Skip to main content
New: Introducing Irys Workbench, agents to take up the work for you. See Workbench
Book a DemoStart Free Trial

Architecture

Why Irys is priced the way it is.

A chatbot sees the current prompt. A matter platform preserves the work. Here's how Irys is built, and why that supports flat pricing.

How Irys is built

Three proprietary layers, and an LLM only at the end.

Irys has three systems that run before any frontier model is called. By the time the frontier model is reached, the prompt is small and the context is tight, so the model call stays efficient.

Your organization's context

Irys
Every matter's facts, parties, timeline, issues, and authorities, structured once and cached per matter.
Every matter in Irys produces a structured representation of facts, parties, timeline, issues, and authorities.
Built once, cached per matter, updated as new documents are ingested.
A question about the Smith case doesn't re-read every document in the matter. It retrieves from the graph.
The platform remembers what your matter is about, what you've asked, what you've drafted, and what's been decided.
That memory is stored in our infrastructure and retrieved in milliseconds, not passed into an LLM prompt on every query.

Irys intelligence layer

Irys
Reasons about what's actually needed, pulls only the relevant material, and compresses context before any model call.
When a query comes in, our system reasons about what's actually needed, pulls only the relevant material from the matter, and compresses context before any LLM call.
Many queries are answered entirely inside our infrastructure without touching a general-purpose model.

Narrow model inference

Model provider
Only the final synthesis step calls a frontier model, with a small prompt and tight context.
Only the final synthesis step calls a frontier model.
By the time it does, the prompt is small and the context is tight, instead of the entire matter being sent to the model on every query.
Model-agnostic by design. When Claude improves, Irys improves. When GPT improves, Irys improves. We're not locked to any single lab.

Verification

Irys
Citations checked against real, still-good-law sources before work reaches you.
Cite Check checks whether cited authorities can be matched to available case records and surfaces treatment signals for review.
Researches real case law with checked citations.

Finished work product


Comes back as a Google Doc, Word draft, Excel workbook, or email draft, ready for review.
Brings finished work back to the matter: a Google Doc, a Word draft, an Excel workbook, or an email draft ready for review.

The market today

We built Irys to be architecturally different.

More context.

The platform remembers what your matter is about, what you've asked, what you've drafted, and what's been decided.

Better work.

The meaningful difference is what each platform builds around the model: legal sources, matter structure, retrieval, citations, integrations, and governance.

Lower model cost.

Many queries are answered entirely inside our infrastructure without touching a general-purpose model.

Less data exposure.

Our system pulls only the relevant material from the matter and compresses context before any LLM call, so less of it ever reaches a third-party model.

What this means for your bill

It's why we can charge flat

Here's the market, as publicly listed.

Legal AI has split into five rough categories. Where a standard price is not published, we say so:

Legal AI market price categories
CategoryTypical priceArchitecture
General consumer AI$20 / monthNot built for legal work
Legal AI point toolsOften not publicly listedContract-focused, Word-native
Legal AI assistantsOften not publicly listedFrontier models plus legal workflows
Irys One$299 / monthFull platform, proprietary architecture
Enterprise legal AINot publicly listedFrontier models plus enterprise deployment, sales-led
Most products in this table, Irys included, use frontier models for part of the work.
The meaningful difference is what each platform builds around the model: legal sources, matter structure, retrieval, citations, integrations, and governance.
That surrounding layer is what determines quality, and it is also what shapes cost structure.

What this actually costs us

Service-like margins at software-like prices.

A heavy Irys power user. 25 matters, hundreds of documents, daily usage, costs us under $20 a month in infrastructure and compute, by Irys internal analysis.
The number stays low because the platform structures each matter once and retrieves only what a task needs, instead of loading large context into a model on every query.
That cost base is structural.
It's not a promotion, it's not a pricing strategy, it's not a race to the bottom.
It's what happens when the proprietary infrastructure does the work before the LLM ever sees the query.
Usage-based pricing is common in AI because model calls carry a real marginal cost.
When a product sends large context to a model on every query, heavy use is expensive, and that pressure tends to surface as metered plans, capacity tiers, add-ons, or overage charges.
Claude's public plans are $20, $100, and $200 per month with defined usage capacities, and some enterprise configurations use credits or usage-based billing.
Confirm current terms with Anthropic before modeling costs for legal work.
Irys is different because our cost base is different.
The proprietary infrastructure does most of the work before a model is called, so by the time a frontier model is reached, the prompt is small and the context is tight.
That is the structural reason we can offer one number per seat with no usage caps and no overage line on the advertised plan.

$299 per seat per month. Every module. Every matter. No token billing.

Read the full comparison breakdown

The structural flip

What happens to pricing when model costs move?

Here's a question worth asking about any AI product: what happens to its economics when frontier model prices move?
A product whose costs are mostly model calls rides the labs' price curve in both directions.
A product whose costs are mostly its own infrastructure does not.
Our costs don't move much with lab price changes, because by Irys internal analysis most of our compute runs in our own infrastructure.
But our pricing doesn't face the same pressure either, because our unit economics were never dependent on the labs.
When the labs cut prices, our margin expands.
When the labs raise prices, we're insulated.

That's the structural difference between running on someone else's infrastructure and running on your own.

What happens when frontier model prices drop 50%: model-dependent AI versus Irys
When frontier prices drop 50%Model-dependent AIIrys
COGS drops−50%−10%
Savings captured byCustomers (competitive pressure)Irys (retained as margin)
Directional outcomeMargin compressionMargin expansion

What this means for you

We're a different kind of company, doing a different kind of work.

We priced Irys One at $299 a month because that's where our costs are.
We didn't price at the enterprise level, where standard pricing is usually not even published, because we don't need to.
We have service-like margins at software-like prices, which means we can pass efficiency on to customers instead of extracting it.
We didn't price at Claude's $20 because we're not a chatbot. We're a platform with a proprietary backend that took years to build and continues to deepen.
Much of the legal AI market does not publish standard pricing at all, and where costs track a frontier lab's price list, the number depends on someone else's infrastructure.

Ours doesn't. $299, published.

See the architecture run.

7-day free trial. Full platform, no credit card. Or book a call and we'll walk through the stack and your unit economics side-by-side.

Start 7-day free trial Book a Demo

Comparisons above are based on publicly available information reviewed July 28, 2026 and may vary by customer, contract, and deployment. Where a vendor does not publish standard pricing, we say so. Irys architecture and efficiency figures are Irys internal analysis. For the detailed technical breakdown, contact us .