ToolJet vs Retool: we built the same apps with AI on both

Same prompts, same rules and, where both platforms offer it, the same model. ToolJet built our eight-page property manager for 21 cents with GPT-6 Luna and 79 cents with GPT-6.1 Sol, without a single approval stop, and both apps saved their data. Retool charged $7.10 on Retool Mini and at least $8.00 on Retool Power, which used up the last of the account’s free credits.

• Same model: on Claude Opus 5.5, available on both platforms, Retool cost 2.6 to 2.8 times as much as ToolJet for the same apps.
• Cost: for the entry and step-up models we tested (GPT-6 Luna and GPT-6.1 Sol against Retool Mini and Retool Power), ToolJet was 10 to 60 times cheaper per app and 16 to 28 times cheaper per follow-up edit. Even with GPT-6.1 Sol, an edit cost less than half of Retool Mini’s.
• Free plan: ToolJet’s free credits build four of our eight-page apps with GPT-6 Luna; Retool’s build three with Retool Mini, although they are worth 25 times more. In our test, Retool’s free $25 covered exactly two expense trackers, two eight-page apps and two edits, then Retool switched off AI app building until the next billing cycle.
• Friction: Retool stopped for approval before database changes, up to eight times in one build, and again the first time each new save action ran in preview; an expense typed before that approval was simply lost. ToolJet asked for no approvals in any of our tests.
• Reliability: in our tests, the apps in one Retool account shared a single Retool Database. In our test, building a second eight-page app on Retool changed the tables under the first, and four of the first app’s pages stopped loading. Both ToolJet eight-page apps kept their own tables and passed every save check.
• Models: ToolJet offers 19 models from five providers, including GPT-6.1 Sol, Claude Opus 5.5 and Gemini 3.6 Flash. Retool’s free plan runs only on Retool’s own two models, and its paid plans add GPT 5.6 and Claude models up to Opus 5.5; GPT-6 and Fable need your own key on Enterprise. The newer models are also cheaper to run: GPT-6.1 Sol costs half as much per token as GPT 5.6 Sol, and GPT-6 Luna costs half as much for input and 2.4 times less for output than GPT 5.6 Luna.
• Ownership: ToolJet’s apps open in a visual builder anyone on the team can edit, and ToolJet is open source with a free self-hosted edition. Retool’s new AI builder generates a React code project, and self-hosting is Enterprise only.
• Upkeep: The generated Retool app we inspected pins 56 declared npm dependencies. Fixes to those app-code dependencies require updating and republishing the app; this is separate from Retool platform patches. Standard ToolJet apps use the shared platform runtime, while custom or external libraries need their own maintenance.
• Your own coding agent: ToolJet’s MCP plugin lets Claude Code or Codex build ToolJet apps without spending ToolJet credits, on every plan including Free. On Retool’s free plan the only route is its MCP server, which still spends Retool credits; Retool’s credit-free CLI is paid-plan only.
• Your own data: on a live Databricks warehouse, ToolJet built a three-page analytics dashboard for $0.11 with GPT-6 Luna and $0.87 with GPT-6.1 Sol, with no approval stops; Retool Mini built one for $5.40 with five. Both showed the same live figures.

We gave both products the same prompts, word for word, and judged the app each one built without touching it by hand. Then we asked each for two changes, and tested whether the finished apps actually save data.

Every cost in this article is in US dollars: the value of the AI credits used, at each product’s paid top-up rate. Both products meter AI in credits, but a credit is worth ten times more on one side: ToolJet sells 100 credits for $1 ($0.01 each), and Retool sells 1,000 credits for $100 ($0.10 each). Our tests ran on free-plan credits, so these figures are what each build’s credits would cost to buy, not invoices or the total cost of owning an app.

The tests

1. Expense tracker (one page, with a design brief). Add expenses with a date, category, amount and note; filter by category and month; show the month’s total. The brief adds a header, a total card with the expense count, the form in its own card, coloured category tags, indigo (#4F46E5) as the primary colour, and set formats for money and dates. The full prompt is at the end of this article.
2. Property manager (eight pages). A property management app for a small landlord: dashboard, properties, units, tenants, leases, rent, maintenance and vendors, with seeded data, working forms and design rules. The full prompt is at the end of this article.
3. Two follow-up edits on the eight-page app: add a Reports page with a CSV export, then switch the theme to navy, drop cents from money and add a new column to the leases list.
4. Save checks on every finished app: add an expense in the tracker, and add a vendor and renew a lease in the eight-page app, each confirmed after a reload.
5. The same models on both platforms: both apps again on Claude Opus 5.5, and on GPT-6.1 Sol (ToolJet) against GPT 5.6 Sol (Retool), using a paid Retool Team account.

The models, two tiers on each side:

Tier	ToolJet	Retool
Entry	GPT-6 Luna, ToolJet’s cheapest model	Retool Mini, Retool’s default
Step up	GPT-6.1 Sol	Retool Power

Stronger models build better apps on both platforms. GPT-6 Luna is the cheapest option in ToolJet’s picker, and GPT-6.1 Sol or Claude Opus 5.5 are the picks when quality matters more than cost.

The rules. One run per prompt and model, no hand edits. When an agent stopped to ask for approval, we approved and counted the stop. Time is wall clock from sending the prompt to a finished app, including any approval stops; our test harness answered each stop within a few seconds, so waiting adds well under a minute to any build. Retool’s tests ran on a new account, one build after another, the way a new customer would use it; ToolJet’s ran in a new workspace. This is a same-prompt test, not a benchmark: one run per model is a fair first impression, not a statistical result, and Retool’s build times in particular varied widely from run to run.

Who ran this. ToolJet ran this test and publishes it here. We report where Retool did better, and every prompt is at the end of the article so anyone can repeat it.

Versions. Retool’s AI app builder (beta) on its cloud Free plan, and ToolJet AI on the release that adds the page compiler. Tested on 1, 2 and 4 October 2026; the same-model builds and the CLI test ran on a paid Retool Team account.

Both free plans are enough to try AI app generation for real; they differ most in who can use the apps and how far the AI credits go.

	ToolJet Free	Retool Free
Price	$0	$0
Builders	2	Up to 5 users in total
End users	Up to 50	Included in the 5
Apps	Up to 10	Unlimited web apps
AI credits per month	100 (worth $1)	250, shared across the account (worth $25 at Retool’s pack price)
Eight-page apps the free credits build	4 with GPT-6 Luna ($0.21 each)	3 with Retool Mini ($7.10 each)
AI app generation	No page limit	No page limit stated
Models on the free plan	GPT-6 Luna, DeepSeek V4.1 Flash and others in the picker	Retool Mini, Retool Power
Extra AI credits	Not on Free; $1 per 100 credits on paid plans	Not on Free; $100 per 1,000 credits on Team and Business
Built-in database	10 tables, 500 rows	Up to 5 GB
Environments	Production	Production only
Build with your own coding agent	Yes: ToolJet MCP plugin for Claude Code, Codex or any MCP client; no ToolJet credits used	MCP server only (beta), which spends Retool credits; the credit-free Retool CLI is paid-plan only
Self-hosting	Free open-source edition	Enterprise plan only

In dollars, Retool’s free plan includes more AI usage ($25 against $1). Counted in apps, ToolJet’s goes further: its 100 free credits build four of our eight-page apps with GPT-6 Luna, while Retool’s 250 build three with Retool Mini. Retool says a prompt costs about 12 credits ($1.20) and an app takes 3 to 6 prompts; the results below show what each build actually cost.

Sources: Retool pricing, ToolJet pricing, both read on 1 October 2026.

ToolJet offers 19 models from five providers, including the newest from each; Retool’s free plan offers its own two models, and its newest third-party models need your own key on Enterprise.

Provider	ToolJet AI	Retool (managed key, Team and up)	Retool (own key, Enterprise only)
OpenAI	GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Terra, GPT-5.6 Sol, GPT-5.6 Luna	GPT 5.5, GPT 5.6 Luna, GPT 5.6 Sol, GPT 5.6 Terra	GPT 6 Astra
Anthropic	Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Fable 5, Opus 5, Sonnet 5, Haiku 4.5	Claude Opus 5.5, Opus 5, Sonnet 5, Opus 4.8	Claude Fable 5; Claude Opus 5.5 and Sonnet 5 through Amazon Bedrock
Google	Gemini 3.6 Flash		Gemini 3.1 Pro (Preview), Gemini 3.5 Flash
xAI	Grok 4.7, Grok 4.6, Grok 4.5
DeepSeek	DeepSeek V4.1 Flash (US-hosted, zero data retention)
Own models		Retool Mini and Retool Power (Free and Team plans, beta)

ToolJet’s picker shows each model’s typical build time and credit cost before you send a prompt, so the trade-off is visible up front. GPT-6 Luna is the cheapest model in the list, at about 5 to 15 credits for a 1 to 5 page app. Stronger models such as GPT-6.1 Sol and Claude Opus 5.5 build better apps and cost more credits per build.

Newer models matter for the bill, not just the release notes. Each new generation in this list costs less per token than the one Retool offers in its place:

On ToolJet	Newest equivalent on Retool	Price per million tokens, input / output	Difference
GPT-6.1 Sol	GPT 5.6 Sol	$2 / $10 against $4 / $20	Half the price; cached input a quarter
GPT-6 Luna	GPT 5.6 Luna	$0.10 / $0.50 against $0.20 / $1.20	Half the input price; output 2.4 times cheaper

Agentic app building re-reads the app and its context at each step, so input and cached-input prices can matter as much to a build’s cost as output prices. We did not measure the token mix of these runs.

ToolJet AI’s model picker shows each model’s typical build time and credits before you send a prompt.

On Retool’s free plan, app building runs only on Retool’s own models. Retool’s docs describe Retool Mini and Retool Power as “open-source models that Retool provides for building apps on the Free plan and Team plan”, in public beta, and since 28 September 2026 other providers’ models are paid-plan only. Retool’s managed models on Team and up now include Claude Opus 5.5; GPT-6 Astra and Claude Fable 5 still need your own key on Enterprise.

Sources: ToolJet AI model picker (screenshot above), Retool supported AI models (read on 4 October 2026), Retool changelog: changes to available AI models on the Free plan (read on 1 October 2026).

Every build finished with the pages we asked for. The differences were in what each one cost, how often it stopped for approval, and whether the finished app could actually save data.

Build	Time	Cost	Approval stops	Add-expense check
ToolJet, GPT-6 Luna	4m 18s	$0.03	0	Passed
ToolJet, GPT-6.1 Sol	2m 25s	$0.22	0	Passed
Retool Mini	2m 8s	$1.80	2	Passed after approving the new save code; the first expense was lost
Retool Power	32m 47s	$4.30	3	Passed after approving the new save code

With a design brief, all four trackers came out close to it: indigo buttons, a total card with the count beside it, coloured category tags and the requested money and date formats, and none opened with errors. The cost did not converge: $0.03 on ToolJet with GPT-6 Luna against $1.80 on Retool Mini, 60 times as much, and $0.22 with GPT-6.1 Sol against $4.30 on Retool Power, about 20 times as much. Retool Power’s build also took 33 minutes; along the way its first attempt to insert sample data failed and its sandbox restarted.

The expense tracker ToolJet built with GPT-6.1 Sol from the design brief, in 2m 25s for $0.22.

Build	Time	Cost	Approval stops	Pages	Save checks
ToolJet, GPT-6 Luna	17m 24s	$0.21	0	8 of 8	Passed
ToolJet, GPT-6.1 Sol	11m 11s	$0.79	0	8 of 8	Passed
Retool Mini	About 56 min	$7.10	8	8 of 8	Add vendor passed after approving the new save code; four pages later broke (see below)
Retool Power	Under 55 min	At least $8.00, all the credits left	Not captured	8 of 8	Failed: add vendor and renew both returned errors

All four builds delivered the eight pages and the seeded data. On ToolJet that cost 21 cents with GPT-6 Luna and 79 cents with GPT-6.1 Sol, and both apps passed every save check. On Retool, one build cost $7.10 to more than $8.00, close to a third of the free plan’s whole month of AI credits each, and Retool Mini’s build took close to an hour, with eight approval stops along the way. Retool Power’s build finished in under 55 minutes; we did not capture its exact time or approval stops.

The dashboard ToolJet built with GPT-6 Luna, for $0.21: occupancy, rent collected against expected, overdue balance, open maintenance, both charts and leases ending in 60 days.

The rent ledger ToolJet built with GPT-6.1 Sol, with monthly, payment history and overdue tabs and a Record payment action.

Retool Power’s build was the one that emptied the account. When it finished, Retool showed: “Your Retool organization has reached its AI credit limit. AI-powered app building is now unavailable until your credits reset at the start of your next billing cycle.” Its app looked finished, but neither save worked: adding a vendor failed with “Cannot read properties of undefined (reading ‘trim’)”, and renewing a lease returned an error and left the end date unchanged.

What the stops look like. During a build, Retool’s agent pauses before database changes such as creating a table, shows the SQL with Deny and Run buttons, and waits; left alone, our first build sat at “Awaiting your review” for 23 minutes. After the build, the first time each new save action runs in the preview, Retool pauses again: “createVendor contains new data access code that needs approval”. The stops are a safeguard, so a person reviews database changes before they run; the cost is that a build cannot finish while nobody is watching. ToolJet’s builds ran start to finish without a stop.

Saving a vendor in Retool’s preview: the app pauses until the builder clicks Approve & run, under a banner saying the account has reached its AI credit limit.

Retool’s Team plan includes Claude Opus 5.5, which ToolJet also offers, so on a Team account we paid for we ran both apps again on the same model. We also paired Retool’s GPT 5.6 Sol with ToolJet’s GPT-6.1 Sol, the newer model from the same family.

Build	ToolJet	Retool
Identical model: Claude Opus 5.5
Expense tracker	1m 40s, $0.72, no stops; saved	3m 44s, $2.00; saved after approving the new save code
Property manager	15m 39s, $3.23, no stops; both save checks passed	13m 27s, counted from our answer to its database question; $8.30; 1 question and 1 approval during the build; both save checks passed after approving each new save action
Same family, different generation: GPT-6.1 Sol (ToolJet) vs GPT 5.6 Sol (Retool)
Expense tracker	2m 25s, $0.22, no stops; saved	2m 1s, $1.30, 1 question; saved after approving the new save code
Property manager	11m 11s, $0.79, no stops; both save checks passed	3m 19s, $1.70; data hard-coded in the front end, so the new vendor was gone after a reload and renewing failed

On the same model, Retool cost 2.6 to 2.8 times as much for the same apps. Retool’s Opus build was careful: it asked which database to use, put its tables under a pm_ prefix so they would not collide with other apps, and both saves worked. (Retool’s time is measured from our answer to that question.) Its GPT 5.6 Sol build finished in three minutes because it skipped the backend: the eight pages read hard-coded sample arrays held in browser memory, even though Retool reported “seeded sample data and interactive management workflows”.

Every screenshot below is the app exactly as the AI left it, with no hand edits, labelled with the model, what the build cost and how long it took.

Expense tracker

Eight-page property manager

Every app here looks finished, and Retool’s are polished. What a screenshot cannot show is whether the app works: two of the Retool builds pictured, Retool Power’s and GPT 5.6 Sol’s, could not keep a new record, while every ToolJet app passed its save checks.

A first build is the cheap part of an internal tool; the changes that follow are where the cost adds up. We ran the two edits on three of the eight-page builds: ToolJet’s GPT-6 Luna and GPT-6.1 Sol apps and Retool Mini’s. ToolJet made each change for under ten cents with GPT-6 Luna, and for about 60 cents with GPT-6.1 Sol, still less than half of what Retool Mini charged. Retool’s edits were faster.

Edit	Platform	Time	Cost	Approval stops
Add a Reports page with a chart, a table and CSV export	ToolJet, GPT-6 Luna	10m 26s	$0.08	0
	ToolJet, GPT-6.1 Sol	8m 1s	$0.65	0
	Retool Mini	1m 10s	$2.30	2
Navy theme, money without cents, a new sorted column	ToolJet, GPT-6 Luna	11m 13s	$0.09	0
	ToolJet, GPT-6.1 Sol	7m 25s	$0.61	0
	Retool Mini	44s	$1.50	0

The Reports edit also asked for costs on the requests already marked done. Retool Mini wrote them in directly. Neither ToolJet model changed existing records during the edit: GPT-6 Luna reported spend from each request’s existing cost estimate and made completed requests editable, and GPT-6.1 Sol added a one-click “Fill completed sample costs” button on the Reports page, so its spend reads $0 until someone clicks it.

At Retool’s prices, ten such edits a month cost about $19 for one app. On ToolJet with GPT-6 Luna they cost under a dollar.

What you get to edit. ToolJet’s AI builds into the same visual builder people use by hand: every page, component, query and event it creates can be selected and changed in the editor, by anyone on the team, without code. Retool’s new AI builder, the one we tested, generates a Vite, React and TypeScript project. You change it by prompting again, which spends credits, or by editing the code yourself in the Code tab.

What about Retool’s classic builder? Retool’s drag-and-drop classic builder is still available, but it is no longer where Retool’s AI lives. Retool removed its AI assistant, Assist, from the classic builder for most customers on 30 September 2026, with certain customers keeping access until 31 December 2026 to migrate, and recommends the new builder for AI-powered building. Retool has not announced a retirement date for the classic builder; its staff have said classic apps will be supported “for at least two years after the new experience reaches GA”. For most customers, AI-assisted building on Retool now uses the code-generating builder we tested; certain customers retain classic Assist until 31 December 2026. Sources: Retool changelog: Assist removal, Retool changelog: upcoming conclusion of public beta for Assist, Retool community: the new Retool app builder is here.

Where your data lives. In our tests, every app in a Retool account used the same Retool Database, and the AI treats it as its own. In our test, the second eight-page app, built with Retool Power, rebuilt the shared tables to suit itself. Afterwards the first app, built with Retool Mini, failed on at least four of its pages, the dashboard included, with errors such as column “rent” does not exist, although nobody had touched it. On ToolJet, the second eight-page app created its own set of tables beside the first, and both apps kept working.

It does not have to happen. Retool’s Claude Opus 5.5 build on the paid account asked which database to use and put its tables under a pm_ prefix, so it collided with nothing.

Retool Mini’s finished property app after Retool Power built a second app in the same account: the dashboard now fails with column “rent” does not exist.

Retool’s new builder writes a React and TypeScript codebase for every app. ToolJet’s AI fills in the same visual spec ToolJet’s builder has run since 2021: pages, components, queries and events. The two approaches suit different people.

What each produces. Retool’s Claude Opus 5.5 build of the property manager came to 43 files and 3,159 new lines of code, on top of a project scaffold of about 7,700 lines, roughly 2,400 of them Retool’s own runtime code for sign-in, backend calls and the preview, with 56 npm packages declared. ToolJet’s Claude Opus 5.5 build of the same app is 8 pages, 109 components, 17 queries and 109 events, each one visible and editable in the builder, with no code files and no packages.

Same model, different bill. On Claude Opus 5.5, the expense tracker cost $2.00 on Retool against $0.72 on ToolJet, and the property manager $8.30 against $3.23. The model was identical; what differed was everything around it, from each platform’s agent and instructions to what it produces. Our test measures the gap; it cannot say how much of it comes from writing a codebase rather than filling in a spec.

Who can check the result. Retool pauses new data-access code until a person approves it, and runs server-side checks on every push, which are sensible safeguards. But approving SQL and TypeScript takes someone who can read them. In our test, two Retool builds looked finished and were not: Retool Power’s app on the free plan could not save, and the GPT 5.6 Sol build kept its data in hard-coded front-end arrays, so nothing survived a reload. A non-developer sees a polished screen either way. In ToolJet, every query and the table it writes to are visible in the builder.

What has to be maintained. The generated Retool apps we inspected carry application code and dependency lists. The package.json of Retool’s Claude Opus 5.5 build pins exact versions, React at 19.2.7 and Vite at 8.2.2 among 56 declared packages, and our CLI-built tracker installed 291 npm packages in all. Fixes to these app-code dependencies require updating the affected versions and republishing the app. That is separate from Retool’s platform and runtime patching; the package count alone does not establish how much maintenance an app will need. Standard ToolJet apps are definitions that use the shared platform’s components and runtime, so platform fixes reach those apps through ToolJet Cloud updates or a self-hosted deployment upgrade. Any custom code or external libraries a team adds still need their own review and maintenance.

Who each approach fits. Code generation fits developers who want React as the source of truth, Git branches and full control. A visual spec fits the operations, finance and support teams who live in internal tools, and the developers who support them, because they can inspect, adjust and govern an app without reading a codebase.

Many teams already pay for Claude Code, Codex or Cursor. Both platforms let those agents build apps, but on Retool’s free plan the only way in still spends Retool credits.

	ToolJet MCP plugin	Retool MCP server	Retool CLI
Who builds the app	Your agent, directly through ToolJet’s APIs	Retool’s own building agent; your agent passes the prompt along	Your agent, writing React code on your machine
Uses the platform’s AI credits	No	Yes	No
Plans	Every plan, including Free on ToolJet Cloud and the free self-hosted edition	All plans, public beta, Retool Cloud	Paid plans only, private beta, Retool Cloud
Setup	Install the plugin (two commands), create an access token, set two environment variables	Add Retool’s MCP URL to your agent and sign in	Node 22, pnpm and git; install the CLI, sign in, install its skill; then init, install, start, push and publish for each app
What you end up with	A normal ToolJet app in the visual builder	A React project in Retool	A React project you maintain and push

Not on Retool’s free plan. Building with your own coding agent without spending Retool credits needs the Retool CLI, and the CLI requires a paid Retool plan. On Retool Free, the only route is the MCP server, where Retool’s building agent spends your AI credits on every build. ToolJet’s MCP plugin works on every ToolJet plan, Free included, and never spends ToolJet credits.

Retool’s MCP server spends credits by design. Retool’s FAQ describes building through your own agent as using two AI services at once: your agent’s tokens, and Retool credits for its building agent. In its words, “Retool AI credits apply to Retool’s app building agent, and there’s no setting to turn that off.” On the free plan, that is the same $25 allowance our eight-page apps used up.

Retool’s credit-free CLI is paid-plan only, and we tested both sides of that. Retool’s docs say “Retool CLI commands do not incur AI credit spend.” On our free account the CLI (@tryretool/cli) signed in and was then refused: “The Retool CLI is not available on the Free plan. Upgrade to a paid Retool plan to use it.” On a Team account we paid for, it worked. Claude Code, running Sonnet 5.5, built the same design-brief expense tracker through it, and Retool’s credit balance stayed at 1,000 of 1,000. Setup took about two minutes (sign-in, retool init, pnpm install), writing the app took Claude Code under two minutes, and retool push built a preview in 22 seconds. The app passed our add-expense check after one “Approve & run” in Retool’s editor.

Rough edges we hit with the CLI. Twice its sign-in stopped working within minutes (“Invalid access token”) while retool auth status still reported us signed in, so we had to sign in again. It has no command to create a table, so the expenses table was made with raw SQL through retool resource explore --allow-mutative. Retool lists Node 22, pnpm and git as prerequisites, and the CLI is still a private beta.

ToolJet’s plugin builds with your agent’s own subscription. Install it in Claude Code with /plugin marketplace add ToolJet/tooljet-mcp and /plugin install tooljet-app-builder@tooljet (Codex has an equivalent), then set TOOLJET_DEPLOYMENT_URL and a personal access token. Your agent creates the tables, pages, queries and events itself, so no ToolJet AI credits are spent, and the result opens in the same visual builder as any other ToolJet app.

We tried it. Claude Code, running Sonnet 5.5, built the design-brief expense tracker through the plugin in 4.9 minutes, with 23 calls to ToolJet’s tools and no ToolJet credits. The app passed the same add-expense check as the other trackers. Claude’s own usage is billed by Anthropic, as it is for any coding agent.

Sources: Retool: use MCP to build apps, Retool AI credits FAQ, Retool CLI, Retool community: CLI private beta, ToolJet MCP, tooljet-mcp on GitHub, read on 4 October 2026.

Both platforms connect to Databricks natively, and both built a working live dashboard; ToolJet’s cost $0.11 with GPT-6 Luna and $0.87 with GPT-6.1 Sol, against Retool Mini’s $5.40.

We connected the same Databricks workspace to both, using Databricks’ sample Bakehouse sales tables, and asked for a three-page sales analytics dashboard: KPIs with period-over-period change, revenue by day with a 7-day moving average, hour and weekday patterns, a franchise scatter chart, filters on every page and a full design brief, all queried live with no seeded data. The prompt is at the end of this article.

Build	Time	Cost	Stops	Result
ToolJet, GPT-6 Luna	12m 49s	$0.11	0	All 3 pages on live data; each page queries when you click Load
Retool Mini	5m 34s	$5.40	5 approvals	All 3 pages on live data, queried on open; the franchise detail panel shows an error until a row is picked
ToolJet, GPT-6.1 Sol	9m 13s	$0.87	0	All 3 pages on live data, queried on open; default date range a day early

Every app read the warehouse correctly: ToolJet’s GPT-6 Luna build and Retool’s showed the same figures, $66,471 of revenue from 3,333 transactions across 48 franchises. The GPT-6 Luna build used ToolJet’s load-on-click option for warehouse queries, so opening it spends no Databricks compute until someone asks for the data; the GPT-6.1 Sol and Retool builds query every time a page opens. Each ToolJet build had one flaw: GPT-6 Luna drew a selected franchise’s revenue by day unsorted, and GPT-6.1 Sol set its default date range a day early, so its Overview left out May 17 until the range was changed.

Setup. ToolJet’s Databricks data source takes the warehouse details and a personal access token, and its connection test passed first time. Retool’s resource takes the same fields; its first test failed on Retool’s side with a driver error, and a retry passed.

For a team that builds with AI every week, ToolJet’s monthly credits cover 5 to 34 times more eight-page builds than Retool’s, tier for tier, and every app beyond the allowance costs cents instead of dollars.

The example below is a team of 5 builders and 50 internal users, on annual billing. “Apps covered” divides each plan’s monthly credits by what our eight-page app cost on each model. These are estimates from single runs, not guaranteed capacity.

Plan	Price per month	AI credits per month	Eight-page builds the credits cover (estimate)	Each build beyond that
ToolJet Pro	$395 (5 × $79)	10,000 (2,000 per builder)	476 on GPT-6 Luna, 127 on GPT-6.1 Sol	$0.21 / $0.79
ToolJet Team	$995 (5 × $199)	15,000 (3,000 per builder)	714 on GPT-6 Luna, 190 on GPT-6.1 Sol	$0.21 / $0.79
Retool Team	$300 (5 × $10 + 50 × $5)	1,000, shared by the account	14 on Retool Mini, at most 12 on Retool Power	$7.10 / $8.00 or more
Retool Business	$1,000 (5 × $50 + 50 × $15)	1,500 to 3,000, shared by the account	21 to 42 on Retool Mini, at most 18 to 37 on Retool Power	$7.10 / $8.00 or more

The build counts measure AI capacity, not apps kept: ToolJet Pro allows up to 10 apps (Team has no app limit), and a team of up to two builders can use ToolJet Basic, which gives the same 2,000 credits per builder for $23, and most of a team’s credits go to edits and smaller builds.

Three things move the totals:

• Users. Retool charges for every internal user ($5 to $15 a month each), so its price climbs with headcount. ToolJet includes end users in the plan: up to 50 on Pro and 100 on Team, with unlimited users on Enterprise.
• Credits. ToolJet’s allowance grows with every paid builder (2,000 credits a month each on Pro, 3,000 on Team) and is shared by the whole workspace; Retool gives each account one fixed pool. Retool’s own sources disagree on its size: the pricing page lists 1,000 credits for Team and 3,000 for Business, including bonus credits it says “may change over time”, while its billing docs say the bonus ran through 30 September 2026. A Team plan we bought on 4 October showed 1,000 credits a month, so the table uses 1,000 for Team and gives Business as 1,500 to 3,000.
• Governance. SSO, audit logs and Git sync are on ToolJet’s Team plan. On Retool, audit logs need Business, and SSO and source control need Enterprise.

Sources: ToolJet pricing, Retool pricing, Retool AI credits billing and usage, read on 2 October 2026.

AI builds the first version; the platform decides whether you can run it for years. The biggest structural difference is hosting: ToolJet is open source and self-hostable for free, while Retool’s self-hosting is an Enterprise feature.

	ToolJet	Retool
Source	Open source, AGPL-3.0	Proprietary
Self-hosting	Free community edition; paid plans self-host too	Enterprise plan only (cloud, self-hosted or air-gapped)
What the AI generates	Apps in ToolJet’s visual builder: components, queries and pages you edit by hand or with AI	A React and TypeScript code project (Vite) with a live preview, edited by prompt or in the Code tab
Built-in database	ToolJet Database (PostgreSQL)	Retool Database (PostgreSQL)
Data sources	100+ databases, APIs and SaaS tools	Databases and APIs on every plan
SSO	Team plan (Google, GitHub, OIDC, LDAP, SAML)	Enterprise plan
Audit logs	Team plan	Business plan
Git	Git sync on Team	Source control on Enterprise
Workflows and agents	Self-hosted plans	500 workflow runs and 20 agent hours a month on Free
Bring your own model key	Not on Cloud plans; Enterprise offers custom AI model options	Enterprise

The second difference is what you own after generation. ToolJet’s AI writes into the same visual builder you would use by hand, so every component, query and event can be changed without code. Retool’s new builder generates code: you get full control if you write React, and a prompt box if you don’t.

Sources: ToolJet pricing, Retool pricing, ToolJet on GitHub, read on 1 October 2026.

Pick ToolJet if your team will build and change apps with AI as a habit. In our test the same eight-page app cost 10 to 34 times less on the entry and step-up models and 2.6 times less on the identical Claude Opus 5.5, the edits after it 16 to 28 times less, nothing stopped for approval, a second app never broke the first, and every app stays editable in a visual builder with no codebase to patch. Retool’s edits were faster, and its builder suits teams that want React code as the starting point; for everyone else, ToolJet delivers more apps and more models for the money.

• Choose ToolJet when AI usage will grow with the team, when you want the newest models from several providers, when non-developers need to edit what the AI built, or when you need to self-host without an Enterprise contract.
• Choose Retool if your team writes React and wants a generated code project as the starting point, and you expect only a handful of AI builds a month.

On both platforms, the model you pick matters. GPT-6 Luna is the cheapest way to build on ToolJet; GPT-6.1 Sol and Claude Opus 5.5 build better apps for more credits.

Explore ToolJet AI App Builder, or see our ToolJet vs Retool platform comparison for deployment and governance details.

Every build and edit in this article, with the credits each one used. ToolJet’s figures come from its credit meter; Retool’s from the account balance before and after each run.

Platform and plan	Build	Model	Time	Credits used	Cost
ToolJet	Expense tracker	GPT-6 Luna	4m 18s	2.9	$0.03
ToolJet	Expense tracker	GPT-6.1 Sol	2m 25s	22	$0.22
ToolJet	Expense tracker	Claude Opus 5.5	1m 40s	72.45	$0.72
ToolJet	Property manager	GPT-6 Luna	17m 24s	21	$0.21
ToolJet	Property manager	GPT-6.1 Sol	11m 11s	78.55	$0.79
ToolJet	Property manager	Claude Opus 5.5	15m 39s	323.17	$3.23
ToolJet	Edit 1	GPT-6 Luna	10m 26s	8.1	$0.08
ToolJet	Edit 2	GPT-6 Luna	11m 13s	9.2	$0.09
ToolJet	Edit 1	GPT-6.1 Sol	8m 1s	64.5	$0.65
ToolJet	Edit 2	GPT-6.1 Sol	7m 25s	61.17	$0.61
ToolJet, MCP plugin	Expense tracker	Claude Code (Sonnet 5.5)	4m 54s, 23 ToolJet tool calls	0	$0 in ToolJet credits
Retool Free	Expense tracker	Retool Mini	2m 8s	18	$1.80
Retool Free	Expense tracker	Retool Power	32m 47s	43	$4.30
Retool Free	Property manager	Retool Mini	About 56 min	71	$7.10
Retool Free	Property manager	Retool Power	Under 55 min	80 (all remaining)	$8.00 or more
Retool Free	Edit 1	Retool Mini	1m 10s	23	$2.30
Retool Free	Edit 2	Retool Mini	44s	15	$1.50
Retool Team	Expense tracker	Claude Opus 5.5	3m 44s	20	$2.00
Retool Team	Property manager	Claude Opus 5.5	13m 27s, from our answer to its database question	83	$8.30
Retool Team	Expense tracker	GPT 5.6 Sol	2m 1s	13	$1.30
Retool Team	Property manager	GPT 5.6 Sol	3m 19s	17	$1.70
Retool Team, CLI	Expense tracker	Claude Code (Sonnet 5.5)	About 2 min of coding	0	$0 in Retool credits
ToolJet	Databricks dashboard	GPT-6 Luna	12m 49s	10.75	$0.11
Retool Team	Databricks dashboard	Retool Mini	5m 34s	54	$5.40
ToolJet	Databricks dashboard	GPT-6.1 Sol	9m 13s	87.44	$0.87

ToolJet credits are $0.01 each and Retool credits $0.10, at each product’s paid top-up price.

Every prompt below was sent word for word to both products.

Expense tracker:

Build an expense tracker. Let me add an expense with date, category (travel, meals, software, office, other), amount and a note. Show all expenses in a table I can filter by category and month, with the month's total at the top.
Design: a clean, modern light theme on a near-white background; a header with the app name and a short subtitle; the month's total as a large figure in a white card, with the number of expenses beside it; the add-expense form in its own card with the fields side by side and one primary button; category shown as coloured tags (travel blue, meals amber, software violet, office teal, other grey); indigo (#4F46E5) as the primary colour; money as $1,234.56 and dates as Mar 4, 2026.

Property manager:

Build a property management app for a small residential landlord, with eight pages and a left sidebar to move between them.
Seed realistic sample data: 4 properties, 32 units, 26 tenants, a lease for each occupied unit, 6 months of rent payments, 20 maintenance requests and 8 vendors.
1. Dashboard: KPI cards for occupancy %, rent collected this month vs expected, total overdue balance and open maintenance requests; a 6-month rent collection chart; occupancy by property as a bar chart; and a list of leases ending in the next 60 days.
2. Properties: a card grid showing each property's name, address, unit count and occupancy %; clicking a card opens a panel listing its units.
3. Units: a table with property, unit number, beds, baths, square feet, monthly rent and status (occupied, vacant, notice given); filters for property and status; rent is editable inline.
4. Tenants: a searchable directory; selecting a tenant shows their contact details, current lease, payment history and open maintenance requests.
5. Leases: a list with tenant, unit, start and end dates, rent, deposit and status (active, expiring, ended); a form to create a lease for a vacant unit, which marks the unit occupied; and a Renew action that extends the end date by 12 months.
6. Rent: a ledger of payments by month; a form to record a payment against a lease; and an overdue list showing days late, with a 5% late fee added once a payment is more than 5 days late.
7. Maintenance: requests grouped by status (new, scheduled, in progress, done) with priority (low, medium, high, urgent); a form to log a request for a unit; and the ability to assign a vendor and change status.
8. Vendors: a list with trade, phone, email and rating (1 to 5); the number of jobs assigned to each; and a form to add a vendor.
Design: a clean, modern light theme on a near-white background, white cards with soft borders, emerald (#059669) as the primary colour for the main buttons and active navigation, status tags in consistent colours (green for good states, amber for warnings, red for overdue or urgent, grey for neutral), money shown as $1,234 and dates as Mar 4, 2026.

Edit 1: “Add a ninth page, Reports, to the sidebar: a chart of monthly rent income against maintenance spend for each property over the last 6 months, a table with the same numbers per property and month, and a button to export that table as CSV. Track a cost on each maintenance request so the spend is real data, and fill costs in for the existing completed requests.”

Edit 2: “Change the primary colour from emerald to navy (#1E3A8A) everywhere in the app, show all money without cents (for example $1,234), and add a ‘Days to lease end’ column to the Leases list, sorted with the soonest first.”

Databricks dashboard (sent after connecting the workspace’s sample tables on each platform):

Build a sales analytics dashboard for Bakehouse, a bakery franchise chain, on my Databricks data. Use my connected Databricks data source. The tables are samples.bakehouse.sales_transactions, samples.bakehouse.sales_franchises and samples.bakehouse.sales_customers; read them to learn their columns. The data is read-only: query Databricks live, and do not create, copy or change any tables, and do not use sample or seeded data.
Three pages with a left sidebar:
1. Overview: a date range filter (defaulting to the full range in the data), a franchise country filter and a payment method filter that drive everything on the page. KPI cards for total revenue, transactions, units sold, average order value and active franchises, each with the change against the previous period of the same length. Charts: revenue by day as a line chart with a 7-day moving average; revenue by product as a horizontal bar chart showing each product's share; revenue by hour of day; revenue by day of week; payment method mix as a donut chart; and the top 10 franchises by revenue as a bar chart.
2. Franchises: a sortable table of every franchise with name, city, country, size, revenue, transactions, average order value and revenue rank, with country and size filters; selecting a franchise shows its revenue by day, its product mix and its busiest hours. Add a scatter chart of franchises by transactions against average order value.
3. Customers: KPI cards for unique customers, revenue per customer and repeat customers (more than one transaction); revenue by customer country; customers by gender; and a table of the top 25 customers by spend with name, country, transactions, total spend and favourite product.
Design: a clean, modern light theme on a near-white background; a header on each page with the page title and a one-line description; KPI cards in a row of white cards with soft borders, the value large and the change as a green or red tag; charts in white cards with a title and a short subtitle, in a two-column grid; amber (#D97706) as the primary colour for buttons, active navigation and the main chart series, with warm neutral colours for other series; money as $1,234 on KPIs and charts and $1,234.56 in tables, and dates as Mar 4, 2026.