A general AI agent takes a task, works in a sandbox for minutes or hours, and hands back a finished file. Two questions matter before you give one real work: what a run costs, and whose servers hold your files while it runs. Credit billing inside a vendor's cloud makes both hard to answer.
Manus is where most people start, and it just raised money again. Its parent company, Butterfly Effect, announced more than $500 million on October 8, 2026, led by Boyu Capital and IDG Capital, months after Chinese regulators made Meta drop its $2 billion acquisition. I tried to price one recurring job on Manus and seven alternatives, and only one hosted vendor publishes enough numbers to do it. This is the condensed version. The full comparison on DevToolLab has every screenshot, the plan details and the GitHub script.
Three reasons teams are shopping around
Manus plans a task, runs it on a sandboxed virtual computer with internet access, and returns a report, website, slide deck or video. Version 2.0, announced September 28, 2026, brought the Cascade agent harness, Cloud Computers you can rent, and Automations that fire on events.
Ownership comes first. Meta announced the deal in December 2025, and Chinese authorities ordered it reversed in April 2026. On August 23 and 24, 2026 (Singapore time), after a two-week window to back things up, Manus deleted data some users had created since December 29, 2025, citing "regulatory requirements in specific jurisdictions." A US team's work was removed on a foreign regulator's say.
The meter comes second. Plans are $20 a month for 4,000 credits, $40 for 8,000 and $200 for 40,000, plus 300 daily credits that only work on Manus 2.0 Lite (pricing page, October 15, 2026). That works out to $5 per 1,000 credits, but Manus says a task's cost "depends on the complexity of the task" and won't estimate one before it starts. Its credits page lists three sample tasks at 200 to 900 credits, so the bill shows up after the work.
Third, model choice came late. Bring-your-own-key arrived with Manus Flex on September 29, 2026, and infrastructure still consumes credits under Flex.
What one task costs, where a vendor will say
The test job is a weekly competitive landscape report with pricing and feature comparisons, run four times a month. The prompt comes from Perplexity's credits page, because Perplexity is the only vendor here that lists typical credit ranges per task type.
| Vendor, entry paid plan | Price | Credits | Credits per $1 | Publishes cost per task |
|---|---|---|---|---|
| Manus | $20/month | 4,000 + 300/day (Lite only) | 200 | Three samples |
| Genspark Plus | $24.99/month | 10,000 | 400 | No |
| Perplexity Max | $200/month | 10,000 | 100 | Yes, by task type |
| Kortix Team | $40/seat/month | 2,500 (compute and managed models) | 62.5 | In hours, not tasks |
Perplexity rates a "Multi-step research report" at 350 to 950 credits (credits page last modified October 13, 2026). At 100 credits per dollar, one report costs $3.50 to $9.50, or $14 to $38 a month at four runs. Genspark sells the most credits per dollar and gives you no idea how many a task uses.
ChatGPT Work: already in your Plus plan
OpenAI launched ChatGPT Work on July 9, 2026, replacing the retired agent mode. It handles long multi-step jobs across your connected apps and files, produces documents, spreadsheets, decks and Sites, and Scheduled Tasks run it on a timer or a trigger.
It comes with Plus at $20 a month, and Plus gets GPT-6 Astra in Work. It runs in the cloud on web and mobile, and the desktop app can read local files once you allow it. OpenAI publishes no per-task allowance ("Work usage will vary by task"), longer sessions need Pro from $100 a month, and the Free and $8 Go plans get it only in the desktop app. It runs OpenAI models only.
Claude: Cowork now lives in the chat app
Anthropic merged its Cowork agent into the Claude chat app on September 16, 2026. A quick question and a multi-hour handoff now share one conversation: schedule the report, then come back to a document and slides.
There is no meter to read. Pro costs $20 a month ($17 billed annually) and Max starts at $100 for 5x or 20x Pro's usage, but neither tier says how much agent work it covers. Pro and Max got the merge first, with Team and Free to follow. On the plus side, your connectors, skills and projects already work for agent tasks, Claude asks before it acts by default, and Slides export to PowerPoint or PDF.
Perplexity Computer: the one with a price list
Perplexity Computer runs multi-step jobs in an isolated cloud sandbox, pulls context from connectors such as Gmail, GitHub and Slack, and works on schedules or triggers.
It is the only hosted agent here that quotes costs before you run anything: 15 to 70 credits for a quick fact, 2,400 to 9,800 for a data dashboard, and 15,000 or more for a full app build. Max accounts can cap monthly spend at $200 by default or up to $5,000, and tasks pause instead of failing when credits run out.
Volume is the expensive part. Max is $200 a month or $2,000 a year for 10,000 credits that don't roll over. Pro has no monthly credits, just a weekly allowance that is still rolling out (Light effort until you buy credits) and a one-time 4,000-credit bonus.
Genspark: cheap credits, thin cost data
Genspark is a hosted workspace built around a Super Agent, with AI Slides, Sheets, Docs and an AI Drive. Plus is $24.99 a month for 10,000 credits and 50 GB of storage, chat with core models is free on paid plans, and a weekly Free Quota covers Standard-mode agent work.
What it won't publish is a per-task cost. Its credits guide, updated October 14, 2026, does say you pay for failed agent retries and for output flagged after it is generated. Free accounts get 100 credits a day up to a lifetime limit. Pro is $249.99 a month and Team is $35 per seat.
Before you self-host: which repos are still alive
Star counts don't show whether anyone still merges code, so the original post includes a short Python script that asks the GitHub API for each self-hosted option's license, latest release and commit count over the past 30 days. Here is what it printed on October 15, 2026:
repo stars license latest release commits/30d
bytedance/deer-flow 83,882 MIT License v2.1.0 (2026-09-24) 873
FoundationAgents/OpenManus 58,661 MIT License v0.3.0 (2025-04-10) 0
kortix-ai/suna 20,276 Elastic License 2.0 v0.13.57 (2026-10-14) 2,909
DeerFlow released v2.1.0 on September 24, 2026, and Kortix shipped v0.13.57 the day before the run. OpenManus has gone quiet, with no commits to its default branch in 30 days and no release since v0.3.0 in April 2025. The script needs no token, and it's in the DevToolLab version of this comparison if you want to point it at other repos.
DeerFlow: the closest self-hosted match
DeerFlow is ByteDance's open-source "super agent harness," which coordinates sub-agents, memory and sandboxes through skills. If the term is new, an agent harness is the loop around the model that both DeerFlow and Manus Cascade implement. The 2.0 rewrite topped GitHub Trending on February 28, 2026.
It is MIT licensed and closer to Manus than either other self-hosted option. Any OpenAI-compatible endpoint works as the model, including OpenRouter or a local vLLM server on localhost:8000, so prompts and files can stay on your network.
Budget for the hardware. The README recommends 4 vCPU, 8 GB of RAM and 25 GB of SSD for Docker development, prefers Linux for servers, and warns that gateway admin access amounts to code execution. Its suggested models are Doubao, DeepSeek and Kimi, so expect to swap in your own.
OpenManus: a clone worth reading, not deploying
The MetaGPT team built OpenManus as a Manus clone with no invite code, and the first prototype took three hours. It is small enough to read in an afternoon: copy config/config.example.toml, set base_url to any OpenAI-compatible API, and run python main.py. Browser control goes through Browser Use CLI 3.0 as an MCP server, with no key in local mode.
That simplicity is also the ceiling. You type tasks into a terminal, the README documents no scheduler or multi-user setup, and the multi-agent run_flow.py mode is labeled unstable. Fork it as reference code rather than rolling it out to a team. MIT, 58.7k stars as of October 15, 2026.
Kortix: agents managed like code
Kortix, whose GitHub repo is still named suna, calls itself "the open-source AI Operating System." Agents, skills, memory and connectors sit in one git repository, each session gets a sandbox on its own branch, and changes merge through a change request that a person approves. Approval gates, per-tool rules and an audit trail make it the pick for teams that review agent output the way they review pull requests, and it takes your own API keys.
Read the license, not just the tagline. Elastic License 2.0 is not an OSI open-source license: you can self-host and modify Kortix, but you can't offer it to others as a managed service. Self-hosting is free, the cloud Free tier includes 200 credits, and managed Team seats cost $40 for 2,500 pooled credits, which Kortix equates to about 125 Agent Computer hours.
All eight side by side
| Tool | Where it runs | Entry paid price | How it meters | Self-host | License |
|---|---|---|---|---|---|
| Manus | Manus cloud | $20/month | Credits, three sample task costs | No | Proprietary |
| ChatGPT Work | OpenAI cloud, desktop app | $20/month (Plus) | Plan limits | No | Proprietary |
| Claude | Anthropic cloud, desktop app | $20/month (Pro) | Plan limits | No | Proprietary |
| Perplexity Computer | Perplexity cloud, or your NVIDIA GPU (Portable Computer) | $200/month (Max) | Credits, 100 per $1 | Portable Computer | Proprietary |
| Genspark | Genspark cloud | $24.99/month | Credits, per-task cost unpublished | No | Proprietary |
| DeerFlow | Your Docker host | $0 plus model tokens | Your provider bill | Yes | MIT |
| OpenManus | Your machine | $0 plus model tokens | Your provider bill | Yes | MIT |
| Kortix | Your infra or Kortix cloud | $0 self-host, $40/seat cloud | Sandbox credits plus your tokens | Yes | Elastic License 2.0 |
Picking one without migrating twice
- Start with what you already pay for. ChatGPT Plus and Claude Pro can both schedule the weekly report at no extra cost, so try it there first.
- Run one identical prompt on every trial, then read each meter. A real run tells you more than any credits-per-dollar table, including the one above.
-
Settle where the data may live. If a contract or your SOC 2 scope keeps data in your own cloud, only DeerFlow, OpenManus, Kortix and Perplexity's Portable Computer qualify. The shortest route is DeerFlow on an 8 vCPU, 16 GB VM in
us-east-1. - Price self-hosting in tokens. Multiply one finished run's input and output tokens by your provider's rates; the LLM Token Cost Calculator does the math. The loop is free and the model is not. To run the report on a schedule yourself, the Cron Expression Generator writes the expression for, say, every Monday at 8:00 AM ET.
- Check the license before you build on it. MIT lets you resell a hosted version. Elastic License 2.0 does not.
Which one fits
- Already on ChatGPT Plus or Claude Pro: use the agent you have. Both schedule the report at no added cost.
- You want a task's price in writing: Perplexity Computer, the only hosted option here with published cost ranges and a monthly spending cap.
- Data can't leave your network: DeerFlow, which is MIT licensed, runs on Docker and works with a local vLLM model.
- Your team reviews agent work like code: Kortix, if Elastic License 2.0 suits your plans.
- Staying on Manus: turn on Flex with an OpenRouter, Fireworks or Modal key, so a provider you already audit bills the inference.
Bottom line
This was the year agents moved into the chat apps: ChatGPT Work took over from agent mode in July, and Claude absorbed Cowork in September. Manus answered with 2.0, Flex and half a billion dollars in new funding. Before you renew any credit plan, run your most common task once and find out what it cost. If the vendor can't tell you, the meter is working for them, not you.
References
- Original article on DevToolLab: 7 Best Manus AI Alternatives in 2026
- TechCrunch: Manus raises over $500M in its first funding round since the split with Meta
- Manus
- ChatGPT Work and Codex, OpenAI Help Center
- Claude pricing
- What is Computer?, Perplexity Help Center
- Genspark
- DeerFlow on GitHub
- OpenManus on GitHub
- Kortix
Top comments (1)
The $14-$38 estimate for four Perplexity reports needs to sit beside the $200 Max subscription and credits that don't roll over. For someone buying the plan just for that weekly report, the cash outlay matters more than the credits consumed. I'd also compare time spent correcting pricing tables and checking sources across the identical-prompt trials: a cheaper run can become the expensive option if someone has to rebuild the report before sharing it.