Starter
A solid foundation for your first deployment.
$99/month
5,000 runs included
Start with Starter- Unlimited agents
- Model call records and evaluations
- Deterministic guardrails
- No overage — move to Growth when ready
Why Krazimo Studio
One governed path for every model call.
Answering is the easy part. Being able to show why it answered, prove it still behaves after the last change, and hand someone a bill they can check — that is the part teams assemble from half a dozen vendors. Here it is one product, at a fraction of what the pieces cost separately.
The same volume, priced in each vendor's own unit: our run, a Lyzr run, an Agentforce action. Ours is one agent turn, flat — a question that takes eight steps bills the same as one that takes two. At forty thousand a month the gap widens rather than closes: $499 against $3,200 and $4,000. Platform fee only. Lyzr states it bills inference separately, as we do; Salesforce does not publish a separate inference charge, so its figure may already include the model.
Lyzr also publish $0.03 a run for VPC and on-premises and label it their best value; the comparison here is cloud to cloud, because on VPC the customer brings and pays for the compute and we do not sell a self-hosted plan today. Lyzr and Salesforce figures from their own published pricing. Salesforce also sells Agentforce at $2.00 per conversation rather than per action; that is a larger unit holding several turns, so it is not shown here — multiplying it against a turn count would overstate what they charge.
You are not buying an agent. You are buying the ability to defend one. Anyone can ship something that answers. The cost lands later — the week a customer quotes your agent back at you and someone has to say where the answer came from, and whether it will happen again.
What running an agent in production takes
Ten things stand between a demo and something you would put in front of a customer.
Eight of them you can buy from someone. Assembled, that is eight products on eight units with eight dashboards, and the code that makes them agree is yours to write and yours to keep working. Here they are one thing, on one bill.
| What you need | What teams buy for it | What you would pay | What you get instead |
|---|---|---|---|
| The agent loop — which tool to call, in what order, when to stop | An agent platform | $19 to $349 a month | The loop, the step budget and the tools |
| Answers drawn from your own documents | A vector database and its pipeline, or a hosted retrieval service | $25 to $100 a month | No vector database at all — the agent opens named pages instead of embedding your library |
| A record of every model call, kept and searchable | An observability tool | $29 to $199 a month | On every call, not a sample |
| Proof it still behaves after you change something | An evaluation platform | $100 to $249 a month | The suites, the checks and the judge |
| Guardrails on the question on the way in, and the answer on the way out | A guardrails service | $15 to $40 a month | A scope screen before the model runs, and fixed rules after it answers |
| Keys, provider routing, failover | A model gateway | 5% of your model spend | We take no share of your model spend |
| Changing what the agent may say without shipping code | A prompt-management tool | $49 a month | Publishing a version serves it, and every version says who published it and can be reverted |
| Not paying twice for the same context | A gateway with caching | $79 to $799 a month | Repeated context bills at a tenth of the input rate |
| All eight | Eight products, eight bills, and the code that holds them together | $316 to $1,785 a monthplus 5% of your model spend to the gateway | $99one bill, to start |
Ballpark monthly figures from each vendor's published entry pricing, for a small team; guardrails services price per thousand checks and are shown here at that volume. They all rise with use, and several vendors — Sierra, Decagon, Voiceflow, Lakera — publish no price at all.
One note on retrieval: at a small library most vector databases are free or nearly free, so this is not where the money goes. It is where the work goes — chunking, embedding, and re-indexing every time a document changes — and where a particular failure comes from, since a nearest-neighbour match always returns something, even when your library has no answer.
If you build on a hyperscaler instead
The agent loop is free there, and we are not going to pretend otherwise.
Amazon meters Runtime, Gateway, Memory and Evaluations but has no line item for agent orchestration. Microsoft states it outright: There is no additional charge for creating or running Foundry-native agents using prompts and workflows.
1 What you pay for instead is every call. Retrieval is $1.00 per 1,000 Retrieve API calls2, or $4.00 per 1,000 Agentic Retrieve calls, plus $1.00 per 1,000 underlying Retrieve calls3. Guardrails are $0.104 for contextual grounding and $0.155 for content filters.
Worked against a real workload at 5,000 runs a month — one agentic retrieval per run, guards on the question and the answer, a two-gigabyte index — that comes to roughly $40 in platform fees against our $99, and about $253 against $499 at forty thousand. The metered path is cheaper than we are at both volumes on this page.
What it does not include is retrieval that could not have written the answer in the first place, guards that are rules rather than scores, or one version number over the library, the policy, the suite and the model. Seven metered services produce seven bills and seven change histories, and no sentence a compliance owner can sign.
The full arithmetic, line by lineAnd the part you would have to build
Two of them are not something you can bolt on later.
Everything above can be bought from somebody. These cannot. There is no price to put beside them because no vendor packages them — which means the only ways to have them are to build them yourself, or to run something built this way from the start.
An agent with no way to answer outside your library
You can buy a score that grades an answer after the model has written it — AWS charges $0.10 per thousand units for one, and its own documentation says the check runs on the output. You cannot buy retrieval that had no way to write it. Ours walks an index and opens the page it names; a path no listing offered is refused by the tool itself.
“the model attempts to only rely on your documents… This option encourages the model to respond using your data only.”
One version number over the documents, the rules and the tests
Model registries version the model. The closest product to this pins the knowledge base association rather than its contents — re-sync the documents and every earlier version quietly changes meaning. Here one version covers the library itself, the policy, the suite that had to pass and the model, records who published it, and can be put back.
“You can name exactly what was serving on the day a customer got an answer, and you can restore it.”
Three things are deliberately not on this list.
A disposition field, citations to source pages, and a meter that also bills are all real here — and all of them can be bought elsewhere, so they belong in the feature table below rather than in a list of differences.
Build it yourself
What would it cost you to build this? Use your own numbers.
Most vendors answer this with a number they made up about your team. We would rather you put your own in — your salaries, your estimates, and the pieces you would actually build. The difference at the end is whatever your numbers say it is.
Your numbers
The last one is the work either way — connecting it to your product and pointing it at your documents. It is on our side of the comparison, not left out of it.
What you would build
Days are what each piece took here. Yours will differ — change them, or switch off anything you would not build.
Build it
$137,302
first year, all in
Buy it
$4,919
first year, all in
Difference, on your numbers$132,383
Where those come from
- Building it — 147 engineer days
- $109,685
- Running it for a year
- $27,618
- Wiring ours in — 5 days
- $3,731
- A year on the entry plan
- $1,188
Every year after the first$27,618 against $1,188
A build keeps costing upkeep and hosting for as long as it runs. The wiring is not paid twice on either side.
Elapsed at 2 engineers: 3.4 months before it serves anyone. Every figure here comes from the numbers on the left, which are yours to change — we have guessed at none of them except how long each piece took us.
Cheaper, and harder to be wrong
Cheaper would not be worth much if the answers were worse.
They are not, and the reason is architectural rather than a matter of trying harder.
It navigates, it does not guess
The agent walks an index and opens the page it names. No embedding, no nearest neighbour — so there is no such thing as a confident answer drawn from an unrelated document, and no vector store to operate or pay for.
Refusal is structural
A path no listing offered is rejected by the tool itself, not discouraged by an instruction. The refusal is a property of the system rather than of the model's mood on the day.
The guards run after generation
Deterministic rules applied to the finished answer, each reported by name with a reason. Every one replaced a prompt instruction that was tried first and did not hold. A fixed rule has no opinion to change.
Proof runs against the real thing
Test cases run against the deployed system, through the same sign-in, checks and guards a real customer's question passes through — not against a copy of it that behaves differently.
What one turn passes through
Simpler to build against
The person who has to live with this is usually not the person who chose it.
So here is the same decision from their side of the desk.
Parse the prose to see whether the agent actually answered
Every response carries decision — answered, declined or escalated
Route an escalation on a field, not a substring match
Show a similarity score and hope the reader trusts it
Every response carries the pages it opened
A citation lands on the document, checkable in one click
Reconcile a vendor invoice against your own logs at month end
Usage is returned on the call that incurred it
The same number that bills. No second record to disagree
Ship a release to change what the agent may say
Publishing a policy version from the console serves it
A wording fix takes a minute and leaves a version history
Stand up an ingestion pipeline and a vector database
Point the agent at a document tree
Nothing to chunk, embed or re-index when a document changes
Write your own meter before you can charge your users
Per-user allowances, and a clear answer when someone reaches theirs
Sell a plan to your own customers on day one
The whole price list
Two plans and a rate. No free tier, and no sales call to see a number.
Model spend is never included and never marked up — it is passed through at cost, or billed straight to your own provider key.
Growth
Built for growthMore capacity. The controls to grow with it.
$499/month
40,000 runs included
Get started with Growth- Everything in Starter
- Per-user allowances and limits
- Policy history and rollback
- Caching and usage on every call
Scale
A simple rate for what comes next.
$0.01/extra run
Beyond Growth’s 40,000 runs
Let’s talk scale- An extension of Growth, not a new plan
- The same rate for every extra run
- No markup on model spend
- Talk to us about your deployment
The developer comparison uses the US Bureau of Labor Statistics median wage for software developers plus its published benefits share — about $93 an hour, fully loaded. Your own number is the one that matters; the arithmetic is there so you can substitute it.
Questions we get
What exactly is a run?
One agent turn: one question in, one governed answer out. Every step it takes to get there — the documents it opens, the tools it calls, the guardrails — is inside that one run, whether it takes two seconds or two minutes. It is the same unit Lyzr publishes, so comparing us is arithmetic rather than argument.
Does a long run cost more than a short one?
No. A turn that walks eight steps bills exactly what a turn that walks two does. Length is our problem to fix, not yours to fund — and it already has a meter, in the model spend you are billed for separately.
Do you take a margin on model spend?
No. It is passed through at cost, or billed straight to your own provider key. This is the position, not an introductory offer — a per-call margin is a toll on your growth and we are not building one.
Is there a per-agent fee?
No, and the reason is worth stating: charging per agent would push you to fold ten jobs into one policy to save the fee, and one policy covering ten jobs is worse governed than ten covering one each.
What about voice, workflows, CRM channels, a marketplace?
Not today. Nothing in the design stands in the way of them — they get built when a deployment needs them, so if yours does, say so and it goes on the list.
Can we put our own model account behind it?
Yes. Point your workspace at your own provider key and inference bills to you directly, at whatever rate you have negotiated. Otherwise it is passed through at what it cost us.
What happens if we want to leave?
The engine that answers is a self-contained container: your library, a config file and a model key, with no database and nothing managed behind it. We intend to publish it under Apache-2.0. The point is that you are not holding a system you cannot run without us.
Sources
Every external figure on this page, with the wording it was read from and the date it was read. Competitor prices change; if one of these is out of date, tell us and it gets corrected here rather than argued about.
- 1Microsoft, Foundry Agent Service pricing
“There is no additional charge for creating or running Foundry-native agents using prompts and workflows.”
azure.microsoft.com/en-us/pricing/details/foundry-ag…read 2026-09-07 - 2AWS, Bedrock pricing — Knowledge Bases
“$1.00 per 1,000 API calls”
aws.amazon.com/bedrock/pricingread 2026-09-07 - 3AWS, Bedrock pricing — Knowledge Bases
“$4.00 per 1,000 Agentic Retrieve API calls + $1.00 per 1,000 underlying Retrieve API calls”
aws.amazon.com/bedrock/pricingread 2026-09-07 - 4AWS, Bedrock pricing — Guardrails
“$0.10 per 1,000 text units”
aws.amazon.com/bedrock/pricingread 2026-09-07 - 5AWS, Bedrock pricing — Guardrails
“$0.15 per 1,000 text units”
aws.amazon.com/bedrock/pricingread 2026-09-07
Stop assembling the stack. Publish an agent you can defend.
A library, a policy and a suite — then a release someone is willing to sign. $99 a month to start, $499 at forty thousand units, and your model spend stays your model spend.
Open the console