The problem: You want to hand off an entire multi-step task — not one question — and get a finished deliverable back without babysitting each step.

Manus gives an AI agent a full virtual computer: a browser, a terminal, and a file system, then lets it plan and execute a broad goal like "research the top ten AI agent platforms, compare pricing, and build me a formatted spreadsheet" end to end, continuing to run in the cloud even after you close the tab. A multi-agent architecture delegates sub-tasks to specialized workers — a browser agent, a code-execution agent, a data-analysis agent — and you can watch a real-time replay of exactly what it did afterward, which is a genuinely useful transparency feature most competitors don't offer. Manus 1.6, released in early 2026, added a Wide Research mode that runs many sub-agents in parallel across sources for deeper investigation.

The company's biggest 2026 story isn't a feature — it's a deal that didn't happen. Meta reportedly agreed to acquire Manus for around $2 billion in late 2025, but Chinese regulators blocked the transaction in April 2026, leaving Manus independent and still shipping its own roadmap, including a new Web App Builder. Pricing runs on credits that several independent reviewers flag as genuinely confusing to predict in advance — a simple lookup might cost 10-20 credits, a complex multi-step research task can burn 200-500, and the interface doesn't always show the cost before you run it.

Pricing: Free (300 daily credits); Basic $19/mo; Plus $39/mo; Pro $199/mo.

The problem: The website you need to automate has no API, requires login and 2FA, and breaks your script every time it redesigns.

Skyvern takes a screenshot of a live webpage and uses a vision-capable LLM to identify what it's looking at and where to click, instead of hard-coded CSS selectors or XPath queries that snap the moment a site changes a class name. That approach scored 85.8% on the WebVoyager benchmark, the highest published result among web automation agents as of this writing, and it comes with built-in handling for CAPTCHAs, 2FA, and login flows that most competing tools leave you to solve yourself. It's Y Combinator-backed (S23), AGPL-3.0 open source with more than 22,000 GitHub stars, and self-hostable via Docker with your choice of LLM backend.

Worth flagging directly: Skyvern's own pricing page and its launch-week blog post disagree with each other about the free tier's exact credit allowance — one says 5,000 credits a month, the other says roughly 1,000 — a real, checkable inconsistency worth confirming directly before planning usage around a specific number. The company also moved away from simple per-step billing in January 2026 toward monthly credits, arguing the old model punished retries and validation steps.

Pricing: Free (self-hosted, open source); Hobby $29/mo; Pro $149/mo; Enterprise custom.

The problem: You're building your own product or agent and need browsing as a capability inside it, not a standalone app to open yourself.

Browser Use is an open-source Python library, not a destination app — it lets any LLM drive a real Playwright-controlled browser session, becoming the closest thing to a standard in this specific niche with more than 100,000 GitHub stars and a state-of-the-art 89.1% score on the WebVoyager benchmark, the highest of any tool with a published, developer-facing benchmark result. There's no GUI to speak of; you write code that defines what the agent does, which is exactly the point for a team embedding browsing into their own product rather than using one as an end-user tool.

The free tier (10 tasks/month, 3 concurrent sessions, no card required) is genuinely usable for real evaluation before committing to the $29/month Dev tier, which adds matching credits and stealth browsers for sites that actively block automated traffic.

Pricing: Free (10 tasks/mo); Dev $29/mo; Business $299/mo.

The problem: You don't just want browsing — you want the research it produces turned directly into a deck, a spreadsheet, or a finished document, without switching apps.

Genspark's Super Agent orchestrates nine underlying LLMs and more than 80 tools, one of which is an agentic AI Browser that navigates and acts on live pages as part of a larger chained task — a single prompt like "research the EV market and build me a 10-slide investor deck" runs research, a browsing pass, and slide generation without you opening a second tool. It's closer to a bundled workspace than a dedicated browsing agent specifically, which is exactly the appeal for someone who'd otherwise juggle three or four separate subscriptions.

The free plan includes 100 credits a day, enough to sample most of the toolset, with Plus at $24.99/month for 10,000 monthly credits. The credit system has drawn real user complaints about clarity — everyday chat and image generation are unlimited on paid plans, but heavier actions like video generation and long agent runs draw from the metered pool in ways that aren't always obvious upfront.

Pricing: Free (100 credits/day); Plus $24.99/mo; Pro $249.99/mo.

The problem: You just want your everyday browser to handle small research and task-completion work, and you don't want to pay anything for it.

Comet is Perplexity's Chromium-based browser with an agentic assistant built into the sidebar, and it's the most accessible tool in this entire roundup: it launched behind a $200/month paywall in July 2025, and Perplexity dropped that paywall entirely on March 18, 2026, making the full agentic experience free across Mac, Windows, iOS, and Android. It reads the current page, cross-references open tabs, and executes multi-step tasks — compare three project management tools, and it opens the tabs, reads each page, and returns a structured comparison.

Paid tiers exist mainly for volume and heavier background automation rather than gating the core experience: Pro at $20/month raises usage limits, and Max at $200/month adds a separate orchestration layer running many models and 400-plus app connectors in the background.

Pricing: Free (full agentic experience); Pro $20/mo; Max $200/mo.

The problem: You already pay for ChatGPT, and you want that same assistant to act on your behalf instead of just answering questions in a chat window.

Atlas is OpenAI's Chromium-based browser with an Agent Mode that takes a single instruction and executes it end to end — researching a topic, building a list, filling out a checkout form — while visibly showing each action as it happens. Its Browser Memories feature reuses ChatGPT's existing memory so stated preferences carry between sessions instead of starting fresh every time, a real continuity advantage for anyone already invested in ChatGPT's ecosystem.

Two real limitations are worth knowing before switching: Atlas remains macOS-only as of mid-2026 with no confirmed Windows date, and independent enterprise security testing found it blocked only 5.8-6% of malicious pages specifically designed to hijack its agent through prompt injection — a genuinely low number that reflects how unsolved this whole category's security problem still is, not that Atlas is uniquely careless. Agent Mode itself requires at least a Plus subscription; the browser shell alone is free.

Pricing: Free browsing; Agent Mode requires ChatGPT Plus ($20/mo)+.

The problem: Your task isn't really "browse a website" — it's "fix this GitHub issue," and that means writing code, running tests, and opening a real pull request, with browsing as just one supporting step.

OpenHands (formerly OpenDevin) is a free, open-source, model-agnostic autonomous software engineering platform running inside a sandboxed Docker environment with its own terminal, code editor, file system, and a real browser for researching documentation mid-task. Given a GitHub issue, it plans a solution, writes and tests code, and opens a pull request, iterating until the task completes or it gets stuck — genuinely broader autonomy than a browsing-only tool, at the cost of a heavier setup than a lightweight browser extension.

Three deployment paths exist: fully self-hosted and free under MIT license, a bring-your-own-key cloud tier capped at 10 daily conversations on the free plan, or custom Enterprise pricing with VPC deployment and SSO. The project has raised an $18.8 million Series A, a real funding signal in a category otherwise dominated by either bootstrapped open-source projects or venture-scale consumer browsers.

Pricing: Free (MIT, self-hosted); free cloud tier (BYOK); custom Enterprise.