AI agent platforms compared: pricing, memory and self-hosting
A flat monthly figure on this market buys access, not the work the agents do, and most of these 13 platforms keep a second meter running underneath. We priced and scored each plan below from what the vendor publishes, and named the metered layer wherever one exists.
What do AI agents cost to run each month?
Entry pricing here spans $10.59 to $59, and $29 a month is typical across the 13, verified July 29, 2026. The flat fee buys access; most of these meter the work too, in tasks, credits, executions or tool calls. No unit converts into another. Autonomous agents and scripted runners share the word, not the billing.
Paid plans here charge the account rather than the person, so headcount barely moves the bill anywhere except Flowise and Devin.
Free access is on offer at 13 vendors here, and on several it is a self-hosted build you stand up and maintain yourself.
Full disclosure is rare here: the count that leaves no tier blank is one out of 13, with thirteen publishing any figure at all. An unpriced enterprise or contract step is the norm on this market, not the exception.
Under most of these flat fees sits a second meter counting tasks, credits per module action, executions, tool calls, predictions, browser time or agents left running.
No exchange rate exists between those units, so two platforms priced the same are not buying you the same thing.
Flowise holds first place with 94. That figure comes from what a vendor publishes and what its users say, and capability does not enter it.
Ranked by a transparency score: pricing transparency 60%, user satisfaction 40%. Capability is not scored. It is the condition grid below. Prices are read from vendor pricing pages and re-checked per product on the dates shown. ComparEdge sells no AI agent platforms and takes no payment for placement. How the score is built.
01 / 06
AI agent platforms ranked: pricing transparency, memory and self-host
Vendors are placed by how openly they price, which is what our score measures. The columns mark autonomous running, persistent memory, multi-agent orchestration and self-hosting; a blank is unlisted, not unable.
Sorted by transparency scorePriced tiers 40 / 53Full disclosure 1 / 13How to read this table
Reading the console
Each row is one platform. Rank comes from our score for pricing openness, never from capability. The axis carries each vendor's cheapest paid tier logarithmically, so distance along it is real money. What it leaves out is the meter underneath: one vendor counts tasks, another a credit per module action, others count executions, tool calls, predictions, browser time or agents left switched on, and the counts do not translate. Read the mark as a door charge. A filled condition means the vendor states that capability in public. A blank means we found no record of it. That is a hole in what the vendor discloses, not a limit on the product.
Price axis
One shared logarithmic axis, $10.59 to $999. A tick further right is genuinely dearer. The tall tick is the cheapest paid tier.
Diamond
A flat platform fee, billed per month, not per seat.
Granted
Listed on the vendor’s own pages on the verify date.
Conditional
Present, but some tiers carry no published price.
Not on the record
Absent from the recorded feature set. It means unlisted, not incapable.
01FlowiseFree tierAll tiers pricedAutonomous run not on the recordPersistent memory listedMulti-agent listedSelf-host listed$35flat94Alternatives to Flowise
02MakeFree tier1 of 5 tiers unpricedAutonomous run not on the recordPersistent memory not on the recordMulti-agent not on the recordSelf-host not on the record$10.59flat83Alternatives to Make
03ComposioFree tier1 of 4 tiers unpricedAutonomous run not on the recordPersistent memory not on the recordMulti-agent not on the recordSelf-host not on the record$29flat83Alternatives to Composio
04n8nFree tier1 of 5 tiers unpricedAutonomous run not on the recordPersistent memory not on the recordMulti-agent not on the recordSelf-host listed$27.48flat82Alternatives to n8n
05OpenAI Agents SDKUsage-priced, public rate cardFree tier1 of 2 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent listedSelf-host listedSales only77Alternatives to OpenAI Agents SDK
06Browser UseFree tier1 of 5 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent not on the recordSelf-host listed$29flat76Alternatives to Browser Use
07Zapier AIFree tier1 of 4 tiers unpricedAutonomous run not on the recordPersistent memory not on the recordMulti-agent not on the recordSelf-host not on the record$29.99flat75Alternatives to Zapier AI
08Relevance AIFree tier1 of 4 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent listedSelf-host not on the record$29flat72Alternatives to Relevance AI
09DifyFree tier2 of 6 tiers unpricedAutonomous run not on the recordPersistent memory listedMulti-agent not on the recordSelf-host listed$59flat72Alternatives to Dify
10DevinFree tier1 of 5 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent not on the recordSelf-host not on the record$20flat66Alternatives to Devin
11LettaFree tier1 of 4 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent not on the recordSelf-host listed$20flat63Alternatives to Letta
12AutoGPTFree tier1 of 4 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent listedSelf-host listed$50flat60Alternatives to AutoGPT
13CrewAIFree to start, paid tier on requestFree tier1 of 2 tiers unpricedAutonomous run listedPersistent memory listedMulti-agent listedSelf-host listedFree52Alternatives to CrewAI
GrantedSome tiers sealedNot on the recordFlat account fee, not per seatScore ranks pricing transparency and user ratings, not capability. Capability is the grid.
A retrieval chatbot you would rather draw as a flow chart than assemble in code, and then host yourself. Flowise is the one platform here with no unpriced tier, and its plans are measured in predictions rather than people.
Critical gapComponent updates frequently cause breaking changes in existing flows.
Plan table and expert take
Flowise: expert take
Teams past five people are the ones who notice this ladder move: Pro includes five users and charges monthly for every one after that, which is unusual on a market that otherwise bills the account. Below that line the meter is predictions, not people.
Where Flowise holds up
Visually construct complex chains with 100+ LangChain/LlamaIndex nodes
Embeddable chat widget for any website with just a code snippet
Fully open-source core (Apache 2.0) allows for total customization
One-click deployment of chatflows as serverless API endpoints
Extensive marketplace of pre-built templates for quick starts
Branching, many-step scenarios across dozens of apps, mapped on a canvas instead of written out in code. Make is a workflow engine with AI steps in it, not an agent that decides for itself, and busy scenarios get hard to trace.
Plan table and expert take
Make: expert take
A credit is spent per module action, not per scenario run, so a long branching flow costs several times what a short one costs on the same trigger. Code execution is worse: two credits for every second it runs.
Where Make holds up
Visually debug workflows with real-time execution flow.
Advanced data handling: arrays, routers, aggregators, error handlers.
Granular control over scheduling, including specific dates and times.
Built-in AI module for text parsing, sentiment analysis, and generation.
Supports custom logic with Make Data Stores, a simple internal database.
Hundreds of third-party tools an agent has to reach, with delegated auth already handled for each one. Composio counts tool calls and moves you up through quota tiers; connected accounts themselves are never charged or capped.
Critical gapThe architecture requires custom bridging when integrating within complex legacy codebases.
Plan table and expert take
Composio: expert take
Buyers who care about support should read the fine print rather than the tier: Slack access opens on monthly spend, not on which plan you picked. The ladder itself is quota-based, counting tool calls, and connected accounts are never charged or capped.
Where Composio holds up
250+ pre-built integrations save huge development time
Handles OAuth and auth complexity
Works with all major AI frameworks
Good developer experience
★★★★★5.0CE scoreG2 4.9 · 7 reviewsFounded 2023Verified July 8, 2026
Automation that has to sit on your own servers, extended with nodes your engineers write themselves. The Community Edition carries no execution ceiling, and the paid tiers above it are built on concurrent executions rather than on people.
Critical gapSelf-hosting demands significant infrastructure maintenance and lacks direct technical vendor support.
Plan table and expert take
n8n: expert take
Teams under twenty people are the ones this ladder is built for: the business tier halves for them on application. And the shop window is priced in euros, so every dollar figure here is a conversion rather than the vendor's own number.
Where n8n holds up
Self-hostable for full data control and security
Source-available code allows for custom extensions
Build multi-step AI agents with memory and tools
Generous free tier for personal and small projects
Visual workflow editor with code node for complex logic
A notebook prototype that has to become a multi-agent service in production, on OpenAI models you already committed to. The SDK itself is free, the source public; the bill arrives from the API underneath, with a surcharge on the data-residency endpoints.
Plan table and expert take
OpenAI Agents SDK: expert take
The surcharge hides in the endpoint, not the plan: routing through the data-residency endpoints adds a tenth to the model rate for anything released after early March. The SDK itself is free and open source, so the entire bill is the API underneath it.
Where OpenAI Agents SDK holds up
The official OpenAI framework, so it tracks new models and platform features first
A built-in agent loop, handoffs, sessions, and guardrails, so you skip writing the orchestration yourself
Native tracing logs every tool call, which makes replaying and diagnosing a bad run straightforward
Free and MIT-licensed, with cost that is purely the model tokens and tools your agents consume
Realtime API integration for voice agents, plus async support for high-throughput batch work
Agents that have to click through real web pages rather than call a tidy API. Browser Use meters three layers over the subscription, browser time, proxy bandwidth and model tokens, and each layer carries its own rate.
Critical gapThe engine frequently fails during non-linear navigation tasks.
Plan table and expert take
Browser Use: expert take
Proxy bandwidth bought through the vendor costs twenty-five times what your own proxy costs, metered separately from browser time and model tokens. Bringing your own model key drops the orchestration markup from 1.2x to 0.2x.
Where Browser Use holds up
Dual-mode: Use locally via Python or via a managed Cloud API
LLM-agnostic: Integrates with any model (OpenAI, Anthropic, local)
Agentic tools for complex tasks like multi-tab coordination
Permissive Apache 2.0 license allows for commercial use
Active community with frequent releases and new integrations
★★★★★4.5CE scoreG2 4.4 · 3,294 reviewsFounded 2024Verified July 16, 2026
The first automation a small team builds once the spreadsheet stops coping. Zapier trades depth for reach, and every price on its page is a starting figure pinned to the lowest task volume, with overage billing on by default.
Critical gapThe platform requires extensive trial and error cycles to stabilize complex multi-step automation logic.
Plan table and expert take
Zapier AI: expert take
Every price on the page is a floor, captured with the task slider at its lowest notch. Overage billing is on by default, and email, SMS, WhatsApp and phone steps bill outside the task allowance entirely.
Where Zapier AI holds up
Build a multi-step automation from one natural-language prompt instead of assembling it by hand
Wires AI actions into an ecosystem of 6,000+ apps, which nothing else matches for breadth
Spin up custom AI chatbots grounded in your own knowledge bases, connected to your live app stack
A real free plan, not a countdown trial, so you can validate the fit before paying
Zapier Tables gives you a built-in database to hold state across complex AI workflows
Ready-made sales and support agents a go-to-market team can point at its own CRM within the week. Relevance meters three things at once on its upper tier: actions, vendor credits, and how many people are allowed to build.
Plan table and expert take
Relevance AI: expert take
One quota is really three on the Team tier: monthly actions, a separate pot of vendor credits, and a cap on how many people may build. Run out of any one of them and the tier is finished, whatever the other two still show.
Where Relevance AI holds up
Build capable agents in a visual, drag-and-drop canvas without writing code
300+ pre-built tools cover web scraping, lead enrichment, and the usual revenue-ops jobs out of the box
Deep native integrations with Salesforce, HubSpot, and Slack, so agents plug into the CRM you already run
A genuinely usable free tier lets you build and test real agents before paying
One-click deploy as an API, a web app, or a scheduled task, with bring-your-own-LLM on paid plans
A retrieval application that has to reach production and might have to live inside your own network. Dify's page carries a switch that swaps the entire plan set, so the cloud line and the self-hosted line are two different things to buy.
Critical gapThe platform requires self-hosting to ensure full data sovereignty.
Plan table and expert take
Dify: expert take
The toggle on the pricing page swaps the whole plan set rather than the billing period. Cloud and self-hosted are separate ladders: the cloud side carries an annual discount, the self-hosted side has none, and its premium step bills by consumption through a marketplace.
Where Dify holds up
Visual prompt orchestration with version control and rollback
Backend-as-a-Service (BaaS) for rapid API deployment
Built-in RAG engine supports multiple data sources out-of-the-box
Generous free self-hosted option with full feature access
Agent-based tools for creating autonomous, function-calling AI
★★★★★4.7CE scoreG2 4.6 · 20 reviewsFounded 2023Verified July 16, 2026
A well-scoped ticket you would rather hand to an agent than to a person, browser and terminal included. Devin's team tier arrives in two parts, a base charge plus a per-participant fee on top, and its quotas refresh on a daily and weekly clock.
Critical gapSlow performance reported during extended sessions.
Plan table and expert take
Devin: expert take
Teams is where this gets misread. The Teams plan starts at $80 a month and adds a charge for every full participant on top. Quotas refresh daily and weekly, and anything past them bills at API rates the pricing page never lists.
Agents that must remember a conversation from three weeks ago without being re-fed the transcript. Letta charges for an agent simply existing, month after month, and meters tool execution by the second, over and above the account fee.
Plan table and expert take
Letta: expert take
An agent costs money simply by existing here: there is a monthly charge per active agent sitting on top of the account fee, whether or not it does anything. Tool execution is metered by the second, and model tokens bill separately at provider rates.
Where Letta holds up
Best framework for persistent memory agents
Open source and transparent
Unique MemGPT architecture
Good for long-running agent tasks
★★★★☆4.0CE scoreG2 3.9 · 1,317 reviewsFounded 2023Verified July 8, 2026
The last two are the open-source pair with the thinnest published plans. AutoGPT prints two cloud prices and leaves its team tier blank; CrewAI prints one free tier and then quotes the rest.
An open sandbox for watching an autonomous agent reason its way through a goal, source code in your hands. The paid cloud adds a visual builder and a marketplace, but the subscription gates the chat and the agents draw on credits.
Plan table and expert take
AutoGPT: expert take
The subscription gates the chat and nothing else. Running agents and automations draws on a separate prepaid credit balance that refills itself, on every tier including Max, and that rate is never printed on the pricing page. Two invoices, not one.
Where AutoGPT holds up
Fully open-source and self-hostable, so you own the whole loop and your data never leaves your machine
Model-agnostic: point it at GPT-5.x, Claude Opus 4.8, or a local open-weight model, rather than one vendor
A large community plugin library extends what an agent can reach without writing the plumbing yourself
The clearest reference implementation of the goal-decompose-execute pattern, good for learning how agents work
The managed cloud build removes the setup tax if you want to try agents without standing up infrastructure
The stage before anyone signs anything, when a crew of role-playing agents still has to prove the idea works. CrewAI is free up to that point and contract-only past it, with no middle tier you can buy yourself.
Critical gapOverkill for many use cases; LangGraph offers wider support.
Plan table and expert take
CrewAI: expert take
Anyone hoping to grow out of the free tier gradually should look again. Basic allows fifty executions a month with no overage at all, and the next step is a negotiated contract carrying fifty development hours inside it. Nothing self-serve sits between them.
Where CrewAI holds up
Role-based agent design with distinct goals, tools, and backstories
Autonomous inter-agent delegation for dynamic task assignment
Flexible process flows: sequential, hierarchical, or consensual
LLM-agnostic: supports OpenAI, Azure, Anthropic, Gemini, and local models
Built-in memory and caching for stateful, efficient agent collaboration
AI agent questions: hidden fees, quota overruns, n8n and Make
Why is Zapier so expensive?
Because the figure on the page is a floor, not a price. Zapier scales by task volume, every published number is captured at the lowest notch of the slider, and overage billing is on unless you switch it off. None of that is peculiar to Zapier: volume sliders and automatic top-ups are how much of this market prices, and the best-known vendor is simply where buyers notice it first.
Is there anything better than Zapier?
Depends what you are optimising. Make maps branching logic more deeply and counts a credit for every module action instead of a task per run. n8n can be hosted by you, and its Community Edition carries no execution ceiling, the reason it keeps surfacing in automation forums. Zapier still reaches more apps than either, and for two-step work that reach is the argument.
Which is better, n8n or Make?
They meter differently, which usually settles it. Make counts credits per module action, so a scenario with many steps burns more than a short one at the same run count, and code execution costs two credits for every second it runs. n8n counts workflow executions and sells concurrency as the tiers rise. If the flows have to stay inside your own network, only n8n does that.
Do these platforms have hidden fees?
Nothing is hidden exactly. The meters are published, they are simply outside the plan price. Expect prepaid credit balances that top themselves up, top-ups once the bundled allowance is gone, proxy bandwidth, tool execution charged by the second, a fee for every agent left active, and overage priced at raw API rates. AutoGPT shows the shape best: the subscription gates chat, the agents draw on credits.
What happens when the included quota runs out?
Three behaviours, and they are worth checking before you sign up. CrewAI stops dead: its free plan allows fifty executions a month and carries no overage at all, so the only way past it is a contract. Zapier and Browser Use charge automatically instead. Dify pushes you onto a bigger plan rather than charging an overage.
Where does a workflow automation end and an AI agent begin?
On the bill, mostly. An agent decides its own next step toward a goal, so its cost tracks how hard the task proves to be. A workflow runner executes the exact path you drew, so its cost tracks how many steps you drew and how often it runs. One is priced by outcome, the other by structure.
Can I run agents for free?
Yes, at 13 of them. A free entry means two different things here. Hosted free tiers are real and tightly capped, usually a few flows, a few hundred actions or a handful of agents. Self-hosted community builds carry no vendor cap and hand you the costs instead: the server, the model keys, and everything the agent touches.
What will five people on one account pay?
About the same as one person pays. Almost every plan bills the account rather than the head, so the calculator above prices by team size against your monthly ceiling. Two exceptions. Flowise includes five users on its top tier and charges monthly for each one beyond that; Devin's team tier adds a per-participant charge over its base fee.
Is annual billing worth it here?
Where it exists, yes. Discounts land between roughly a seventh and a third below the monthly rate, with Zapier cutting deepest. Make is the exception worth reading twice: its annual option is not a percentage at all but a prepaid credit pack that stops expiring monthly and expires after twelve months instead. Several vendors, Devin and Flowise among them, publish no annual option.
Which of these actually run without a person watching?
Read the autonomous-run column in the grid, not the product page. The workflow platforms sit blank there by design, not by oversight: they run the path you wired and never pick one. The frameworks, the coding agent and the browser drivers record it as granted. Neither answer is better in the abstract, but buying an agent when you needed a deterministic runner is the common mistake here.
Can these prices be compared directly?
The subscription line compares. The meter under it does not. One vendor counts tasks, another a credit per module action, another executions, tool calls, predictions, seconds of browser time or agents kept alive. Nothing converts those into each other, so two platforms sharing a monthly figure can land on completely different totals once real work starts.
What do you still pay when you self-host?
The software is the free part. AutoGPT's self-host build needs your own model keys plus your own compute, sandbox, mail, CAPTCHA and antivirus. Flowise's free edition is limited by whatever machine you put it on. Dify's self-hosted line is a separate plan set with no annual option and a marketplace-billed premium step. Free means unmetered by the vendor, not unpaid.
Field note 01
A flat AI agent price is not a predictable one
Flat pricing is what this market advertises, and it is mostly true of the subscription line. It is rarely true of the invoice. Under most of the plans here sits a second meter that starts counting the moment an agent does anything: tasks, credits per module action, executions, tool calls, predictions, seconds of browser time, gigabytes through a proxy, or a standing charge for every agent left switched on. Each of those rates is published. None of them lives in the number the vendor leads with.
AutoGPT is the sharpest version of this. The subscription there gates the chat interface and nothing else, while running agents and automations draws on a prepaid credit balance that tops itself up, on every tier including the most expensive one. That is not a plan with a usage limit. It is two separate invoices, and the second has no published ceiling.
Field note 02
Before you pay for an AI agent platform, price the unit
Begin with the unit rather than the price. Find what the vendor counts, count how much of it one real workload consumes over a week, and only then look at which plan covers that. A monthly figure that looks cheap against a rival means nothing when one vendor counts a task per run and the other counts a credit per step inside that run.
Three switches are worth finding before you pay. Overage either charges automatically or stops the work dead, and both behaviours live on this page, failing in opposite directions. A free tier hosted by the vendor is a different offer from a community build on your own hardware, which moves the cost instead of removing it. The annual option deserves a look too, because here it can be a genuine discount or a prepaid pack with an expiry date.
The verdict on AI agent platformsSigned review · Updated
Oleh KemFounder & Lead AnalystComparEdge Editorial
Flat pricing is the promise on this market, and it holds for about a month. The subscription buys access. The work runs on a second counter underneath, and that counter is where a busy month goes wrong.
Find what the vendor counts, then run one real workload against it for a week before you commit. Tasks, credits per module action, executions, tool calls, agents left switched on: none of these convert into each other. The same monthly figure on two platforms can end the month in very different places. Check one more thing after that. Overage either bills silently or stops the work dead, both behaviours live here, and they fail in opposite directions. Disclosure is what the ranking measures, and the honest vendors here publish the second counter too.
MethodWe read 53 plans across thirteen platforms and recorded the metered layer wherever the vendor publishes one, last checked .
DisclosureCollection is tool-assisted; every verdict is written and signed by a human analyst.
05 / 06
Not sure which? Answer one, take a shortlist.Pick the line that sounds like your team. Each one opens the vendor built for it.
How this review is made. Prices are read from vendor pricing pages and re-checked on the dates shown against each product. Condition columns reflect the feature set recorded on the vendor’s own pages on that date. ComparEdge sells no AI agent platforms and takes no vendor payment for placement. Where a vendor publishes nothing, this page says so rather than estimating. Ranking is by transparency score: pricing transparency 60%, user satisfaction 40%. What a product can do is shown in the condition columns and carries no weight in the number.
Oleh Kem
