Chatbase alternatives, nine AI support widgets and what you pay

Chatbase is where most founders start, because everybody has heard of it.

Then you go and look at what else is out there, and the prices stop making sense. Chatbase sells message credits and Botpress sells conversations. ChatBot.com sells resolutions and counts a whole chat as one. Crisp sells you dollars of model usage and leaves the arithmetic to you. Put those four pricing pages side by side and there is no way to tell which is cheapest.

We read the nine pricing pages. For each tool you get the free plan, the first paid step, and what the meter is actually counting. Deacon is ours and it is in the table with the rest.

Every figure here was read from the vendor’s own pricing or docs pages on 17 September 2026, and each section links the page it came from. Where a vendor does not say something, we say that instead of guessing.

The nine, side by side

Chatbase

This is the one you are measuring everything else against, so it matters exactly where the free plan stops.

You get 50 message credits a month, one member and 1 MB of training content. Hobby is $40 a month for 700 credits, and more recharge at $40 per 1,000. Credits are priced by model, so the better the model you pick, the faster the tank empties.

Two things catch people out. Handing a conversation to a person is an AI Action, and the free plan has none of those, so on the free tier the bot answers and that is where it ends. Chatbase also deletes an agent after 14 days of inactivity on the free plan. That is easy to trip over while you are still making up your mind.

It trains on files, pasted text, a crawl, a sitemap, single links, question and answer pairs, Notion, and tickets from Zendesk or Salesforce. That is a wide list. On not knowing, Chatbase says the agent “answers from what you’ve given it, and says so when something isn’t covered”. Whether an answer names its source is not stated on the pages we read. Figures from chatbase.co/pricing.

Crisp

Crisp is not really a widget. It is a shared inbox with a widget on the front, and the AI is called Hugo.

That shapes the pricing. The free plan gives you the chat widget, two seats, and no AI credits at all, so the thing you came for does not run on it. Mini is $45 a month per workspace with $5 of AI credits, which Crisp puts at around 90 automated conversations. Credits are spent by tokens rather than by the message, and Crisp reckons a full Hugo conversation costs around $0.05 to $0.10.

Buy it if you want the inbox. Most of what Mini costs is the inbox. On not knowing, Crisp says “If the AI can’t find the answer, it’ll say it and your teams have just to follow-up”. Figures from crisp.chat/en/pricing.

Chatling

Chatling has the most generous free plan here, and it suits somebody who just wants to see whether any of this works.

200 AI credits a month on one seat, with 500,000 characters of knowledge. Starter is $25 a month, or $21 billed annually. Watch the meter though, because a reply costs anywhere between 0.5 and 25 credits depending on which model you point it at. Two hundred credits is either a lot of answers or a handful. Run out and the widget shows an insufficient credits error.

Handover is on every plan including the free one. Most of this group puts it behind a paid plan. It trains on a crawl of up to 1,000 pages, a sitemap, a list of URLs, PDFs, DOCX and TXT, pasted text, FAQs, Zoho, Zendesk and Notion, and there is a setting to show the sources behind an answer. Figures from chatling.ai/pricing.

Tidio

Tidio is a live chat tool first. Lyro is the AI agent sold next to it, and the two are metered separately, which is where the confusion starts.

A billable conversation is one a human on your team replied in. A Lyro conversation is one the AI replied in at least once, however many times it replies after that. So one customer with one problem can cost you on both meters.

The free plan is 50 conversations a month plus a one-off 50 Lyro conversations that never renew. Starter is $24.17 a month billed annually for 100 billable conversations, and Lyro is a separate add-on from $32.50 a month for 50 AI conversations, or $0.50 each. Tidio says nothing comes off your Lyro quota if Lyro hands over immediately, so a question it could not answer is not billed as one it could.

On not knowing, Tidio says Lyro forwards the question to a human agent. Figures from tidio.com/pricing.

Botpress

Botpress is a builder rather than a widget you paste in, and at $189 a month for the first paid plan it is not really chasing a small team’s money.

If you do look at it, the free plan is 25 conversations on three seats. Plus is $189 a month, or $150 billed annually, with 250 conversations and $25 of AI usage included. We read the Plus price again on 20 September 2026. It is the one vendor here that defines its unit without any wriggle room, and any exchange with at least two messages from the end user in the billing month counts as one conversation. More come in packs of 100 for $65, and model spend is metered on top of all that.

It imports websites, blogs, FAQs, PDFs, spreadsheets and documents up to 50 MB each, and says its agents cite their sources and escalate on rules you write. Figures from botpress.com/pricing.

Dante AI

Good free plan, and then a catch two rows down the pricing page.

100 messages a month on one seat, no card, 150,000 characters of knowledge. Starter is $40 a month, or $33 billed yearly, for up to 2,000 messages on two seats. A credit is one answer, and every model on the free plan costs exactly one, so the bill is easy to predict.

The catch is that the inbox a person takes over in, and lead capture, both start on Advanced at $120 a month. The two cheaper plans answer questions and do nothing else with them.

It trains on your website, price lists, PDFs, policies, opening hours, FAQs and anything you type in. Dante AI says an answer shows you the page it came from, and that when it does not know, it says so. Figures from dante-ai.com/pricing.

DocsBot AI

The one to look at if your documentation is scattered across a lot of different places. It lists more than thirty seven source types, including sitemaps, RSS, Office files, Notion, Google Drive, Zendesk Guide, helpdesk tickets and GitHub.

Free is one bot, 50 source pages and 100 AI credits a month. Personal is $49 a month, or $41 billed annually. A standard message starts at one credit and heavier work costs more.

Read this bit before you put it on a busy page. When the credits run out, the bot stops responding and shows visitors an error message. It does not apologise and it does not offer to take their email. DocsBot says every answer links to its source documents, and escalation to a support ticket starts on Standard. Figures from docsbot.ai/pricing.

SiteGPT

No free plan, and the trial asks for a card up front. Starter is $59 a month, or $39 billed annually, for 4,000 messages, 1,000 pages and one member.

SiteGPT’s own blog counts a visitor’s question and the AI’s reply together as one message, so 4,000 messages is 4,000 answers. That holds on its lightest model, and a stronger one costs ten times as much, so the same plan gives 400. Extra messages are $59 for another 5,000.

It trains on a website link, a sitemap, YouTube videos, most text formats, pasted text, and Notion, Google Drive, Dropbox, OneDrive, Box or GitHub. A visitor can move to a live agent with a button, and SiteGPT emails you a weekly digest of the questions it could not answer with one-click fixes. That last part is useful. Figures from sitegpt.ai/pricing.

ChatBot.com

This one is priced per user as well as per resolution, so it is the hardest here to line up against the others.

There is no free plan, just a fourteen-day trial that asks for no card. Essential is $25 a user a month, or $19 billed annually, and carries 10 AI resolutions a month. Another 50 cost $49.50.

A resolution, in their words, is a customer question the AI solved on its own, and several questions in one chat still count as one. That is a fair definition, and it makes ten a smaller number than it first looks. Your team can step in and take over at any point, with the visitor’s name, email, location and current page on screen. It trains on help articles, internal docs, URLs and PDFs up to 50 MB a file. Whether an answer names its source is not stated. We have a longer piece on this one in Deacon vs ChatBot. Figures from chatbot.com/pricing.

Deacon

Ours. It answers on your website from the documentation you already have, it installs with one line of HTML, and there is no inbox for anyone to sit in.

The free plan is 50 answers a month on one seat, with no card and no time limit. Nothing expires and no card is asked for. The AI answers on it, handing a conversation to you works on it, and nothing gets deleted if you go quiet for a fortnight. Above it there is one flat price a month with the team’s seats included, so nobody is charged per credit, per resolution or per model. The figures are on the pricing page.

Every answer is written from the passages Deacon found for that question, and when the answer came from a page on your site it links to it.

When your documentation does not cover something, Deacon says so, takes the visitor’s email and hands you the conversation. Every one of those questions is listed in your dashboard with a box beside it, and what you write there becomes the answer the next person gets.

There is more in the dashboard than the transcripts. It groups every conversation into topics, so you can see what people keep asking about, and it counts your visitors and sign-ups on the same screen without setting a cookie.

What to check before you pick one

Four things decide what you actually pay, and the headline price is only one of them.

Start with the unit. A message, a credit, a conversation and a resolution are four different amounts of work. Several charge more for a reply from a stronger model, ten times more on SiteGPT, and one charges you on two meters for a single customer.

Then work out what the free plan is really for. Some are a permanent small allowance you can build on. Others are a trial wearing a free plan’s clothes. Chatbase deletes an inactive agent after fourteen days, and Tidio’s AI quota is a one-off rather than something that comes back each month.

Check where handing over to a person starts, because on several of these it is not where the answering starts. On Chatbase it is an AI Action the free plan does not have. On Dante AI the inbox is two plans up. On Deacon it is on the free plan. If the reason you want a widget is to catch what you cannot answer, that line matters more than any message count.

Last, ask what happens when the meter runs out, because that is a real afternoon you will have one day. DocsBot stops answering and shows an error. Crisp stops the AI and sends new conversations to your operators. Chatbase and Chatling sell you more credits. Deacon keeps talking to the visitor, tells them it cannot look into it right now, and takes their email so the question still reaches you.