Julius AI pricing: Plus $20, Pro $45, and the semantic layer only on Business at $450

Julius pricing at a glance (from Julius's own pricing page)

• Free — $0, with an allowance Julius describes three different ways. The billing docs' FAQ says a one-time grant of 100 credits that does not renew; the credits table on that same page says "25 welcome bonus + 25 daily credits"; the pricing page says only "free daily credits" with no amount. None of the three reconciles with the others, so treat the Free allowance as small and read it off your own Usage page. The full quotations are further down.
• Plus — $20/mo, or $16/mo billed yearly. 2,000 credits a month (24,000 a year on annual), 1 seat.
• Pro — $45/mo, or $37/mo billed yearly. 5,000 credits a month (60,000 a year on annual), 1 seat, expanded context window.
• Max — $200/mo, or $166/mo billed yearly. 25,000 credits a month (300,000 a year).
• Ultra — $500/mo, or $416/mo billed yearly. 70,000 credits a month (840,000 a year). Julius's billing docs list this same tier as Max ($500) rather than Ultra, so expect the naming to differ between the docs and the pricing page.
• Business — $450/mo, or $375/mo billed yearly. 60,000 credits a month (720,000 a year) and up to 50 team members. Three larger bands live inside that same card rather than on cards of their own. The page's own card data groups them under the Business card's credit selector and titles them Growth, Growth2 and Growth3 — $750/mo (100,000 credits a month), $1,400/mo (200,000) and $2,500/mo (400,000), all read on 28 August 2026; all three of their buttons read "Get started with Growth". They are published, just not visible until you open that selector — worth knowing if your volume is in that range, and worth knowing that Julius's billing docs call all four of these levels "Business" while the pricing page titles three of them Growth, Growth2 and Growth3.
• Enterprise — custom pricing.
• Annual billing is presented as a lower effective monthly price rather than a separate plan, and annual credits are granted up front for the whole year rather than monthly.

Credits, not messages: what changed and why it matters

Julius used to meter usage by counting messages. It now measures credits, and its documentation is direct about the reason: different tasks cost different amounts depending on the compute involved, so a simple message costs less than running a full analysis or executing code.

The practical consequence is that Julius's own answer to how much a message costs is that there is no single answer. Its billing FAQ says cost depends on the message complexity and which tools are invoked — a quick question costs far less than a multi-step analysis that pulls data, runs code and generates a deliverable. The FAQ elsewhere puts it more bluntly: every action in Julius costs credits, and the cost depends on what you are doing, which AI model is used, and how complex the task is. Built-in tools, including the data connectors, draw on the same credit pool.

Metering by compute means you cannot compute your monthly cost in advance from a price list. You can only measure it after the fact, on Julius's Usage page, which tracks credit consumption in real time.

What burns credits fastest

Julius names three drivers in its documentation, and they are worth reading carefully because they map directly onto how analysts actually work.

• Action type — a plain message costs less than running an analysis or executing code. Since the point of an AI data analyst is to run analysis, most of your real usage sits at the expensive end of this scale rather than the cheap end.
• Model choice — Julius calls model selection one of the biggest drivers of credit consumption, and says its faster in-house and mid-tier models consume credits far more slowly than frontier models. If you leave the model picker on the most capable option, your credits go faster. This is the single biggest lever you control inside Julius.
• Complexity — longer outputs, bigger datasets and multi-step tasks add up faster. A question that requires four steps and a chart is not four times a one-liner; it is whatever the compute cost turns out to be.

The billing rules people get caught by

None of these are hidden — they are all in Julius's published billing and refund policies — but they are the mechanics that determine whether a plan feels fair or frustrating, and they are easy to skip on the way to checkout.

• Credits do not roll over. Unused credits expire at the end of your billing cycle. If your work is seasonal — heavy at month-end, quiet mid-month — you pay for a flat grant against uneven demand.
• You cannot buy more credits. Julius's docs state that purchasing additional credits or topping up your balance is not supported at the moment, and that if you need more usage right away you may consider upgrading your plan. That makes running out a plan-change decision, not a small incremental purchase — the next step above Pro at $45 is Max at $200.
• Credits are non-refundable and non-transferable. They have no cash value, cannot move between accounts, and are forfeited at the end of the current period if you cancel. Credits associated with a reversed payment are revoked.
• Refunds are reviewed case by case, not granted automatically. Julius's refund policy asks you to email its team with your name, purchase email, receipt number and reason, and states explicitly that reviewing your plan, billing interval, pricing and renewal terms before checkout is your responsibility. Approved refunds go back to the original payment method and can take up to ten business days.
• Upgrading resets your billing cycle to the upgrade date, and takes effect immediately. Leftover credits from the previous plan stack with the new plan's credits, but they keep their original expiry — the end of the billing period in which they were granted.
• Downgrading takes effect at the end of the current period, and the unused portion of your current plan is applied as a credit toward your next invoice rather than refunded.
• Switching billing cadence is asymmetric. Monthly to annual takes effect immediately and resets the cycle; annual to monthly takes effect at the end of the annual period unless you email support to expedite it. Julius's billing-cycles page also says to contact its team to switch cadences, which suggests the self-serve path may not always be available.
• Mobile subscriptions cancel elsewhere. If you subscribed through the App Store or Google Play, cancelling on the website does not stop those charges — you have to cancel through Apple or Google.
• Payments run through Stripe, with cards, digital wallets, bank redirects, debits and transfers supported depending on region. Monthly billing is described as every 30 days rather than on a calendar month.

Julius AI pricing: what is gated behind which price

Cost is not only the number on the plan — it is the number on the plan that includes the thing you need. Julius's own plan-comparison table draws these lines, and a few of them will decide your tier for you before credit volume ever enters the conversation.

Database connectors are the big one. On Free, the listed connectors are Google Drive, OneDrive and SharePoint — file sources. Snowflake, BigQuery and Postgres appear from Plus upward, so the cheapest Julius plan that connects a real database is Plus at $20/mo ($16/mo annual). Export formats — slides, HTML artifacts, charts and images — are also listed from Plus upward rather than on Free. The sandbox memory ladder that used to sit on this page — 2 GB of RAM on Free, 32 GB on paid plans, 64 GB on Enterprise, read 29 July 2026 — is no longer published anywhere on the pricing page: on 20 August 2026, and again on 28 August 2026, the words sandbox, GB, RAM, Python and notebook return zero matches across the whole page. For notebooks that is now confirmed at the source rather than inferred from an absence: julius.ai/docs/notebooks-sunset says, word for word, "Notebooks have been sunset" (read 28 August 2026). If sandbox size decides your purchase, ask Julius directly rather than trusting any figure that predates that change.

Two features are listed only from Business ($450/mo) upward: the Julius Slack agent, and Semantic Schema Learning. Scheduled runs are capped at 3 on Pro and unlimited from Business. Collaboration is ticked on Business and Enterprise only, and Plus and Pro each include one seat: adding team members starts at Business.

So Plus and Pro are individual plans, and the moment you need more than one person in the same workspace, the price step is from $45 to $450. Intellrise Team covers the same move, five seats with the semantic layer and a shared workspace, for $100.

One more thing about that ladder that is easy to misread: it is not monotonic. Business at $450/mo carries 60,000 credits a month, while Ultra at $500/mo carries 70,000. Business is not the tier above Ultra — it is the team tier, priced for shared seats rather than for volume. If you are buying capacity rather than seats, paying more can get you fewer credits.

Where Julius's own pages disagree with each other

Each disagreement below lands on exactly a number or a feature you would need in order to forecast a bill or pick a tier — what a tier includes, whether you can buy more when you run out, how many credits a year contains, and whether something you are being sold still exists. In each case the same question, asked of Julius's own published material, comes back with two different answers — sometimes from two pages, and three of the five times from two places on one page, which is the harder version to catch. The conclusion: you cannot compute your monthly cost from this table, and neither can they. Every quotation below was re-read at its source on 28 August 2026, and none of them had been reconciled on that date.

• The $450 tier is sold on a capability the $20 tier already lists. The Business card on julius.ai/pricing (read 14 August 2026 and again on 28 August 2026) names, word for word, "Postgres, BigQuery, Snowflake + more data connectors" as a reason to be on Business. The tier-comparison table on that same page fills the Data connectors row for Plus ($20/mo) with, word for word: "Google Drive, OneDrive, Sharepoint, Snowflake, Bigquery, Postgres". What Plus does not cover is the step immediately after connecting. On that same Julius comparison table, Semantic Schema Learning — the thing that remembers that your gmv is ex-VAT and that status 3 means refunded — is marked unavailable on Free, Plus and Pro, and available from Business at $450/mo. So on Plus you connect to Postgres and then re-explain your own columns, session after session; on Intellrise that layer is on Pro at $29/mo. If you are already paying for Plus, that gap, not the connector, is the thing worth fourteen days of your attention.
• One document says you cannot buy credits, and also that your purchase is non-refundable. In julius.ai/docs/billing/credits-overview (read 14 August 2026 and again on 28 August 2026), the answer to "Can I buy more credits?" is, word for word: "At the moment, purchasing additional credits or topping up your balance is not supported." The same document's answer to "What is the credit refund policy?" reads, word for word: "Credits are non-refundable, whether you bought them as a top-up or received them through your subscription." What it means for you: a refund policy governing a kind of purchase the same document says you cannot make, so the page cannot settle the question for you, and that is the question you need answered on the afternoon you run dry mid-analysis, not the morning you sign up.
• Pro's annual allowance disagrees with itself by 10,000 credits. The Pro card on julius.ai/pricing (read 14 August 2026 and again on 28 August 2026), switched to annual billing, reads "60,000 credits per year", which is consistent with the monthly view's "5,000 credits monthly" at twelve months. The FAQ in julius.ai/docs/billing/credits-overview says, word for word: "For example, a Pro annual subscriber gets all 50,000 credits for the year on day one." That document's own table also lists Pro at 5,000 a month — so the document contradicts the table printed inside it. What it means for you: before committing to twelve months, get the 10,000-credit difference confirmed in writing by a human at Julius rather than planning a year of work against a figure two Julius pages cannot agree on.
• The Free allowance has three published descriptions, not one. The credits table lists Free as "25 welcome bonus + 25 daily credits"; the prose under it says "The Free plan's 25 welcome bonus credits are a one-time grant and don't renew"; and the FAQ on the same page says "Free plan users receive a one-time grant of 100 credits when they sign up; these credits don't renew." So the one-time grant is either 25 or 100, and daily credits either exist or do not, depending on which line you read. All three quotations in this bullet were read on 14 August 2026, re-read on 28 August 2026, and were unreconciled on both days.
• Julius's pricing page is still selling a feature its own docs say is gone. The FAQ on julius.ai/pricing, under "Can I give Julius my business context?", answers word for word: "Yes, upload your data dictionary, train Custom Agents on your company schemas and tables, or include your DBT connection information." julius.ai/docs/custom-agents-sunset says, word for word: "Custom Agents have been sunset". Both read 28 August 2026. What it means for you: if training an agent on your schema is part of why you are buying, that is the sentence to get confirmed before you pay rather than after.

What a flat plan actually costs you when you are wrong

The question you ask first is almost never the question you needed. You ask for revenue by month, see the grain is wrong, ask again by order date instead of ship date, notice the join has doubled every row with two shipments, fix that, and somewhere around the seventh attempt you finally have something you would put in front of another person. That is not sloppiness — that is what analysis is: mostly rework, and most of the rework lands early, when the data is least familiar and the questions are worth the most.

A credit balance bills all seven of those attempts, not just the one you keep. The problem is you cannot see the meter before you spend it: the week you understand the data least is the week you ask the most questions, and it is also the week you can least predict what they will cost. On Intellrise there is no balance to watch and no counter in the corner of the screen, because Intellrise does not resell model capacity and so has nothing to count. Being wrong eight times in a row is a fact about the data, not a budgeting decision you have to justify.

The same four questions, answered for Intellrise. What a tier includes: Intellrise Pro is $29/mo, or $24/mo billed annually, and the semantic layer, every export format and every source type sit inside it. Whether you can buy more when you run out: there is nothing to run out of, because Intellrise sells no credits at all; model tokens are billed by your AI provider at its list price. How many credits a year contains: none, on either cadence. Whether what you just read still exists: every quotation on this page carries the day it was read, while Julius's pricing page carries no date of its own at all — last updated, as of and effective date all return zero matches (28 August 2026). You can check the first three of those for yourself in fourteen days of Intellrise Pro, no card required.

The structural difference: a meter prices your curiosity

Here is the part that matters more than any single number. Real analysis is iterative. You ask a question, look at the answer, notice the number is wrong because a filter was off, ask again, split it by region, realise the region field has three spellings, fix that, re-run, then ask for it as a chart. That loop is the work — not overhead on the work.

Under a compute-metered model, every turn of that loop has a price, and so does every retry that produced nothing useful. The cost of being thorough and the cost of being wrong are the same cost. That makes metering a tax on exploration. The incentive points one way: batching questions instead of asking them as they occur, leaving a cheaper model selected on work that needed the better one, not going back to check the surprising number. The thing to notice is the exchange rate — the money saved is small, and the analysis is worse.

A flat plan changes the shape of the bill rather than only its size. The subscription is fixed, and the variable part is raw token cost paid directly to your AI provider with no per-message markup in between, so a retry is charged at provider cost, and a cheap model can take routine questions and a strong one the hard ones without a plan change in the middle.

How Intellrise prices the same job

• Free — $0. One database source plus Google Sheets (uncounted), 3 dashboards with unlimited charts, unlimited reports and every export format (CSV, Excel, PDF, PPTX, Word), AI report editing, AI custom instructions, and editing dbt schema and field descriptions. Questions run on an AI key you add once; a free Gemini key works.
• Pro — $29/mo, or about $24/mo billed annually. Unlimited data sources — the count never enters the price — plus file upload, unlimited dashboards with full customisation (drag and resize, chart type and colour, edit the SQL), AI Semantic Schema Learning, per-key AI configuration, read-only share links, one-off email delivery of a report or dashboard, and a daily or weekly email schedule for a report.
• Every new account starts on a 14-day Pro trial with no card. After it lapses, the free tier stays free — it is the landing state, not an expiring demo.
• No credit meter, no message cap and no per-question markup on any plan, because Intellrise does not resell model capacity. Tokens are billed directly by your AI provider — Gemini, OpenAI, Anthropic, DeepSeek, MiniMax or any OpenAI-compatible endpoint — at their list price.
• Team — $100/mo including five seats, $20/mo for each seat after that, or $1,000/year. On sale since 16 September 2026: one organisation whose members share data sources, dashboards and reports, fall back to the owner's AI key if they have not added one of their own, and are managed with owner, admin and member roles. Five seats is the minimum, so a three-person team also pays $100. Against Julius Business at $450 a month, a five-person team on Intellrise pays $100 — the full comparison is on our Julius AI Business plan comparison.
• If, on the other hand, you are the one person who writes the queries — and "sharing" in your week means sending a chart rather than handing over a seat — that gap never touches your work. What does touch it is the bill, and the bill is the part you can settle for yourself in a fortnight, on your own database, without a card.

The arithmetic

The two models are not directly comparable question-for-question: Julius does not publish a credit cost per action, so there is no way to turn 5,000 credits into a number of questions from the outside. What can be shown is what the Intellrise side costs, so you can put your own numbers on both.

Take a schema-aware question against a connected database: the request carries a system prompt, your schema and semantic-layer context, and the conversation so far. Call it roughly 12,000 input tokens and 800 output tokens for a typical turn — more for a long multi-step analysis, less for a follow-up. At a cheap flash-tier model published at $0.30 per million input tokens and $2.50 per million output, that is 12,000 x $0.30/1M = $0.0036 in, plus 800 x $2.50/1M = $0.0020 out. About half a cent a question.

Run 500 questions in a month and the token bill is about $2.80. On Intellrise Pro that is $29 + $2.80, roughly $32/mo, or about $27/mo on annual billing. Push the same 500 questions through a frontier model published at $5 per million in and $25 per million out and the token bill becomes about $0.08 a question, so $40 — roughly $69/mo all in. A mid-tier model at $3 in and $15 out lands near $53/mo; a small model at $1 and $5 near $37/mo. Those are published list rates at the time of writing and they move, so substitute your provider's current numbers, but the shape holds: model choice moves your bill by a factor of ten. All totals are before any sales tax or VAT, which is added to the $29 at checkout where it applies.

Now scale up. Go to 2,000 questions a month on the cheap model and the token bill is about $11 — $40/mo all in, because the variable part is linear. That is the whole argument for a flat plan: the curve is a straight line, not a staircase. On a credit model, the answer to exhausting your grant is a tier change, because top-ups are not supported — and above Pro at $45 the next rung is $200.

One more piece of arithmetic: if you are an independent consultant carrying three clients, is that one Pro account or three? It is one. Pro sets no cap on the number of data sources, so three separate Postgres databases, on three different hosts, belonging to three different clients, sit on the same $29 and the same token bill above — the count does not enter the price anywhere. Our Terms put the condition where it belongs rather than on the plan: you are responsible for having the right to connect and query the data sources you add.

Everything you connect lives in one workspace, so each question's schema context lists the tables and columns of all three and a single query can join across them, because cross-source questions are half the point. A client who requires strict separation gets an account of their own.

Priced out: each isolated client is another $29/mo account, so two are $58/mo and three are $87/mo, each with its own login and its own AI key. So the earlier number is conditional, and the condition is not yours to decide: if your clients' data sits in databases they own, and you are the only person asking the questions, and none of them requires separation on paper, it is $29 plus your own token bill, once. If one of the three does require it, it is $58 plus tokens. Work out which of your clients would ask before you decide whether this is cheaper than what you pay now — and if that arithmetic is what decides it, write to us first rather than paying twice to find out.

What "never re-explain your columns" means in practice

Watch what actually happens when someone points a general chat tool at a real warehouse. They paste the schema of a 40-table Postgres, and then they type the footnotes: gmv is ex-VAT; status = 3 means refunded, not cancelled; orders_v2 is the table everyone uses and orders is a 2023 archive nobody dared drop; the finance week starts on a Monday. None of that is the question you came to ask, and all of it has to be typed before you can ask it. Next morning, new session, blank slate, footnotes gone. Whatever the per-question price is, that retyping is the tax you are actually paying.

Intellrise treats those footnotes as part of the database rather than part of the chat. On connection it annotates the schema table by table and column by column, and when you correct it in conversation — no, gmv excludes VAT — the correction is written into the semantic layer instead of into a transcript that expires. It persists across later sessions. You can open that layer and edit any definition by hand, and if you already maintain descriptions in dbt, Intellrise works from that shape rather than asking you to keep a second copy of your own vocabulary somewhere else. On Julius's own comparison table this capability, Semantic Schema Learning, is marked available on Business and Enterprise and unavailable on Free, Plus and Pro; on Intellrise it is on Pro at $29/mo.

The fourteen-day trial is the test: connect the messiest database you own rather than a clean sample, teach it your three worst column names on day one, then come back on day three and ask a question that depends on all three. If you have to explain any of them a second time, you have your answer, and finding out cost you no card details.

One AI key covers the whole team, at the provider's price

On Team, the owner adds one key and it covers every colleague who has not added their own, so a new teammate goes straight to asking questions.

Token spend is billed by the provider at its list price and itemised in tokens rather than in a unit Intellrise invented, with no reseller margin in between. Gemini, OpenAI, Anthropic, DeepSeek, MiniMax and any OpenAI-compatible endpoint are supported.

What does not run out

Intellrise Pro is $29/mo, or about $24/mo billed annually, and nothing in it meters your questions: no credit grant to spend down, no message cap, no per-question markup. You pay your AI provider directly at their list price, so a heavy month costs what the tokens cost rather than forcing a plan change. Julius credits do not roll over and cannot be topped up, so when a grant empties mid-analysis the only way on is a bigger plan.

How to actually decide, in about twenty minutes

• Write down your monthly question volume honestly, including retries and dead ends: retries and abandoned questions are the easiest part to leave out, and this figure is the input both pricing models are most sensitive to.
• Decide whether you need a database connector or only files. Julius connects a database from Plus at $20/mo; Intellrise Free connects one at $0.
• If you are already on Julius, open the Usage page and look at what a representative question of yours actually consumed. That single measurement tells you more than any pricing page, and it is the only way to convert credits into your own work.
• Count seats. If more than one person needs the same workspace today, that decides your tier — and in Julius's ladder it is a step from $45 to $450.
• Price the Intellrise side with your provider's current published token rates and your own volume, not with the numbers in this article. Add the flat plan. Compare that total to the Julius plan whose grant your measured usage fits inside.
• Test the part that is supposed to compound. On Intellrise that is the semantic layer, so correct one badly named column on day one and check on day three whether the correction survived — a tool that forgets your vocabulary is charging you to retype it, whatever the meter says. Intellrise puts that layer on Pro at $29/mo; on Julius's comparison table it is a Business feature at $450/mo.
• Then test both on one real question from your own data, because a cost model you cannot get a correct answer out of is not cheap at any price.

Try it before you pay for anything

There is a public demo at https://intellrise.com/demo/ that needs no signup and no card — a synthetic sample dataset, the exact SQL each chart runs, and the exact numbers it returns, so you can judge the substance rather than a feature list. Each query on it is the SQL the AI wrote, verbatim, after two column definitions were saved once. If it looks right, every new account starts on a 14-day Pro trial with no card, and after that the free tier keeps working: one database source plus Google Sheets, three dashboards, and unlimited reports and exports.

The two things Intellrise is built around are what the flat price buys. First, one question goes from a readable query plan to a chart, a dashboard and an exportable report, in one place. Second, it keeps a semantic layer you correct once — the AI annotates your schema, learns what your columns and tables mean from your conversations, and stores those definitions against the source, with a dbt mode and a manual editor — and they survive later sessions and a schema refresh. Read the first pass and fix what it got wrong; after that you do not re-explain your data, and you do not pay a higher tier to stop re-explaining it.