Top 10 Claude Cowork Alternatives (July 2026 Guide)

The honest July 2026 comparison of every serious alternative to Claude Cowork, with verified pricing, current benchmarks, and a decision framework built for how agents actually work now.

Claude Cowork went from a Mac-only research preview to general availability on every paid Claude plan in under four months. On April 9, 2026, Anthropic dropped the beta badge and shipped Cowork to every subscriber from the $20/month Pro plan upward - Claude pricing. The product that started 2026 as an expensive curiosity for early adopters is now the mainstream default for autonomous desktop work, running on macOS, Windows, ChromeOS, and in web and mobile beta - Claude Cowork.

But here's the problem: the alternatives landscape changed even faster than Cowork did, and almost everything written about it in 2025 is now wrong. OpenAI Operator no longer exists as a standalone product (it was shut down August 31, 2025 and absorbed into ChatGPT). Google Project Mariner was discontinued on May 4, 2026. Meta's acquisition of Manus was blocked by Chinese regulators, and Meta cut ties entirely. Microsoft shipped a competitor that is literally built on Claude Cowork's own technology platform. And autonomous agents quietly crossed the human baseline on the industry's hardest benchmark in December 2025, killing the old "agents fail two out of three times" narrative for good.

This guide breaks down exactly what changed, re-baselines what Claude Cowork actually is in July 2026, and ranks the 10 real alternatives that survived the past year of shutdowns, acquisitions, and consolidation. Every price, every benchmark, and every availability claim is sourced from an official page or primary report as of July 8, 2026. Where the old consensus was wrong, we say so explicitly. When we first published this guide in January 2026, several of the entries below did not exist in their current form; this is a full rewrite against the July 2026 reality, not a cosmetic date bump.

Contents

1. What Changed Since January 2026
2. Claude Cowork in July 2026: The New Baseline
3. ChatGPT Agent and Atlas (OpenAI)
4. Microsoft Copilot Cowork
5. Google Gemini Agent, Spark, and Chrome Auto Browse
6. O-mega
7. OpenClaw
8. Perplexity Comet
9. Manus
10. Simular Agent S3
11. Amazon Nova Act and Alexa+
12. xAI Grok and Grok Build
13. The Agentic Browser Wave
14. Benchmark Reality: Agents Crossed the Human Baseline
15. Enterprise Consolidation: watsonx, Artemis, and ServiceNow
16. Verified Pricing in July 2026
17. Safety, Governance, and the Export-Control Moment
18. Decision Framework: Choosing in July 2026

The July 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 now enormous ($0 for Comet and OpenClaw versus $300/month for Grok Build). Reach gets 20% because agents only help where you actually work. Governance gets 25% because 2026 is the year agent permissions, isolation, and audit trails became 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 Since January 2026

If you last researched Cowork alternatives in early 2026, your mental map is obsolete. The first half of 2026 brought the most violent consolidation the agent market has seen, and no single competing article documents the full graveyard. This section is the changelog: what died, what merged, what shipped, and what got blocked, with dates you can verify against primary sources.

The structural force behind all of it is simple: standalone agent products lost to agents embedded in surfaces people already use. Operator died as a separate website and was reborn inside ChatGPT. Mariner died as a separate experiment and was reborn inside Chrome and the Gemini app. Even Anthropic's own strategy reflects the same physics: Cowork stopped being a separate Mac preview and became a mode of the Claude apps everyone already had. When distribution beats novelty, the standalone agent surface is the first casualty, and the second half of 2026 will be judged on which embedded agent executes most reliably, not on who has the flashiest demo.

The verified timeline of the consolidation:

• August 31, 2025: OpenAI shuts down standalone Operator, folding its Computer-Using Agent into ChatGPT Agent - Wikipedia
• December 15, 2025: ServiceNow completes its acquisition of Moveworks - Moveworks
• April 9, 2026: Claude Cowork hits GA on every paid Claude plan
• April 27, 2026: China's NDRC blocks Meta's acquisition of Manus; Meta formally cuts ties June 15, 2026 - Wikipedia
• May 4, 2026: Google discontinues Project Mariner, moving its capabilities into Gemini Agent and Chrome - Digital Trends

Two more dates complete the picture. On March 9, 2026, Microsoft announced Copilot Cowork, built by bringing the same technology platform that powers Claude Cowork into Microsoft 365, and took it to general availability on June 16, 2026 - Microsoft. And on January 28, 2026, Chrome gained Auto Browse, giving Google's browser native autonomous task execution for AI Pro and Ultra subscribers - TechCrunch. The pattern across every one of these events is identical: the agent moved closer to where the user already was, and the standalone surface disappeared.

Why this matters for your buying decision: any comparison that still ranks "Operator" or "Mariner" as live alternatives was written before the market restructured, and its pricing and capability claims are almost certainly equally stale. How to apply it: check the date on every agent comparison you read, and treat anything citing a standalone Operator URL or a Mariner waitlist as a signal that the entire article predates the consolidation. The profiles in this guide start from the post-consolidation reality, and where a dead product's capability moved somewhere new, we follow it to its new home rather than pretending the category disappeared.

2. Claude Cowork in July 2026: The New Baseline

You cannot evaluate alternatives to something you misunderstand, and most of the internet still describes Claude Cowork as it existed in its January 2026 research preview: Mac-only, expensive, waitlisted. That product no longer exists. The July 2026 Cowork is a general-availability feature of every paid Claude plan, and it sets the baseline every alternative in this guide is measured against.

Start with access and platforms. Cowork reached general availability on April 9, 2026 and is included in Claude Pro at $20/month ($17/month billed annually), Max 5x at $100/month, Max 20x at $200/month, Team at $20 per seat on standard or $100 per seat on premium, and Enterprise - Claude pricing. It runs on macOS, Windows (x64 and arm64), and ChromeOS, with web and mobile in beta, and supports scheduled unattended runs and parallel task processing - Claude Cowork. The one honest caveat Anthropic itself documents: Cowork consumes usage limits faster than Chat mode, because autonomous multi-step work burns far more tokens than conversation. Anthropic even doubled the 5-hour usage limits at no charge for Pro, Max, and Team users through July 5, 2026 to drive adoption - The New Stack. For a full walkthrough of the product itself, see our Claude Cowork desktop, web, and mobile guide.

Under the hood, Cowork runs on Anthropic's current flagship, Claude Fable 5, which launched June 9, 2026 with a 1M-token context window priced at $10/$50 per million input and output tokens - Anthropic. Fable 5's first month was itself a governance story: the model was suspended under US export controls from June 12 to June 30 and returned globally on July 1, 2026 across Claude.ai, Claude Code, and Claude Cowork. That three-week suspension (covered in depth in section 17) is the clearest signal yet that frontier agent capability is now a regulated commodity, which directly affects how you should think about single-vendor concentration risk. We benchmarked Fable 5 against its peers in our Fable 5 and Mythos 5 benchmarks breakdown.

What does Cowork actually do that a chatbot does not? It works in a real folder on your machine (or a cloud sandbox on web and mobile), reads and writes files, drives your browser through a connector, runs scheduled jobs while you sleep, and executes multi-step plans with checkpoints you can inspect mid-run. The mental model that matters: Cowork is a local-machine agent with your file system as its workspace. Every alternative below makes a different architectural bet (cloud sandbox, browser, messaging channel, or OS-level workspace), and that bet, more than any benchmark score, determines which tool fits your work. Our Cowork pricing and agent ecosystem analysis covers the plan-by-plan economics in detail.

So why look at alternatives at all? Four honest reasons survive contact with the July 2026 product. First, usage ceilings: heavy Cowork use hits plan limits, and scaling to Max 20x costs $200/month. Second, ecosystem gravity: if your work lives in Microsoft 365 or Google Workspace, an agent native to those suites touches your real documents with less friction than a general-purpose desktop agent. Third, unattended scale: Cowork schedules tasks and parallelizes them, but it is still fundamentally one user's copilot, not a fleet of independent workers with their own accounts and browsers. Fourth, sovereignty: the June export-control suspension proved that access to a hosted frontier agent can be interrupted by geopolitics, which makes open-source and multi-vendor strategies rational for some organizations. Each profile below tells you which of these four gaps it actually closes, and which it does not.

3. ChatGPT Agent and Atlas (OpenAI)

The single most important correction to the old version of this guide: OpenAI Operator is gone. The standalone Operator surface, which we originally ranked first as a "$200/month US-only research preview," was shut down on August 31, 2025, and its Computer-Using Agent technology was folded into ChatGPT Agent - Wikipedia. The Operator website no longer exists. What replaced it is both cheaper and more broadly available: agent mode now ships to Plus subscribers at $20/month, not just the $200 Pro tier, and OpenAI added a worldwide ChatGPT Go tier at $8/month in January 2026 for lighter usage - Wikipedia. We tracked the original Operator's pricing evolution in our ChatGPT Operator pricing analysis, which now reads as a historical document of how fast this market reprices.

ChatGPT Agent works fundamentally differently from Cowork. Where Cowork operates on your local machine with your files, ChatGPT Agent runs in a cloud sandbox with its own virtual browser and computer, executing research, form-filling, purchases, and document generation on OpenAI's infrastructure. You watch it work, interrupt it, or take over the browser when a login or payment needs your hands. The practical consequence: ChatGPT Agent cannot reorganize your local folders or touch files you have not uploaded, but it also cannot break anything on your machine, which makes it the lower-stakes entry point for people nervous about desktop autonomy.

The second half of OpenAI's answer is ChatGPT Atlas, the Chromium-based agentic browser launched October 21, 2025 on macOS, which brings agent mode to Plus, Pro, and Business subscribers directly inside a browser they use all day - Wikipedia. In March 2026, OpenAI said it will merge Atlas, the desktop app, and Codex into a single desktop application, a clear signal it intends to meet Cowork head-on at the OS level rather than staying browser-bound. The honest caveats: as of May 2026, Atlas still had no Windows or mobile version, and agent mode inside Atlas inherits the same usage-limit pressure Cowork users know well.

It is worth being precise about what actually improved in the Operator-to-Agent transition, because "absorbed into ChatGPT" undersells it. The original Operator was a browser in a box: it could click and type on websites, and that was the whole product. ChatGPT Agent inherits that computer use but wraps it in the rest of ChatGPT's machinery: connectors into Gmail, Drive, SharePoint, and calendars, so the agent can pull real context before it acts; memory across sessions, so recurring tasks do not start from zero; and scheduled tasks, so a research digest or a price check can run on a cadence without you initiating it. The failure modes moved accordingly. Operator's classic failure was getting lost on a website; Agent's classic failure is overreach, confidently assembling a deliverable from sources it misread, which is why the takeover-mode and approval-gate design matters more than it did when the agent could do less.

Model-wise, OpenAI is mid-transition as this guide publishes. The GPT-5.6 family (Sol as flagship, Terra mid-tier, Luna budget) was previewed to roughly 20 trusted partners on June 26, 2026 under a US government safety review, with the full public release landing July 10, 2026 - CNBC. Our GPT-5.6 benchmark and pricing breakdown covers what the new family changes for agent workloads. Best for: anyone who wants the most capable general agent at the lowest mainstream price, does not need local file-system access, and values the largest ecosystem of connectors and the fastest iteration cadence in the industry.

Pricing - Wikipedia:

4. Microsoft Copilot Cowork

This entry did not exist when we first published this guide, and it carries the strangest lineage in the entire market: Copilot Cowork is built on Claude Cowork's own technology platform. Microsoft announced it on March 9, 2026, explicitly describing it as bringing "the technology platform that powers Claude Cowork" into Microsoft 365 Copilot through its partnership with Anthropic, backed by Anthropic's $30 billion Azure compute commitment - Microsoft. It entered the Frontier preview program March 30, reached general availability June 16, 2026, and bills on consumption at $0.01 per Copilot Credit rather than a flat seat price. In June 2026, Claude's Opus-class models also became selectable inside Copilot Chat, making Microsoft simultaneously Anthropic's biggest distribution partner and its most direct competitor.

The strategic meaning deserves a moment of first-principles attention, because no competing comparison explains it clearly. Anthropic proved that the scarce asset in this market is not the agent harness (Microsoft licensed that) but the model quality and the distribution surface. Microsoft owns the distribution: Copilot Cowork operates natively across Word, Excel, PowerPoint, Outlook, Teams, and SharePoint, where the world's enterprise documents already live. If your organization runs on Microsoft 365, Copilot Cowork touches your real work products without any file migration, which is precisely the friction that keeps desktop agents from being adopted by non-technical staff. Our Copilot Cowork complete analysis dissects the architecture and the Anthropic deal in depth.

The consumption billing deserves arithmetic before you commit a department to it, because it is the one dimension where Copilot Cowork and Claude Cowork genuinely diverge. At $0.01 per Copilot Credit, light users cost almost nothing, which is why consumption billing is attractive for organizations where 80% of staff will use agents twice a week. But an enthusiastic power user running long agentic sessions daily can quietly out-spend a $200/month Max subscription, and unlike a flat plan, nothing in the billing model tells them to stop. The practical adoption pattern that works: pilot with a spending alert per user, measure the credit burn of your five most common workflows for a month, and only then decide which population belongs on consumption billing versus which heavy users would be cheaper on Anthropic's flat plans, given that the underlying agent technology is, unusually, the same.

Microsoft's second front is the operating system itself. Windows 11's Copilot Actions and Agent Workspace entered Insider preview in November 2025: each agent runs under a separate non-admin Windows account inside an isolated workspace session, so agent activity is auditable, permission-scoped, and interruptible at the OS level - Windows Forum. This is the most serious OS-level answer yet to the question every security team asks about desktop agents: "what exactly can it touch, and how do I prove it?" No other vendor in this guide runs agents as first-class, isolated OS principals.

The third piece is small, fast, and local: Fara-7B, released November 24, 2025, is a 7-billion-parameter, MIT-licensed computer-use model that scores 73.5% on WebVoyager, beating OpenAI's computer-use-preview at 70.9% and UI-TARS-1.5-7B at 66.4%, with a quantized version that runs on-device on Copilot+ PCs - Microsoft Research. An on-device agent model means sensitive automation that never leaves the laptop. Best for: Microsoft 365 organizations that want Cowork-class autonomy inside their existing compliance boundary, and security teams that need OS-level agent isolation. The trade-off is consumption billing: $0.01 per Copilot Credit is easy to start and hard to forecast, a sharp contrast with Claude's flat monthly plans.

5. Google Gemini Agent, Spark, and Chrome Auto Browse

Our old #2 entry, Project Mariner, is dead. Google discontinued it on May 4, 2026, ending a 17-month standalone experiment - Wikipedia. But treating that as Google's exit from the category would be exactly the kind of stale reading this refresh exists to correct: Mariner's capabilities were redistributed into Gemini Agent in the Gemini app, Gemini Spark, and Chrome's Auto Browse, which together form a more coherent (if more fragmented-sounding) offering than Mariner ever was. We covered Mariner's original launch back when it was news - our Mariner launch coverage, and the contrast between that announcement and today's lineup is a case study in how fast Google reorganizes.

The headline product is Gemini Spark, Google's 24/7 personal AI agent for Google AI Ultra subscribers (US, 18+, English-only for now). Spark's differentiator is genuinely structural: it works tasks in the background even when your devices are off, connected to Gmail, Calendar, Docs, Sheets, and Slides - Google. Where Cowork's scheduled runs still conceptually belong to your machine and your session, Spark is a standing background worker attached to your Google identity. For people whose entire life runs on Workspace, that is the closest thing on this list to an always-on chief of staff.

The architectural distinction between Spark and everything else in this guide is worth spelling out, because it predicts which tasks each handles well. Cowork, Copilot Cowork, and OpenClaw are device-attached: their power comes from proximity to your files and applications, and they conceptually stop when the machine does. Spark is identity-attached: it belongs to your Google account, not your hardware, so "email me a summary of every contract that landed overnight and draft the three replies I always send" keeps running from Google's side regardless of what your devices are doing. The cost of that convenience is scope: Spark can only be brilliant inside the Google surfaces it is wired into, and Google's agent story remains fragmented across three products where Anthropic ships one. Expect that fragmentation to compress the same way Bard, Duet, and Assistant eventually compressed into Gemini.

The mass-market piece is Chrome Auto Browse, launched January 28, 2026: Gemini autonomously books flights, shops, and files expenses inside Chrome for AI Pro and Ultra subscribers in the US, pausing for human approval on sensitive actions like payments - TechCrunch. Because it ships inside the world's dominant browser, Auto Browse is the single widest distribution of agentic capability in existence, and its approval-gate design is discussed further in section 17.

On models and money: the current lineup is Gemini 3.1 Pro as the primary paid model, Gemini 3.5 Flash (launched May 19, 2026) for speed, and Gemini Omni for video - Google release notes. Google AI Ultra now has two tiers: $99.99/month with 5x Pro limits, and roughly $200/month (EUR 219.99) with 20x limits - Google subscriptions. Our Gemini 3.1 Pro complete guide profiles the flagship in detail. Note what is finally gone from Google's own materials: Bard and SGE, names our original article still used, were retired long ago; anything still using them is describing a Google that no longer exists. Best for: Google Workspace households and companies, and anyone who wants an agent that keeps working when the laptop lid is closed. The honest limits: Spark is US-only and English-only, and the best capability sits behind the $99.99+ Ultra tiers.

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, the same one we made when we first published this guide, and the market has since moved 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 and posts 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 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. 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. 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 Agent 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 original version of this guide speculated about open-source agents with the phrase "Llama 3 (if it's out by 2026)", which was embarrassing even then (Llama 3 shipped in April 2024) and is doubly so now, because the open-source agent story of 2026 has a name every developer knows: OpenClaw. The MIT-licensed personal agent amassed hundreds of thousands of GitHub stars and millions of users within months of exploding into public awareness - GitHub. Its creator, Peter Steinberger, subsequently joined OpenAI, which now sponsors the project under a non-profit foundation structure that keeps the code independent - OpenClaw.

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, Slack, and iMessage. There is no subscription, no usage ceiling imposed by a vendor, 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 is 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 which models you pick, 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. Our OpenClaw setup guide covers the configuration work involved, and it is not trivial.

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 (Cowork, Copilot Cowork, or a workforce platform like O-mega) eliminates the entire category of self-hosting risk, which is exactly the trade our OpenClaw workforce guide walks through. Eigent deserves a mention in the same breath: an open-source multi-agent workforce desktop app for users who want OpenClaw-style ownership with a more orchestrated, parallel-worker model.

8. Perplexity Comet

If the question is "what is the cheapest way to get a real agent working for me today," the answer in July 2026 is Perplexity Comet, because the price is zero. Perplexity removed Comet's paywall in October 2025, shipped Android in November 2025 and iOS in March 2026, and as of April 2026 the agentic browser is free on all four major platforms - Wikipedia. For a category that started 2025 gated behind $200/month subscriptions, a free, cross-platform agentic browser is the single most aggressive distribution move any vendor has made.

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. The March 2026 update pushed it beyond browsing into artifact generation: Comet now produces PowerPoints, spreadsheets, and dashboards directly from research prompts, collapsing the "research, then make the deliverable" workflow into one step. For students, analysts, and anyone whose output is documents-about-information, that covers a surprising share of daily work.

In practice, Comet pairs with a desktop agent rather than replacing one, and the combination is better than either alone. A workflow we see repeatedly: Comet handles the exploratory phase (compare fifteen vendors, collect pricing pages, pull reviews, generate the first-pass comparison spreadsheet), and the local agent handles the execution phase against your own files (merge the findings into the existing procurement doc, update the budget model, draft the internal memo in your house style). Treating "which agent" as an either-or question is a 2025 habit; the zero price of one side of that pairing makes the both-and answer free to adopt.

The honest limits define its rank. Comet does not touch your local file system, does not run scheduled unattended jobs the way Cowork 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. Perplexity monetizes through Pro at $20/month and the premium Max tier at $200/month for heavier model access, so the free tier is an on-ramp, not the whole product. There is also a strategic question a first-principles reader should hold: giving away an agentic browser is a user-acquisition play, and free products attached to venture-scale burn rates historically change terms as monetization pressure arrives.

Best for: individuals who want to feel what agentic browsing is like without paying anything, mobile-first users (it is the only agentic browser shipping on both Android and iOS), 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."

9. Manus

Entry #10 of our original guide claimed Meta had acquired Manus. That claim is now wrong twice over: China's NDRC blocked the acquisition on April 27, 2026, and Meta officially cut ties on June 15, 2026 - Wikipedia. Manus (built by Butterfly Effect, roughly 100 employees, headquartered in Singapore) continues as an independent company with approximately a $125M annual revenue run rate as of December 2025. The failed acquisition is arguably the most consequential regulatory event in the agent market to date, and it left Manus in a strange, instructive position: proof that a startup can hold its own surface against Big Tech, and proof that geopolitics now reaches into agent-market M&A.

The product itself remains the reference implementation of the cloud sandbox agent. 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. Manus 1.6 (December 15, 2025) added Chat Mode for lightweight interactions, Wide Research, which fans a task out to parallel sub-agents for large-scale comparison and collection work, and a Lite/1.6/Max model-tier selector. Wide Research is the feature to understand: 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 credits are the fine print that determines real cost - NoCode MBA: Free grants 300 daily credits, Standard is $20/month, Customizable $40/month, and Extended $200/month, with heavier tasks consuming credits faster. Credit systems are harder to forecast than flat plans, and complex tasks can burn through an allocation quickly, which is the recurring complaint in user reports. Governance is the other soft spot: for regulated industries, a consumer cloud agent with credit opacity and a Singapore jurisdiction profile is a harder procurement conversation than a Microsoft or AWS entry.

Against ChatGPT Agent, 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 is something ChatGPT Agent does not offer at any tier, and for genuinely wide tasks it is the difference between an afternoon and a coffee break. ChatGPT wins on ecosystem depth: connectors, memory, and the fact that the agent lives inside an app your team already pays for and has already cleared with IT. The pricing tiebreak favors whoever matches your duty cycle: at light usage Manus's free 300 daily credits beat a $20 subscription, while at daily heavy usage the $20 flat ChatGPT Plus plan is more predictable than watching a credit meter.

Best for: individuals and small teams who want autonomous, deliverable-producing cloud work (reports, sites, data jobs) without owning any infrastructure, and researchers who benefit from Wide Research's parallel fan-out. It is the closest like-for-like rival to ChatGPT Agent, differentiated by its parallelism and its independence, and its post-Meta survival makes it a safer bet than it looked in March.

10. Simular Agent S3

Our original guide cited Simular's Agent S2 at 34.5% on OSWorld as the open-source state of the art, and framed the whole category with "even the best agents fail two out of three times." Both statements are now badly obsolete, and correcting them changes the strategic picture. Agent S3, released October 2, 2025, scores 62.6% on OSWorld single-run and 69.9% with its Behavior Best-of-N (bBoN) method, which runs multiple rollouts and picks the best - Simular. Then, on December 16, 2025, Simular announced its agent had reached 72.6%, the first result to exceed OSWorld's human baseline of 72.36% - Simular.

Read that again, because it is the sentence that retires an era of agent skepticism: on the benchmark built specifically to measure real computer use across real desktop applications, an agent now outperforms the average human. Section 14 charts the full trajectory, but the headline belongs here: the "agents are a cute demo that fails most of the time" position, defensible in early 2025, is no longer supported by the data.

As a product, Agent S3 is a compositional open framework, not a consumer app. You deploy it with the models you choose, instrument it, and build on top; there is no installer, no subscription, and no support line. That gives it a strange profile in this ranking: the highest capability score in the table sitting next to the weakest reach and governance scores, because a research framework has no managed permission model, no enterprise controls, and no non-technical on-ramp. It is the raw engine of the human-baseline milestone, available to anyone willing to do the integration work.

Best for: engineering teams building their own computer-use products who want state-of-the-art open scaffolding, and researchers who need reproducible agent baselines. It is not a Cowork replacement for an individual knowledge worker, but it is the clearest demonstration that frontier computer-use capability is not locked inside proprietary vendors, which matters for anyone modeling where prices go next. Our computer-use benchmarks ranking tracks the full leaderboard as it moves.

11. Amazon Nova Act and Alexa+

When we first covered it, Amazon Nova Act was a research preview and we guessed at its trajectory. The guessing is over: Nova Act reached general availability on December 2, 2025, claiming over 90% task reliability at scale, built on the Nova 2 Lite model trained via reinforcement learning in synthetic "web gyms" - AWS. It comes with published pricing ($4.75 per agent-hour, each parallel agent billed separately), a Playground at nova.amazon.com/act, IDE extensions, and deployment through Bedrock AgentCore, currently limited to the US East (N. Virginia) region - AWS pricing.

Nova Act's positioning is deliberately narrow, and that narrowness is its argument. Amazon is not selling a personal coworker; it is selling reliable UI automation as an engineering primitive. The 90%-reliability framing targets the RPA replacement market: workflows that run thousands of times (order processing, portal data entry, QA flows) where a 60%-reliable general agent is useless but a 90%+-reliable scoped agent replaces brittle, selector-based RPA scripts. The per-agent-hour meter makes cost scale with actual execution, which finance teams can model, unlike a fleet of seat licenses.

The RPA-replacement math is the part worth taking to a spreadsheet. A traditional RPA deployment carries a license plus the ongoing engineering tax of selector maintenance: every UI change in the target application breaks the script, and someone bills hours to fix it. 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 an RPA bot. The crossover point runs the other way too: 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+, which became available to everyone in the US on February 4, 2026: free for roughly 180 million Prime members, $19.99/month for non-Prime, and now live in 10 countries, with Amazon reporting users have 2-3x more conversations than with old Alexa - TechCrunch. Alexa+ books, orders, and coordinates smart-home and shopping tasks conversationally. It is not a desktop work agent, but for household-operations automation it reaches more people than every other entry in this guide combined.

Best for: engineering teams on AWS automating high-volume, well-defined UI workflows, and organizations replacing legacy RPA with something maintainable. The trade-offs are equally clear: $4.75 per agent-hour compounds quickly for always-on use (a single agent running 8 hours a day costs roughly $1,140/month), the single-region availability constrains latency-sensitive and data-residency use cases, and nothing here helps an individual knowledge worker the way Cowork does.

12. xAI Grok and Grok Build

Our original guide contained a flat factual error, describing xAI as "Meta's independent AI initiative (led by Elon Musk)." xAI has never been affiliated with Meta, and the correction comes with a bigger structural update: xAI merged with SpaceX in February 2026, folding the AI lab into the aerospace company's balance sheet and compute ambitions. The agent-relevant products have also moved well past the chatbot era: Grok's multi-agent mode runs 4 or 16 parallel specialized agents on a task, and Grok Build, launched in beta on May 15, 2026, is a terminal-based agentic coding CLI exclusive to the $300/month SuperGrok Heavy tier - CIO Dive. Grok 5 is training on the 1-gigawatt Colossus 2 cluster.

The parallel-agent architecture is the genuinely interesting idea. Where Cowork runs one careful agent with checkpoints, Grok's heavy modes throw a committee of specialized agents at a problem and synthesize their outputs, trading cost and compute for breadth and speed. 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. Grok Build extends the same aggression to coding: terminal-native, autonomous, and positioned against Claude Code rather than against Cowork's general desktop work. Our Claude Code pricing comparison shows what that rivalry costs on each side.

The weaknesses are what the scoring table says they are. Price: $300/month for Grok Build access is the steepest entry fee in this guide, nearly double Claude Max 20x, with no mid-tier agent offering between the standard subscription and Heavy. Surface area: Grok lives in the X app, the standalone apps, and a developer CLI; there is no desktop coworker product, no OS integration, and no file-system agent. Governance: enterprise controls and audit tooling remain young, and the SpaceX merger adds organizational flux that procurement teams price as risk.

Best for: developers and researchers who want maximum parallel reasoning power and are already inside the X/xAI ecosystem, plus teams betting that Colossus 2-scale compute will translate into a capability lead worth paying early for. For mainstream desktop autonomy at sane prices, it is not yet a Cowork substitute, which is why it anchors this ranking rather than leading it.

13. The Agentic Browser Wave

The original version of this guide had no browser category at all, which in hindsight was its biggest structural omission: agentic browsers are where most consumer agent usage actually happens in 2026. The logic is first-principles simple. 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. Three products now define the category, and all three launched or transformed after our original publication.

ChatGPT Atlas (October 21, 2025, macOS, Chromium-based) put agent mode inside a first-party OpenAI browser for Plus, Pro, and Business subscribers, and OpenAI's March 2026 decision to merge Atlas with the desktop app and Codex signals where it is heading: one agentic desktop surface - Wikipedia. Perplexity Comet went free on all four platforms and pushed into artifact generation, as covered in section 8. Chrome Auto Browse (January 28, 2026) is the incumbency play: rather than asking users to switch browsers, Google gave the browser they already use autonomous task execution with human-approval gates on sensitive actions - TechCrunch.

How should you think about browsers versus Cowork-style desktop agents? The browser agent's ceiling is the web: it cannot reorganize local files, run local scripts, or operate desktop applications. Its floor, however, is much higher for everyday tasks: no setup, no folder permissions, and the agent sees exactly what you see. The decision rule that holds up in practice: 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 OS-level agent; if it should run without you present at all, you need a cloud or workforce agent.

There is also a security dimension unique to this category. Browser agents inherit your logged-in sessions, which makes them powerful and makes prompt injection via malicious page content the category's defining 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 (Chrome's human-approval pauses are the clearest example), but the honest state of play is that browser-agent security is a moving frontier, and 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: Agents Crossed the Human Baseline

The original version of this guide told readers that state-of-the-art agents scored around a third on OSWorld and "failed 2 out of 3 times," and it built its recommendations on that assumption. That was accurate when written and is now the single most misleading claim still circulating in older comparisons, so this section replaces it with the July 2026 data and what the data actually means for buying decisions.

OSWorld is the benchmark that matters for Cowork-class agents: hundreds of real tasks across real desktop applications on a real operating system, scored on outcomes. The progression tells the story better than any narrative. The 2024 baselines sat near 12%. OpenAI's Operator reached 32.6% in early 2025, and Simular's Agent S2 hit 34.5%, the numbers our original article cited as state of the art. Agent S3 then jumped to 62.6% single-run and 69.9% with Behavior Best-of-N in October 2025 - Simular. In December 2025, Simular's agent reached 72.6%, crossing OSWorld's measured human baseline of 72.36% for the first time - Simular.

The gap between the 62.6% single-run score and the 69.9% bBoN score carries its own lesson about where agent economics are heading. Behavior Best-of-N does not make the agent smarter; it spends more inference compute on the same task, running multiple full attempts and selecting the best behavior. That seven-point gain purchased with compute is the agent-world version of the test-time-scaling result that reshaped reasoning models: reliability is now partially a dial you pay to turn, not a fixed property of the model. For buyers, that reframes vendor pricing tiers. A "higher limits" tier is not just more tasks per month; on platforms that parallelize attempts, it is potentially a higher success rate on the tasks you already run, which is why comparing platforms on single-attempt anecdotes increasingly misleads.

The web-navigation picture moved just as far, and it moved toward small local models, which is the underappreciated result. Microsoft's Fara-7B, a 7-billion-parameter open model that runs quantized on a Copilot+ laptop, scores 73.5% on WebVoyager, ahead of OpenAI's computer-use-preview at 70.9% and UI-TARS-1.5-7B at 66.4% - Microsoft Research. A tiny on-device model beating a frontier lab's hosted computer-use model on a major benchmark means competent web agency no longer requires a datacenter, with everything that implies for privacy-sensitive automation and marginal cost.

Now the necessary honesty about what these numbers do and do not mean. Crossing the human baseline on OSWorld does not mean agents beat humans at your job; it means that on a fixed distribution of scoped desktop tasks, the best research systems now complete more of them than the average measured human. Production reliability on your workflows still depends on task specification, site weirdness, and error recovery, which is why Amazon leads with a "90%+ reliability on scoped workflows" pitch rather than a benchmark score, and why every serious platform keeps human checkpoints in the loop. The correct update from this data is directional, and it is large: the failure-rate objection that justified waiting in 2025 has expired, per-task economics now favor delegation for a growing class of work, and the differentiator among the tools in this guide has shifted from "can it do the task at all" to reliability, governance, and cost per completed task. Our computer-use benchmark rankings and model benchmark and pricing tracker follow these numbers as they move.

15. Enterprise Consolidation: watsonx, Artemis, and ServiceNow

The enterprise entries from our original list all still exist, but every one of them changed ownership, platform, or pricing since publication, and the pattern is uniform enough to state as a thesis: the enterprise agent platforms got acquired or re-platformed, because enterprises are not buying individual agents, they are buying control planes that manage fleets of them. If Cowork and its consumer rivals answer "who does my work," this section answers "who governs a thousand agents doing the company's work."

Moveworks is no longer a standalone alternative. ServiceNow completed the acquisition on December 15, 2025, positioning Moveworks as the "AI-native front door" for employees inside the ServiceNow platform - Moveworks. If you evaluated Moveworks in 2025 as an independent employee-support agent, the 2026 evaluation is really a ServiceNow platform decision, with everything that implies about contract structure and ecosystem lock.

Kore.ai re-platformed. The company took a strategic growth investment led by AllianceBernstein in January 2026, bringing total funding to roughly $620.9M as of April 2026, and launched its Artemis Agent Platform in May 2026, built around a declarative Agent Blueprint Language for defining agent behavior as versionable specifications - Kore.ai. The blueprint-language approach is the tell for where enterprise agent management is heading: agents defined like infrastructure-as-code, reviewed and audited like code, rather than configured through dashboards.

IBM watsonx Orchestrate got real pricing and a bigger ambition. Our original guide guessed it was in a "limited release or early adopter phase"; it now has public pricing starting at $530/month for the Essentials edition - IBM, and at IBM Think 2026 (May 5, 2026) IBM announced a next-generation version positioned explicitly as an agent control plane for orchestrating and governing thousands of agents across vendors, currently in private preview. That framing (govern everyone's agents, including competitors') is IBM's classic middleware playbook applied to the agent era.

What should a buyer do with this? Treat these platforms as a different layer of the stack, not as Cowork substitutes. 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 consolidation of 2025-2026 (ServiceNow buying, IBM re-platforming, Kore.ai raising to $620.9M) is the market pricing in that layered future. The practical advice: 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.

16. Verified Pricing in July 2026

Pricing in this market changes fast enough that our original article's numbers were stale within a quarter, so this section states current prices with sources and, more usefully, explains 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) cap your spend and meter you with usage limits. Consumption billing (Copilot Cowork's $0.01 per Copilot Credit, Nova Act's $4.75 per agent-hour) scales cost with execution, which is efficient for spiky workloads and dangerous for always-on ones. Credit systems (Manus) sit in between, flat-priced but internally metered, with the burn rate as the fine print.

Two readings of this data earn their place in a buying memo. First, the $20 price point won: Claude Cowork, ChatGPT agent mode, Manus Standard, Google AI Pro, and Perplexity Pro all cluster there, which means for an individual, capability and fit (not price) should drive the choice among the mainstream options. Second, consumption billing deserves arithmetic before commitment: Nova Act at $4.75/agent-hour is cheap for a 10-minute daily job ($24/month) and brutal for an always-on worker ($3,400/month per agent), while a flat-rate workforce platform or a Max-tier subscription holds its price no matter how many hours the agents run. Match the billing model to your duty cycle, not to the marketing page.

17. Safety, Governance, and the Export-Control Moment

A theme now exists that no January 2026 comparison could have covered: governments gate frontier agent releases. In June 2026, Claude Fable 5, the model under Cowork, was suspended under US export controls from June 12 to June 30, returning globally on July 1 - Anthropic. In the same window, OpenAI's GPT-5.6 family was previewed to about 20 trusted partners under a US government safety review, with public release following on July 10, 2026 - CNBC. Agent capability has joined the category of technologies whose availability is a policy variable, and a procurement process that ignores this is modeling 2024, not 2026.

The practical consequences for buyers are concrete. If your operations depend on one hosted frontier model, a three-week suspension is an outage you cannot engineer around after the fact; the mitigations are multi-vendor readiness (can your workflows fail over to another model family?), open-source fallbacks (OpenClaw, Agent S3, and Fara-7B run regardless of any vendor's regulatory posture), and contractual clarity about what your vendor owes you during a government-imposed pause. This is not alarmism; it is the observed behavior of the market in the past 30 days, and the June episode ended well precisely because Anthropic and regulators resolved it quickly.

The second governance frontier is agent containment at the OS and browser level, where the platforms have quietly converged on the same design philosophy: 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 - Windows Forum. Chrome's Auto Browse pauses for human approval on payments and other sensitive actions. Claude Cowork checkpoints multi-step plans for inspection. Workforce platforms log every agent session for review. The design consensus is identity-shaped: the agent is a junior employee with scoped permissions and a supervisor, not a superuser.

The two governance frontiers interact in a way that is easy to miss. OS-level containment protects you from your agent; export-control 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, which is what June demonstrated: no amount of workspace isolation helps when the model behind every workspace is paused. The mature posture treats 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 model like Fara-7B for the workflows that must not stop) rehearsed rather than theoretical.

Why this matters and how to apply it: governance has become the primary differentiator among tools whose raw capability is converging (section 14 showed capability converging above the human baseline). When you evaluate any platform in this guide, ask three questions in order. What can the agent touch, stated as an explicit permission boundary rather than a marketing sentence? 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 (Microsoft at the OS level, AWS at the IAM level, Anthropic at the plan level, workforce platforms at the per-agent level) are the ones treating agents as production infrastructure rather than demos.

18. Decision Framework: Choosing in July 2026

Eighteen months of shutdowns, mergers, and benchmark milestones compress into one primary question: does the work need you present, or not? That fork, more than any feature list, sorts the July 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. Everything else is a second-order refinement of where the work's inputs live.

Applying the framework to the ranking: if your work is your machine and your files, Claude Cowork at $20/month is now the default rather than the premium option, with Copilot Cowork taking its place wherever Microsoft 365 is the system of record. If your work is the web with you watching, ChatGPT Plus with Atlas is the strongest paid option and Comet is free, so trying the category costs nothing. If the work should run without you, match the shape: Gemini Spark for Google-native background tasks, Manus or ChatGPT Agent for deliverable-producing cloud runs, 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. Engineering teams building their own products start from Agent S3 or Fara-7B; sovereignty-first users start from OpenClaw.

Three findings from this refresh are worth restating as conclusions because they invert the advice we gave in the original version. First, the exclusivity premise died: Cowork is no longer the expensive gated option (it costs the same $20 as ChatGPT Plus), so "cheaper alternative to Cowork" is no longer a meaningful category; fit is. Second, capability stopped being the bottleneck: with agents past the OSWorld human baseline and a 7B model winning WebVoyager, the differentiators that matter now are governance, billing model, and where the agent lives. Third, resilience became a feature: after a blocked acquisition, a discontinued Google product, a shuttered OpenAI product, and an export-control suspension, the rational posture is one primary platform plus one fallback path you have actually tested.

This guide was researched and written by Yuma Heymans (@yumahey), founder of O-mega and co-founder of HeroHunt.ai, who has spent 2026 building and operating the kind of autonomous agent workforces this guide describes, including watching several of the platforms above run overnight jobs on real business operations. The comparisons here reflect that operating experience, not just vendor documentation.

This guide reflects the AI agent landscape as of July 8, 2026. Pricing, availability, and model versions in this market change monthly (sometimes weekly): verify current details on the linked official pages before purchasing.