The pitch for autonomous agents in 2024 was "give an AI a goal, it figures out the steps". The reality through most of 2025 was "give an AI a goal, it gets stuck after step three". Manus has spent the last year quietly closing that gap, and the May 2026 releases — 1.6 Max and Wide Research — are the first time I would call the agent experience genuinely productive for hard work.
I have been using Manus for the past few months. This is what I have learned.
_Manus runs on a per-credit model. If you are evaluating it, my referral link gives both of us bonus credits — useful when you are testing whether the platform suits your work._
Why Manus instead of OpenAI Operator or Project Mariner
| Spec | Architecture | Sub-agents? | App builder? | Pricing model |
|---|---|---|---|---|
| Manus 1.6 Max | Single + sub-agents | ✅ Wide Research | ✅ | Per-credit |
| OpenAI Operator | Single agent | ❌ | ❌ | ChatGPT Pro |
| Claude Agent SDK | Single + orchestrator | ✅ (SDK) | ❌ | API per-token |
| Google Project Mariner | Single agent | ❌ | ❌ | Bundled Ultra |
The current autonomous-agent space has four serious entries. They have very different architectures:
- OpenAI Operator — a single agent driving a browser. Good for short tasks. Ships as part of ChatGPT Pro.
- Claude Agent SDK — Anthropic's developer toolkit. You build the agent. Powerful but you write the orchestration.
- Google Project Mariner — single agent, embedded in Chrome. Browser-task focused.
- Manus 1.6 Max — single agent with built-in sub-agent collaboration. Plus an integrated app/website builder. Plus Stripe and SEO. Plus desktop access.
Manus is the only one of the four where "spin up an agent that builds a working app with payment integration" works end-to-end out of the box.
The 1.6 Max architecture
Manus 1.6 Max[4] is the architectural upgrade that landed in late April / early May. The key changes:
- Longer effective context. The agent maintains state across hour-plus runs without losing the thread. This was the single biggest weakness of earlier versions.
- Tool-call accuracy. The agent picks the right tool more often. Anecdotally, the rate of "agent picks Google when it should have used a database query" has gone from ~15% to ~3%.
- Web research depth. Deeper crawls, better source de-duplication, cleaner citations in output.
For the day-to-day "use Manus as your research analyst" workflow, these are quiet quality improvements that compound. Nothing flashy. The kind of thing you only notice when you go back to the old version.
Wide Research — the headline new capability
Wide Research[1] is the genuinely novel feature. Instead of one Manus agent working on a task, Wide Research spawns multiple general-purpose Manus instances that collaborate as sub-agents. Crucially — they are not constrained to predefined sub-roles. Each sub-agent is a full Manus.
What that means in practice:
- A "compare ten options" task previously had to be done sequentially or with a hand-built orchestration. Now you ask once, Wide Research spins up parallel agents, each one does the deep work on one option, and the main agent synthesises.
- The output quality on comparison tasks has gone from "good summary" to "I would have written this if I had a week".
- Token cost goes up (you are paying for N parallel agents), but wall-clock time drops dramatically.
The Manus pitch is that Wide Research is the autonomous-agent equivalent of moving from single-core to multi-core. That is salesy, but for the right kind of task — competitive analysis, literature review, code-base auditing — it is actually accurate.
Web App Builder + Stripe + SEO
The other under-discussed change is that Manus now builds full apps[4]. Web App Builder includes a database, Stripe payment integration, and SEO defaults. You ask in plain English; Manus generates the app, sets up the database, wires Stripe, deploys.
For prototyping or small-scale SaaS, this is the closest thing to "describe the product, get the product" that exists right now. Lovable, Cursor, and v0 do parts of this — but Manus is the only one that gives you Stripe and a deployed URL out of the box.
The output is real code (Next.js or similar), not a vendor-locked black box.
Platform scale
Source: manus.im / arxiv 2505.02024
Manus has now processed over 147 trillion tokens cumulatively and created over 80 million virtual computer instances[3]. To put that in perspective — that token volume is comparable to what OpenAI's API processed in its first two years. The platform is genuinely massive now.
The growth is also Meta-backed. Manus was acquired by Meta on December 29, 2025 for an estimated $2-3 billion[5] — Meta paid roughly four times Manus's $500m post-money valuation from April 2025. The acquisition has given the team capital and compute to ship faster, and the spring releases show it.
What this means for engineers
Three takeaways:
- For research-heavy work, switch to Wide Research. If your job involves comparing options, synthesising sources, or auditing things — Manus's parallel sub-agent architecture will save you hours per task.
- For "build the prototype" work, try the Web App Builder. Especially if your prototype includes payments or a real database. The Stripe + SEO defaults remove the boring 60% of prototype work.
- For agentic engineering work in your own code, stay on Claude Agent SDK. Manus is the right tool for self-contained tasks; Claude Agent SDK is the right tool when you need fine-grained control over what the agent can touch.
The verdict
Manus 1.6 Max + Wide Research is the most consequential agent release of May 2026 — more impactful than any single model upgrade. The Meta acquisition gave the team capital and compute to ship faster, and the May releases show it. If you have not tried Manus since their 2025 launch, the platform is meaningfully different now.
Try it via my referral link — we both get bonus credits — and run one Wide Research task on a topic you would normally spend a day researching. The "give it a goal, get an agent" promise is closer to true than it has ever been.