Updates for 5/16

Exited invite-only beta and is now publicly available with a free tier. New capabilities: slide creation, Web App Builder, desktop app with local file access, and Slack / WhatsApp / Telegram integrations.

Until last month, Manus required an invite code to access. The service had been quietly expanding its waitlist since around March, but for most people it was firmly in the 'heard of it, never touched it' category. Today that changes — a four-tier public plan with a free entry point is now live for anyone.

The shift from invite-only to public matters because real-world reviews and benchmarks will now flood in. Slack and WhatsApp integrations lower friction for embedding Manus into existing workflows. The catch: a single deep-research task can consume 900–1,000+ credits, which can blow through the free daily allowance (300 credits) in one shot. If you want to evaluate it seriously, starting at Pro Entry ($20/mo) will give you a clearer picture of actual usage.

Free: $0/mo (300 daily credits, 1 concurrent task) Pro Entry: $20/mo (4,000 credits) Pro Mid: $40/mo (8,000 credits) Pro Top: $200/mo (40,000 credits) Team: from $20/seat/mo (min 2 members) A single deep-research task can consume 900–1,000+ credits

With Manus behind an invite wall, there was no official pricing — just fragmented speculation online. The 'how much does it actually cost?' question had no clear answer. That changes today with the public launch and official plan announcement.

The free daily cap of 300 credits sounds reasonable until a deep-research task consumes 900–1,000+ credits in one go. For a few light tasks a week, the free tier works. For regular business research, Pro Entry ($20/mo, 4,000 credits) is the realistic starting point. One frustrating detail: credit cost is not shown before a task starts, so real spend remains hard to predict. That is worth knowing before committing to a plan.

The standalone Operator site has been fully deprecated; ChatGPT Agent is now the unified agentic experience combining web browsing, deep research, and file operations in a single interface.

Six months ago, Operator was the 'browse the web' service and ChatGPT was the 'chat with AI' service — two separate products with overlapping use cases. Switching between them was awkward, and it was not always clear which to reach for. The standalone Operator site is now gone.

Web browsing, deep research, and file operations now live under one roof in the ChatGPT interface. Tasks like 'research this topic and summarize it into a document' become single-session workflows. The tradeoff: Plus users are now on a 40 agent messages/month cap, so heavy usage will hit the ceiling quickly. If you were using Operator frequently, expect to rethink your plan tier.

Plus: $20/mo (40 agent messages/mo) Pro: $200/mo (400 agent messages/mo) Team: ~$30/user/mo Enterprise / Edu: rolling out soon

The old Operator had an informal 'included with Plus' feel, with no hard per-message limits prominently communicated. Now, with the ChatGPT Agent unification, explicit monthly caps are front and center.

40 agent messages on Plus means roughly 1–2 tasks per day before hitting the wall. Anyone running daily research workflows will need Pro ($200/mo, 400 messages) or to look elsewhere. Occasional users — a few searches a week — will be fine on Plus. Compared to Manus Pro Top ($200/mo, 40,000 credits), the structures are incompatible for direct comparison, but the pricing landscape now has enough concrete numbers to make real decisions.

Devin 2.0 added Interactive Planning (collaborative pre-task plan editing), Devin Wiki, and parallel session support.

The original Devin model was 'hand it a task, walk away, review the PR.' When the direction was off, you found out at the end — after the agent had already burned through credits running the wrong implementation. For large refactors or new features, this 'surprise at the finish line' pattern was a real friction point.

Interactive Planning lets you see Devin's proposed approach before it starts executing and correct course before any code is written. For expensive or high-stakes tasks, that is where it pays off most. Parallel sessions mean multiple tasks can run simultaneously, raising throughput noticeably. At $2 per ACU (roughly 15 minutes of autonomous work), fixing misalignment early rather than late is also a cost issue — not just a quality one. For small, low-risk tasks you would throw at it anyway, the new flow does not change much.

Workspace Agents added: connects to email and doc repos, can edit spreadsheets and fill forms within existing workspaces

Until last month, ChatGPT Agent (ex-Operator) was primarily a browser-navigation tool — good at fetching and reading web content, but unable to touch your actual work files or email inbox. Anything involving internal data still required manual steps afterward.

Connecting to email and document repositories turns 'research and summarize' into 'research, summarize, and update the spreadsheet' — all in one instruction. That last mile of writing changes back into your tools is where a lot of manual work was hiding. The 40-message monthly cap on Plus has not changed, so power users will exhaust this fast. Workspace setup also requires some initial configuration, so expect a bit of friction before it runs smoothly.

Rebranded as AGI, Inc. with AGI-0 as its flagship product — a native Android agent (private beta) that controls any app on your smartphone via voice or text. 30,000+ on the waitlist. Original browser-automation API continues for enterprise users.

MultiOn was primarily known for its browser automation API — a developer-facing service with limited consumer presence. 'Use it on my own phone' was not really the pitch. The rebrand to AGI, Inc. and launch of AGI-0 signals a hard pivot toward everyday smartphone automation as the core product.

Controlling Uber, DoorDash, or LinkedIn directly by voice — without switching apps — is the kind of interaction AGI-0 is aiming for. The 30,000+ waitlist suggests real demand. But it is still Android-only and in private beta, with iOS support unconfirmed. This is a 'watch this space' moment, not a 'add to your workflow this week' one. Enterprise users on the legacy browser automation API are unaffected.

2026 updates: DLAM (plug R1 into PC via USB for software-install-free computer control) and OpenClaw alpha (voice-driven open-source agent for local environments).

Rabbit R1 launched in 2024 with ambitious claims about LAM (Large Action Model) controlling smartphone apps — claims that largely failed to deliver. It pivoted to being an Android-based LLM assistant, but the 'why not just use your phone' criticism never fully went away. Expanding into PC control is the next attempt to carve out a distinct use case.

DLAM's USB-plug-in approach for PC control removes the software installation step — lower friction than most alternatives. OpenClaw, the open-source voice-driven local agent, is aimed at developers more than general users. Both are early-stage: DLAM is new and OpenClaw is alpha. Given Rabbit R1's credibility track record, these features deserve cautious optimism rather than excitement. Developers interested in local agent frameworks might find OpenClaw worth watching.