Pricing Report 2026 on the Agentic Data Industry

We spent a week reading every pricing page in our market. Fifty-two vendors: agentic data platforms, AI analysts, data catalogs, observability tools, semantic layers, and the coding agents that set what buyers think an AI teammate should cost. Every figure below was pulled from a vendor's own live page and then re-checked by someone whose job was to prove it wrong. Ten claims failed that second pass and are not in this post.

Every price here was verified on 9 August 2026. Two of them change on 1 September; we have flagged those.

What we found, in five lines

How the fifty-two were read. Every vendor's own live pricing page, or its AWS Marketplace listing where that is the only place it publishes, read on 9 August 2026 and re-checked by a second pass whose job was to prove the first one wrong.

Twenty-two of the fifty-two publish no paid price on their own site. 30 of the 52 publish a dollar figure for a paid plan on their own site. 16 publish none anywhere we looked, on their own site or on a Marketplace listing. For the remaining 6 the only public figure is a 12-month AWS Marketplace contract price, with nothing on their own pricing page: Atlan, Alation, Collibra, Collate, DataHub Cloud and Sifflet. That is the whole set, 30 plus 16 plus 6. A free tier, a trial, a credit grant and a third-party estimate of what other buyers paid are none of them a published price, and none of them are counted as one here.

• •Nineteen of the fifty-two define their own billing unit - OLU, ACU, ACC, SPU, KT, DSO, "agent review", "monitored vCore-hour". Four vendors use the word "credit" for four different things, from $0.0025 to $2.00 each.
• •Ten sit between you and your model provider and apply a multiplier or a conversion rate to your inference. One of them publishes the multiplier on its own pricing page: ~1.4× token cost.
• •Six pass model costs through at cost or require your own key. Credit where it is due - Cube, Altimate, Datus, Nao, Tributary and data.world.
• •A metered platform charges you more as it works better. On PromptQL's own published per-task figures, a team running 120,000 agent tasks a year pays roughly $42,000 above the underlying token cost. That gap is a multiple of your token bill, so it grows every time your agents do more.
• •Four criteria separate the pricing models: no markup on model spend, no proprietary billing unit, no per-seat charge, and a full Apache-2.0 core that is the product rather than a limited trial. Among platforms selling a paid product with a published price, we have not found one that clears all four.

Here is what we found in full.

Nineteen vendors define their own billing unit

Not a price. A private unit of account.

PromptQL bills in OLUs - Operational Language Units. TextQL bills in ACUs, Agent Compute Units. Dot bills in Agent Compute Credits. Soda has Soda Processing Units. data.world has Knowledge Transactions. definity bills per monitored vCore-hour. Recce bills per agent review, a unit its pricing page never defines.

And then there is the word "credit", which four vendors use to mean four incompatible things:

• •Zep - $0.0025 a credit ($25 per 10,000, Flex overage)
• •Monte Carlo - $0.18 a credit on Start, $0.28 on Scale
• •Hex - $0.50 a credit (a 50-credit pack is $25)
• •Cleric - $2.00 a credit

That is a spread of 800× between the cheapest "credit" and the dearest.

Monte Carlo actually publishes two live rates it calls a credit - $0.18-$0.28 on its order forms and $0.01 per unit on AWS Marketplace. They are 18 to 28 times apart and cannot be the same thing.

Every one of these vendors documents its unit, and most of them do it clearly. The difficulty is structural rather than deceptive: a unit that only one vendor uses cannot be compared against any other vendor's, and it is hard to convert into a monthly number before you have run a month.

Ten of them sit between you and your model provider

This is the part that surprised us, and it is the part that should change how you evaluate this category.

When an agent does work for you, something has to pay a model provider. The question is whether the vendor stands in the middle of that transaction and applies a multiplier or a conversion rate. Ten of the vendors we checked do. We are describing a pricing structure, not an accusation - several of them publish the mechanism plainly, and one publishes the multiple itself.

PromptQL publishes the multiplier on its own pricing page, which we respect:

"$0.14 per OLU, introductory - that's our at-cost price (1× the underlying token cost, no markup), for a limited time. The standard rate is $0.20 per OLU (~1.4× token cost)." - promptql.io/pricing and its pricing FAQ, 9 August 2026

Read that carefully. The at-cost rate is explicitly temporary. The standard rate - the one you will be paying - is roughly 1.4× what the tokens cost. On a team spending $30,000 a year on inference, that multiplier is about $12,000 a year, which is more than most software in your stack costs outright.

Cursor charges its rate on tokens you already bought:

"On Teams and Enterprise plans, third-party model requests include a Cursor Token Rate of $0.25 per million tokens. This rate applies on top of model API pricing for included usage, on-demand usage, and BYOK usage." - cursor.com/docs/account/pricing, 9 August 2026

Bring your own key, and you still pay the meter.

TextQL converts inference into ACUs at a published per-model rate. Claude Sonnet 5 costs 1,100 ACU per million input tokens today. On 1 September 2026 that becomes 1,650 - a 50% increase in the price of the same work, with no change to your usage.

GitHub Copilot is the most honest about the mechanism: "1 AI credit = $0.01 USD." Hex grants credits per seat and sells overage at $0.50 each. Dot sells them at $1.80. Athenic sells 250 for $1. Basedash charges $1,000/month of platform fee and another $1,000/month of AI credits. Atlan's docs say plainly that "each agent run consumes AI credits, and credit usage is tracked per tenant."

To be fair, and this matters: six vendors already do the right thing. Cube states it clearly - "the price of AI tokens is passed through from the underlying AI service provider without any markup." Altimate says "BYOK is always free and unlimited." Datus, Nao, Tributary and data.world all pass through or require your own key. If you take nothing else from this post, take the habit of asking every vendor in this category which group they are in, and getting the answer in writing.

What the market actually charges

Setting the meters aside, here is the money. Everything below is a vendor-published list price, verified on 9 August 2026.

Enterprise contracts, published by the vendor on AWS Marketplace

These vendors publish nothing on their own websites. They publish contract prices on AWS Marketplace, where anyone can read them.

Twenty monthly active users for $75,000 works out at $312 per person per month, before any agent has run.

Flat platform subscriptions

Notice how many of these numbers have a second number attached to them.

Per seat

$19 to $100 a month, clustering hard at $20-$50: GitHub Copilot Business $19, Claude Team $20, Paradime SPARK $20, ThoughtSpot Essentials $25, Altimate Pro $29, Hex Professional $36, Cursor Teams $40, Cube Starter $40, Hex Team $75, Cube Premium $80, Claude Team Premium $100.

That band is why data buyers think an AI teammate costs $40 a month. It is a reasonable instinct for a tool a person drives. It is the wrong instinct for a workforce that runs without one.

Why we can say, hands down, the best price in the market

Having spent two thousand words on other people's pricing pages, it would be poor form to be coy about our own - so here is where we are, including the part that does not flatter us.

We are not the fastest way to start. If you want to click something and be running before lunch, several of the vendors above will let you, and we will not talk you out of it. Slower to start is not the same as more expensive to run, though, and those two get conflated constantly in this category. The free tiers above are quicker to enter. They are not cheaper.

The whole rate card is three lines. A Deployment Sprint at $7,500, fixed fee, six weeks with a named engineer, credited in full against your first year. Then Scale from $1,000 a month and Enterprise from $3,000 a month, both billed annually, both with unlimited seats. "From" because we charge per workspace and a large estate runs more than one - never because you ran more agents.

Set that against the chart above. The catalogs start at $48,000 to $100,000 a year for the same job. Lightdash is $36,000 for a narrower one. Our top rung, at $36,000, starts where their entry contracts do.

Why the way in is paid

There is no free hosted tier, and that is a deliberate choice rather than an oversight.

The expensive part of getting agents into a data stack is not the software - it is a person. Six weeks of an engineer costs what an engineer costs. A vendor who gives that away is recovering it somewhere you cannot see, and in this market that somewhere is usually a token multiplier.

So we charge for the deployment and credit it back in full. If you continue, the sprint cost you nothing. If you do not, you keep the open-source core and everything built on top of it - your context graph, your configuration, the eleven Apache-2.0 agents - running on your own infrastructure. The managed pieces stop: the proprietary agents beyond those eleven, the hosted Conductor, the control plane.

We would rather that be on the invoice than in the margin.

Why open source is the free tier

Every free tier in this category is a metered sample of a hosted product - a one-time credit grant, three seats, three hundred credits, twenty credits a month. Those are trials with a clock on them.

Ours is the product. Eleven specialist agents, each an MCP server, running on your infrastructure, against your warehouse, on your model key, with no cap on anything and no expiry. For an enterprise it is a materially better evaluation than a hosted free tier could ever be: nothing leaves your network, so there is no data processing agreement, no vendor security review and no procurement conversation before you find out whether it works.

Four commitments

One: we will not meter your agents. No credits, no runs, no tasks, no questions, no invented unit. A flat platform fee per workspace, and that is the whole meter.

The reason is specific to what we build. Our product is a standing workforce - the Autonomous Data-Conductor is meant to run continuously across pipelines, incidents, quality, governance and cost. Under a per-run meter, the better it gets at working unsupervised, the more you pay for the same headcount, and your rational response is to throttle the thing you bought. You can see that anxiety in the vocabulary the metered vendors use on their own pages: prepaid balances, auto-pause at zero, overage rates, top-up thresholds, per-user quotas, spending alerts. One of them reassures customers that work "auto-pauses if your balance hits zero … no surprise bills." That sentence exists because the fear exists. We would rather sell the absence of it.

Two: we will not mark up your model spend. You bring your own key. Your provider bills you directly, at your rate, at every tier and in the open-source core. We do not resell tokens and we do not add margin to what your provider charges you.

There is no unit of ours between you and your inference. You should be able to switch models on a Tuesday because a better one shipped on Monday, and see the whole difference on your provider's bill. We are not alone in this and it would be dishonest to imply otherwise - several vendors above pass tokens through at cost or require your own key. What is unusual is holding it at *every* tier. Bring-your-own-model is commonly an enterprise entitlement, sitting behind the same wall as SSO and audit trails. Ours is not gated at all.

The precise version, so you can hold us to it: we will never resell your model tokens and never add margin to your provider's rate. If we ever offer to consolidate model billing onto one invoice as a convenience, it will be at cost, and it will be optional.

Three: we will not charge per seat. Unlimited users, at every tier.

Seats measure the wrong thing. The work is done by agents; the people are approving it, reviewing it and deciding what happens next. Charging for the fifth engineer who starts reading the receipts taxes exactly the adoption that makes the platform worth having. The going rate for an AI teammate is $19 to $100 a month per person, and that is reasonable for a tool a person drives. It is the wrong instinct for a workforce that runs without one.

Four: we will not lock you in. Bring your own model, catalog, coding agent or context layer. The core is Apache 2.0 and whole - no enterprise-only files hidden in the source tree, no licence-key check, nothing that stops working after a trial period. There is no notice period and no exit fee, and what you built on the open-source core stays yours. Leaving is always a decision you get to make.

Scoped precisely, because a promise this size should not be quoted in half: on the way out you keep the open-source core and everything built on top of it - your context graph, your configuration, the eleven Apache-2.0 agents - running on your own infrastructure. What stops is the managed platform: the proprietary agents beyond those eleven, the hosted Conductor, the control plane.

The whole-year arithmetic

Here is the claim stated so you can attack it. Take any workload, add up the platform fee and the model spend, and the cheapest way to run it is one of ours. Not the cheapest to sign up for. The cheapest to run.

One assumption carries the whole chart, and it is ours. The task mix below is blended 70/25/5 across PromptQL's three published anchors, which is our assumption and not theirs, and it gives 5.9 OLUs per task. The crossover moves with your task mix.

In every row, the lowest number is one of ours - and it is not close at volume.

We will be precise about the one place a metered platform wins, because it is the first thing a competitor will point at. Below roughly 34,000 agent tasks a year, PromptQL's consumption model is cheaper than our Scale subscription. That is real, and the crossover moves with your task mix. It is also not an argument against the claim, because below that volume the right answer is our open-source core, which costs nothing at all and still beats the meter. The flat fee is for when you have outgrown that.

So the honest, complete version is this: there is no workload we have found where another platform in this survey is cheaper to run than the best-fitting Data Workers option. If you find one, send us the numbers and we will publish them here alongside ours.

We are not the only company that passes model costs through - Cube, Altimate, Datus, Nao, Tributary and data.world all do, and they deserve credit for it. Among the platforms in this set that sell a paid product, we have not found one that does all four: no markup, no invented unit, no seat charge, and an Apache-2.0 core the paid product is built on rather than beside. Cube and Altimate clear the first two and then charge per seat. Datus comes closest of anyone - Apache-2.0 core, bring-your-own-key, no billing unit - and publishes no price at all above its free tier, so there is nothing yet to compare. If you find a platform that clears all four with a price on the page, tell us and we will update this post.

The arithmetic that actually decides this

The chart above gives the totals. This is where they come from, and why the gap widens rather than closes. Take PromptQL's own published anchors - under 2 OLUs for a simple task, about 10 for a complex report, about 40 for a deep investigation. Blend those at 70/25/5 - that workload mix is ours, not theirs - and the average agent task costs 5.9 OLUs. At the standard $0.20 rate, that is $1.18 a task. At the at-cost rate PromptQL itself defines as 1× the token price, the same work costs $0.83.

The 35 cents in between is not a fixed cost. It scales with exactly one thing: how much work your agents do.

On that 70/25/5 assumption, a team running 60,000 agent tasks a year pays about $21,000 above the underlying token cost. At 120,000 tasks, about $42,000. PromptQL describes the $0.20 rate as "all-in - tokens + infra + sandbox hosting", so some of that gap is real infrastructure rather than pure margin. That is a fair point and we will not pretend otherwise. What does not change is the shape: whatever the gap is buying, it is charged as a multiple of your token bill, so it grows every time your agents do more.

That is the number a flat platform fee has to beat, and it is not a close contest. Our entire subscription costs less than the multiplier does, and unlike the multiplier, it stops growing.

This is the real argument, and it is not about being a few dollars cheaper. If you buy a metered platform and then succeed with it - if you actually get to a standing workforce that watches your pipelines overnight, catches the incident at 3am, and files the fix with a receipt - your reward is a bigger bill every single month. The rational response is to run fewer agents. Which means the pricing model is quietly arguing against the product.

We would rather sell you a number that does not care how autonomous you get.

How to check any of this yourself

Everything above is on a public page. If you are evaluating anyone in this category, including us, these four questions are worth asking in writing:

• •Is there a unit between me and my model provider, and what is its dollar rate today? If the answer involves a name you have not heard of, ask what multiple of the underlying token cost it represents.
• •What happens to my bill if my agent volume triples? Every vendor has an answer. Some of them are a straight line through the origin.
• •Can I bring my own model key on the plan I am actually buying? Several vendors offer bring-your-own-model only on Enterprise contracts. Check the tier, not the marketing page.
• •Is the introductory rate the rate? PromptQL's own page says the current one is "for a limited time" and names the standard rate 43% above it. That is admirably clear, and it is the kind of thing worth finding before you sign, not after.

Every figure in this post was taken from the vendor's own live pricing page or AWS Marketplace listing on 9 August 2026, and independently re-verified before publication. Two are time-sensitive: TextQL's Sonnet 5 ACU rate rises on 1 September 2026, and GitHub Copilot's promotional credit allowance ends the same day. Prices change; we would rather be corrected than wrong. If we have something out of date, email us and we will fix it and say so.

On the figures. Everything above is a publicly listed price or an AWS Marketplace contract price as published by the vendor on 9 August 2026. Negotiated enterprise contracts routinely differ from list. This analysis uses only what the vendors publish themselves. If we have something wrong, email us and we will correct it and say so.

What "best price" means here. The lowest total cost of running agents against your data for a year - platform fee, plus model spend, plus whatever the meter adds on top. Not the lowest number on a signup page. We show that arithmetic in full above, with the assumptions marked, so you can run it against your own numbers and tell us if we are wrong.