or workforce agent handles the execution phase against your own files and systems. Treating "which agent" as an either-or question is a 2025 habit; the zero price of one side of the pairing makes the both-and answer free to adopt. Our browser use agents review ranks the wider browser-agent field this logic sits inside.

The monetization sits above the free browser: Perplexity Pro at $20/month ($200/year) and Max at $200/month ($2,000/year) unlock heavier model access - Fello AI, and Perplexity shifted to a subscription-first model in February 2026, discontinuing its AI-integrated advertising strategy - Wikipedia. That pivot answers half of the strategic question we raised in July (how does a free agentic browser get paid for) in the reader's favor: subscriptions, not ads injected into agent behavior. The other half stands: free products attached to venture-scale burn rates historically change terms as monetization pressure arrives, so enjoy the free tier and keep your workflows portable.

The honest limits define its rank. Comet does not touch your local file system, does not run scheduled unattended jobs the way Cowork's background sessions or Spark do, and its governance story (what did the agent do, under which permissions, with what audit trail) is thinner than the enterprise entries in this guide. Best for: individuals who want to feel what agentic browsing is like without paying anything, mobile-first users, and research-heavy workflows that end in a document. It replaces the Cowork use case of "go find out and put it in a deck," not the use case of "operate my computer."

10. Manus

Manus remains the reference implementation of the cloud sandbox agent, and its 2026 story remains the market's best case study in regulatory risk: China's NDRC blocked Meta's acquisition in April 2026 and Meta cut ties that June, leaving Manus (built by Butterfly Effect, headquartered in Singapore) independent. You hand Manus a goal; it spins up a virtual computer, plans, browses, writes code, and produces deliverables, showing you its working session as it goes. Its signature feature is still Wide Research, which fans a task out to parallel sub-agents for large-scale comparison and collection work: where Cowork parallelizes your tasks, Manus parallelizes within a task, throwing dozens of sub-agents at a research sweep simultaneously.

Pricing is credit-based, and the tier names consolidated under a "Pro" banner since our July revision - Lindy: Free grants 300 daily credits plus 1,000 starter credits, Pro Standard is $20/month (4,000 credits), Pro Customizable $40/month (8,000 credits, includes Wide Research), Pro Extended $200/month (40,000 credits), and Team runs $20 per seat with a 2-seat minimum and shared credit pools. The structural weakness is unchanged and worth stating precisely: complex or long-running tasks consume more credits, there is no pre-run cost estimate, and a single deep-research job can burn through a meaningful slice of a monthly allocation before you know what it cost. Credit opacity is the recurring complaint in user reports, and it is the fine print that determines real cost.

Against ChatGPT Work, now its closest rival, the comparison comes down to two asymmetries. Manus wins on parallel breadth: Wide Research fanning dozens of sub-agents across a market-scan task has no equivalent inside ChatGPT at any tier, and for genuinely wide tasks it is the difference between an afternoon and a coffee break. OpenAI wins on ecosystem depth and price predictability: ChatGPT Work ships inside an app your team already pays for and has already cleared with IT, meters against a flat plan instead of an opaque credit pool, and delivers the same finished-artifact experience for the common case. The pricing tiebreak favors whoever matches your duty cycle: at light usage Manus's free 300 daily credits beat a $20 subscription; at daily heavy usage the flat plan is the safer bill.

Best for: individuals and small teams who want autonomous, deliverable-producing cloud work without owning any infrastructure, and researchers who genuinely benefit from Wide Research's fan-out. Governance remains the soft spot for regulated buyers: a consumer cloud agent with credit opacity and a Singapore jurisdiction profile is a harder procurement conversation than a Microsoft or AWS entry, and that, more than capability, is what caps its rank here.

11. Amazon Nova Act and Alexa+

Amazon's position in this market is deliberately narrow, and that narrowness is its argument. Nova Act, generally available since December 2025, is not a personal coworker but reliable UI automation as an engineering primitive: an SDK claiming over 90% task reliability at scale, with a Playground, IDE extensions, and deployment through Bedrock AgentCore, currently limited to the US East (N. Virginia) region - AWS. Pricing is pure consumption at $4.75 per agent-hour, with each parallel agent billed separately and human-in-the-loop wait time excluded from the meter - AWS pricing. The 90%-reliability framing targets the RPA replacement market: workflows that run thousands of times, where a 60%-reliable general agent is useless but a 90%+-reliable scoped agent replaces brittle, selector-based scripts.

The RPA-replacement math is the part worth taking to a spreadsheet, because the per-hour meter cuts both ways. A Nova Act workflow that runs 30 minutes a day costs about $71/month in agent-hours and, because the agent perceives the UI rather than hard-coding selectors, it survives cosmetic UI changes that would kill a traditional RPA bot along with the engineering tax of fixing it. Run the meter the other way and the ceiling appears just as fast: an agent that must be live 24/7 costs roughly $3,400/month, at which point a flat-rate platform or a queue-based design (batch the work, run the agent an hour a day) is the sane architecture. Nova Act rewards teams that think in workflows-per-hour, not in headcount-replacement metaphors.

The consumer side of Amazon's agent story is Alexa+, free for Prime members and $19.99/month otherwise, which books, orders, and coordinates smart-home and shopping tasks conversationally - TechCrunch. It is not a desktop work agent and does not pretend to be, but for household-operations automation it reaches more people than every other entry in this guide combined, and it is the closest thing to agentic capability that most non-technical consumers will touch this year.

Best for: engineering teams on AWS automating high-volume, well-defined UI workflows, and organizations replacing legacy RPA with something maintainable under AWS IAM governance. The trade-offs are equally clear: per-agent-hour billing compounds quickly for always-on use, the single-region availability constrains latency-sensitive and data-residency use cases, and nothing here helps an individual knowledge worker the way Cowork, ChatGPT Work, or a workforce platform does.

12. xAI Grok 4.5 and Grok Build

The xAI entry needed the most correction of any profile in this refresh, on three separate axes. First, the model: our July revision said only that "Grok 5 is training on Colossus 2," missing that Grok 4.5 launched July 8, 2026 as xAI's current flagship, positioned specifically for coding and agentic work on a 1.5-trillion-parameter foundation, with training that incorporated real developer session data - Fello AI. API pricing is aggressive: $2/$6 per million tokens with a 500K context window, cached input at $0.30 per million, and rates that double above 200K tokens per request - Fello AI. Our Grok 4.5 benchmarks and pricing profile covers the model in depth.

Second, the plans: we previously wrote that no mid-tier existed between the standard subscription and Heavy, which is now flatly wrong. The current ladder runs SuperGrok Lite at $10/month, SuperGrok at $30/month (with Grok 4.5 in staged rollout), SuperGrok Plus at $100/month, and SuperGrok Heavy at $300/month, which remains "the only consumer plan with confirmed full access to Grok 4.5" - Fello AI. That last clause is the catch that keeps this entry at the bottom of the ranking: the mid-tiers exist now, but the full current-generation agent capability is still gated behind the steepest consumer price in this guide. Grok Build, the agentic coding tool, now reaches developers through the xAI API rather than being a Heavy-exclusive perk, positioning it against Claude Code; our Claude Code pricing comparison shows what that rivalry costs on each side.

Third, the corporate structure, which procurement teams price as risk: SpaceX acquired xAI on February 2, 2026 in an all-stock transaction that valued SpaceX at $1 trillion and xAI at $250 billion, a combined $1.25 trillion, making xAI a wholly owned subsidiary - Wikipedia. Musk has since signaled that xAI stops existing as a standalone brand, moving under a SpaceX AI unit. A frontier lab inside an aerospace company is a genuinely novel structure, and the honest read is that nobody yet knows what it means for product continuity, enterprise contracts, or governance; "post-merger flux" is not a criticism, it is a factual description of the current state.

The genuinely interesting technical idea carries over from July: Grok's heavy modes throw a committee of parallel specialized agents at a problem and synthesize their outputs, trading compute for breadth. For open-ended research and hard coding problems, the fan-out approach produces results single-agent systems miss; for routine multi-step execution it is expensive overkill. Best for: developers who want an aggressive agentic model at $2/$6 API pricing, and teams betting that xAI's compute scale converts into a durable capability lead. For mainstream desktop autonomy at sane subscription prices, it is still not a Cowork substitute, which is why it anchors this ranking rather than leading it.

13. The Agentic Browser Shakeout

Our July revision called agentic browsers "where most consumer agent usage actually happens," and one month later the category is consolidating exactly the way standalone agents did a year earlier. The scoreboard as of August 5: ChatGPT Atlas shuts down August 9, 2026, absorbed into OpenAI's unified desktop application - Wikipedia. Chrome Auto Browse continues as the incumbency play, agentic execution inside the browser people already use, with human-approval gates on sensitive actions - TechCrunch. And Perplexity Comet becomes the last major standalone agentic browser, free on macOS, Windows, Android, and iOS - Wikipedia.

The first-principles logic that created the category still holds: the browser is where web-work already lives, sessions and logins included, so an agent embedded there gets capability "for free" that a desktop agent has to build through connectors. What July clarified is that this logic favors incumbent surfaces, not new browsers. Asking users to switch browsers turned out to be nearly as hard for AI companies as it was for everyone else who tried it in the past twenty years; giving the existing browser agentic powers (Google's play) or folding the browser into a bigger desktop surface (OpenAI's play) is where the distribution physics point. Comet's survival as the exception is explained by its price: free removed the switching cost that killed the paid pitch.

The decision rule for readers is unchanged and worth restating. If a task's inputs and outputs both live on the web (booking, shopping, form-filling, research, expense portals), a browser agent is the lowest-friction executor. If the task touches local files or desktop software, you need Cowork, Copilot Cowork, or an open-source desktop. If it should run without you present at all, you need a cloud or workforce agent. The category boundaries blurred slightly this summer (Cowork's web version runs in a browser tab, Comet generates documents), but the underlying question (where does the work live, and does it need you present) still sorts every product cleanly.

The security dimension remains the category's defining caveat. Browser agents inherit your logged-in sessions, which makes them powerful and makes prompt injection via malicious page content their signature attack vector: a hostile page can try to instruct the agent operating inside your authenticated browser. Vendors mitigate with action confirmations, site scoping, and payment gates, but the honest state of play is that the approval prompts exist because the vendors themselves do not fully trust their agents on the open web yet. Weigh convenience against blast radius accordingly: a browser agent with access to your email session can do most of what a stolen cookie can.

14. Benchmark Reality: The Road to 90%

Every previous version of this guide had a benchmark headline that expired. The original said agents "fail two out of three times." The July revision celebrated agents crossing OSWorld's 72.36% human baseline. Both are now history rather than news, because the frontier moved again, hard. On July 27, 2026, Intelligence Indeed, an enterprise automation company from Zhejiang founded in 2018 with over 6,000 enterprise clients, announced its Z-Agent had reached 90.2% on OSWorld (325.59 points across 361 tasks), the first agent past 90% on the benchmark built to measure real computer use - GlobeNewswire. The same release traces the arc: roughly 12% when OSWorld launched in 2024, 72.6% at the end of 2025, 83.6% by May 2026, and now 90.2%.

The independently tracked leaderboard tells the same story with the names you can actually buy. On OSWorld-Verified, the externally validated variant, the top of the August table reads: Qwen3.8 Max at 86.1%, Claude Fable 5 at 85.0%, Claude Opus 4.8 at 83.4%, Gemini 3.6 Flash at 83.0%, Claude Sonnet 5 at 81.2%, and Meta's Muse Spark 1.1 at 80.8% - LLM Stats. Two things about that list deserve attention. First, every model on it clears the human baseline by eight points or more; "crossed the human baseline" went from milestone to table stakes in seven months. Second, the top spot belongs to an open-weight-adjacent Chinese lab and the 90% announcement to a Chinese enterprise automation vendor, which quietly ends the assumption that computer-use capability is an American duopoly. Anthropic's own OSWorld 2.0 claim for Opus 5 (best capability at any given cost) adds the economic dimension: the frontier is now contested on cost per completed task, not raw score - Anthropic.

The web-navigation picture still carries the underappreciated result from late 2025: Microsoft's Fara-7B, a 7-billion-parameter open model that runs quantized on a Copilot+ laptop, scores 73.5% on WebVoyager, ahead of hosted computer-use models many times its size - Microsoft Research. Competent web agency no longer requires a datacenter, which matters for privacy-sensitive automation, for marginal cost, and for the open-source desktop category in section 8, whose entire premise is that good-enough agent capability now runs on hardware you own.

Now the necessary honesty about what these numbers do and do not mean, updated for the 90% era. A 90.2% benchmark score does not mean an agent completes 90% of your workflows; it means that on a fixed distribution of scoped desktop tasks, the best system now fails one time in ten rather than two times in three. Production reliability still depends on task specification, site weirdness, and error recovery, which is why every serious platform keeps human checkpoints in the loop and why Amazon still leads with scoped-workflow reliability rather than benchmark scores. But the directional update is large and buyers should make it: the failure-rate objection that justified waiting expired in 2025, and the 2026 differentiators among the tools in this guide are reliability on your tasks, governance, and cost per completed task. Our computer-use benchmarks guide tracks the leaderboards as they move, and our best LLM for AI agents ranking maps model choice onto agent workloads.

15. Enterprise Consolidation: watsonx, Artemis, and ServiceNow

The enterprise layer changed the least since July, which is itself information: while the consumer agent market repriced weekly, the control-plane layer settled into the shape the 2025 acquisitions predicted. The thesis carries over intact: enterprises are not buying individual agents, they are buying control planes that manage fleets of them. If Cowork and its rivals answer "who does my work," this layer answers "who governs a thousand agents doing the company's work," and the two purchases should be negotiated separately.

The three landmarks are unchanged. Moveworks operates inside ServiceNow following the acquisition that completed December 15, 2025, making any Moveworks evaluation a ServiceNow platform decision - Moveworks. Kore.ai runs its Artemis Agent Platform around a declarative Agent Blueprint Language, agents defined as versionable specifications rather than dashboard configurations - Kore.ai. And IBM watsonx Orchestrate sells platform subscriptions from $530/month for its Essentials edition - IBM, with its next generation positioned explicitly as a cross-vendor agent control plane: govern everyone's agents, including competitors', the classic IBM middleware playbook applied to the agent era.

What did change is that Microsoft entered this layer with a bundle price, and it recalibrates everyone else's pitch. The E7 Frontier Suite at $99/user/month (section 4) packages agent observability, security, and governance with the productivity suite itself - Microsoft, which forces the standalone control planes to justify their existence on cross-vendor coverage: E7 governs Microsoft's agents beautifully and everyone else's barely, and a real enterprise in 2026 runs agents from four or five vendors simultaneously. That is the gap watsonx and Artemis live in, and it is a durable one for as long as the agent market itself stays plural.

The practical advice for buyers is the same as July, sharpened by E7's arrival: pick your execution agents on capability and cost from the top-10 ranking, and negotiate your governance layer separately, because bundling both decisions with one vendor is how you end up paying control-plane prices for mediocre agents, or (the E7 version of the mistake) letting a governance bundle default you into a single vendor's agents across the whole estate. A 50,000-employee company will plausibly run Copilot Cowork for M365 work, browser agents for web tasks, workforce platforms like O-mega for autonomous operations, and a control plane above all of it for identity, policy, and audit. The layered future is here; buy it in layers.

16. Verified Pricing in August 2026

Pricing in this market changes fast enough that our July table was stale in three places within a month (Grok's tiers, Manus's tier names, Microsoft's E7 bundle), so this section states current prices with the sources they were verified against, and, more usefully, the three billing models they cluster into, because the model predicts your bill better than the sticker does. Flat subscriptions (Claude, ChatGPT, Google, Perplexity, O-mega, xAI's SuperGrok ladder) cap your spend and meter you with usage limits. Consumption billing (Copilot Cowork's $0.01 per credit, Nova Act's $4.75 per agent-hour) scales cost with execution, efficient for spiky workloads and dangerous for always-on ones. Credit systems (Manus, Eigent's hosted tiers) sit in between, flat-priced but internally metered, with the burn rate as the fine print.

Three readings of this data earn their place in a buying memo. First, free got serious: the zero-cost column now contains not just a browser (Comet) but two genuine Cowork-pattern desktops (OpenClaw, Eigent) and ChatGPT Work's free desktop tier, which means "try before you buy" now covers the entire category and any paid commitment should follow a free-tier bake-off, not a feature-page comparison. Second, the $20 price point still won the mainstream: Claude Pro, ChatGPT Plus, Manus Pro Standard, and Perplexity Pro all cluster there, so capability and fit, not price, should drive the choice among hosted options. Third, consumption billing still deserves arithmetic before commitment: Nova Act at $4.75/agent-hour is cheap for a 10-minute daily job and brutal for an always-on worker, and Copilot Credits have the same shape at finer grain. Match the billing model to your duty cycle, not to the marketing page.

17. Safety, Governance, and Concentration Risk

June and July taught this market two different governance lessons, and mature buyers now hold both. The June lesson: access to frontier agents is a policy variable. Claude Fable 5 spent three weeks suspended under US export controls before returning July 1, and GPT-5.6 launched June 26 as a restricted preview for a small group of trusted partners under government restrictions before its public release cleared on July 9 - Wikipedia. Both episodes resolved quickly, which is exactly why they should be read as precedent rather than crisis: the mechanism now exists, it has been used, and a procurement process that ignores it is modeling 2024.

The July lesson is the counterweight: the fallback options got real. The mitigations we recommended in July (multi-vendor readiness, open-source fallbacks, contractual clarity about vendor obligations during a pause) all strengthened in one month. Opus 5's arrival at Fable-class capability for a third of the cost widened the intra-vendor fallback inside Anthropic's own lineup - Anthropic. The OSWorld-Verified top six now spans four independent labs on two continents - LLM Stats, so a workflow that fails over between model families gives up far less capability than it did a year ago. And the open-source desktop wave (section 8) turned "run it yourself" from an OpenClaw-shaped technical adventure into a category with multiple maintained, model-agnostic products. Concentration risk did not disappear; the cost of hedging it collapsed.

The second governance frontier, agent containment, continues converging on one design philosophy across every serious vendor: run agents as constrained principals whose actions are observable and interruptible. Windows 11's Agent Workspace runs each agent as a separate non-admin account in an isolated session - BleepingComputer. Chrome's Auto Browse and Gemini Spark pause for human approval on sensitive actions, with Spark's app connectors off by default - Google. Claude Cowork checkpoints multi-step plans for inspection. Workforce platforms log every agent session for review. Microsoft now sells the whole posture as a SKU in E7. The design consensus is identity-shaped: the agent is a junior employee with scoped permissions and a supervisor, not a superuser.

The two frontiers interact in a way that is easy to miss, and the June-July sequence made it concrete. OS-level containment protects you from your agent; availability policy protects (or interrupts) your access to the agent itself. An organization can be perfectly hardened on the first axis and completely exposed on the second: no amount of workspace isolation helps when the model behind every workspace is paused. Treat them as separate line items in the same risk register, one owned by security engineering and one owned by vendor management, with the fallback plan (a tested second model family, or a local stack like Eigent with Fara-class models for the workflows that must not stop) rehearsed rather than theoretical. When you evaluate any platform in this guide, ask three questions in order: what can the agent touch, stated as an explicit permission boundary; what record exists of what it did and who can review it; and what happens, technically and contractually, when the model behind it is suspended, deprecated, or degraded. The vendors that answer all three crisply are the ones treating agents as production infrastructure rather than demos.

18. Decision Framework: Choosing in August 2026

Eighteen months of shutdowns, mergers, and benchmark milestones still compress into one primary question: does the work need you present, or not? That fork, more than any feature list, sorts the August 2026 market. Attended work (you are there, steering) belongs to local agents and browser agents. Unattended work (it runs at 3 a.m., you review the output) belongs to cloud agents, background agents, and workforce platforms. July blurred the fork's edges (Cowork's background sessions run unattended, ChatGPT Work runs for hours) without breaking it, because both remain one person's task queue, which is a different thing than standing roles. Everything else is a second-order refinement of where the work's inputs live and who you trust to hold the keys.

Applying the framework to the ranking: if your work is your machine and your files, Claude Cowork at $20/month remains the default, with Copilot Cowork taking its place wherever Microsoft 365 is the system of record and Eigent taking it wherever control beats convenience. If your work is the web with you watching, Chrome Auto Browse rides the browser you already use and Comet is free, so trying the category costs nothing. If the work should run without you, match the shape: ChatGPT Work for deliverable-producing background runs, Gemini Spark and Workspace Studio for Google-native standing tasks, Manus for research fan-outs, O-mega when the real need is a set of standing roles executing around the clock, and Nova Act when the job is high-volume scoped automation with an SLA. Sovereignty-first users start from OpenClaw or Eigent; developers wanting raw model aggression at API prices look at Grok 4.5.

Three conclusions from this refresh update the ones we drew in July. First, the free tier became the honest starting point: with ChatGPT Work free on desktop, Comet free everywhere, and two open-source Cowork desktops at zero, every reader can now run a real bake-off before paying anyone, and should. Second, the benchmark era of this decision ended: when six models from four labs all clear the human baseline by eight-plus points and the frontier argument is cost per completed task, picking by leaderboard position is picking on noise; billing model, governance, and where the agent lives are the real differentiators. Third, hedging became cheap: after an export-control suspension, a blocked acquisition, and three product shutdowns in twelve months, the rational posture is one primary platform plus one fallback you have actually tested, and for the first time the fallback can be free, open source, and running on your own hardware by the weekend.

This guide was researched and written by Yuma Heymans (@yumahey), founder of O-mega and co-founder of HeroHunt.ai, who spent July watching half of this article's previous edition go stale in real time while the agent workforces he operates kept running through every model launch, which is roughly the argument for multi-vendor resilience in one sentence.

This guide reflects the AI agent landscape as of August 5, 2026. Pricing, availability, and model versions in this market change monthly (sometimes weekly): verify current details on the linked official pages before purchasing.