Pay for agents the way you pay for people.

Lyzr prices the digital workforce in agent-months: a pool of governed agents in production, budgeted the way a finance team already budgets headcount. Governance, training and hardware are flat. Nothing is metered on tokens.

Build your estimate

All prices annual and prepaid. Deployed in your own cloud, on Optimus, or Lyzr-hosted.

Three layers. One registry.

Govern the estate at a flat price. Build and orchestrate on agent-months. Train on a separate pool. Everything reports into the same control plane.

Govern
See, permission and control every agent your company runs.
Flat, unlimited usage
The control plane

Governs every agent in your estate, Lyzr-built or not: identities, policy, evals, observability and lifecycle. Agents from Copilot, Bedrock, LangGraph and other frameworks come under management the moment you turn policy on. Unlimited agents. Unlimited users.

$250K
per year, unlimited
The default stack

Everything in Opencontroller, plus basecode: monitors the coding agents in your pipelines and the agents running on individual laptops across the company, so every autonomous process is on the record. Runs entirely inside your own cloud.

$500K
per year, unlimited
Build and orchestrate
Create agents, compose them into workflows, promote them into production.
Priced in agent-months
Agent Studio, Architect and SuperFlow
Developers, business users, orchestration

Build in Studio, build in Architect, wire it together in SuperFlow. You buy a pool of agent-months and every governed agent in production draws from it each month, weighted by what it is allowed to do. Agents in development draw nothing.

Each block includes two Lyzr forward deployed engineers for year one. Unused agent-months roll over on multi-year terms.

$250K
per block of 150 agent-months
Beyond the block: $1,700 per agent-month, or the next block
Train
Own the model, not just the agent.
Separate training pool
Openfinetuner
Fine-tuning and post-training

Fine-tune and post-train open-weight models on your own data for agentic tasks, inside your own cloud. Training draws on its own pool of tokens and never touches your agent-months.

$250K
for 2.5T training tokens

What an agent-month is

One governed agent, running in production, for one month. The same object as an FTE-month in a labor plan. A governed agent has its own identity, its own policy and its own deployment lifecycle. Runs, sessions, replicas, versions and clones are not separate agents, just as a person working overtime is not a second person.

A
Retrieve

Read only. Answers questions from knowledge, drafts, looks things up. No side effects.

0.25

A quarter of a $50-an-hour worker. About $25K of labor a year.

B
Act

Writes to enterprise systems: SAP, ServiceNow, Guidewire, Salesforce. No shell, filesystem or browser.

1.0

One $50-an-hour worker. About $100K of labor a year.

C
Privileged

Filesystem, code execution, browser or computer use, payments, identity administration. Runs on its own dedicated sandbox and is priced as standing capacity: there is no part-time rate for Class C.

2.0

One $100-an-hour engineer or analyst. About $200K of labor a year.

Interns are free

You do not put someone on the payroll on the day you interview them. Agents in development, in Architect, in Studio or in a sandbox pointed at UAT draw nothing. You pay when an agent is promoted into production with an identity, a policy and a lifecycle.

One workflow is one agent

A SuperFlow workflow with twenty nodes draws once a month, at the class of its most privileged node. A shared service used by six workflows is billed once for the estate. A specialist only its parent can call draws nothing. Good architecture is the cheapest architecture.

Activity moves the rate

Payroll has part-timers, and so does this. An agent that completes fewer than 100 units of work in a month draws a quarter of its class rate. One that completes more than 100,000 draws one class up. Both are decided monthly by the controller, for that month only. Class C is always full rate.

Draw per month by activity band
Class
Part-time, under 100 runs
Standard
Busy, over 100,000 runs
A Retrieve
0.06
0.25
1.0
B Act
0.25
1.0
2.0
C Privileged
2.0
2.0
3.0

A run is one unit of work completed, not one wake-up: a request answered, a ticket closed, a claim decided, a record processed. A nightly job that clears 10,000 claims is 10,000 runs, not one. Your controller counts them from its own traces, so the band is a fact, not a negotiation, and the number on the invoice is the number on your screen.

One block, one year, one insurer

A mid-sized insurer buys one block of 150 agent-months. Over two quarters, working with the two Lyzr engineers included in the block, they promote 19 agents into production.

Agents in production
Weight
Draw
2 privileged agents

A claims file handler reading scanned documents in a sandbox and writing decisions to the claims system. A coding agent maintaining internal tools.

2 × 2.0
4.0
5 acting agents

Purchase-request validator writing to SAP. ServiceNow ticket resolver. Invoice matcher. Employee onboarding provisioner. Renewal quote assembler.

5 × 1.0
5.0
12 retrieve agents

Policy questions, HR handbook, product lookups, contract clause search and the rest of the read-only estate.

12 × 0.25
3.0
Monthly draw
12.0

Complete the stack

The layer your employees use every day, and the box that runs all of it with no traffic leaving the building.

The consumption layer

Chat, search and personal agents for everyone in the company, running on your own models and your own data. Unlimited users. Personal agents are free. When a team promotes a Cowork agent into a shared workbench it becomes a governed agent and draws from your agent-months.

$100K
per year, unlimited users
The on-prem appliance

The full Lyzr stack on hardware you own, inside your own data center, air-gapped if you need it. Year one covers the hardware alone. The software layers above license on top, the same way they do in the cloud.

$500K
year one, hardware

Never on the meter

Users
Knowledge bases
Observability and traces
The control plane
Agents in development
Tokens

There is no per-seat charge for the people who use agents and no per-seat charge for the people who build them. The only thing on the meter is the digital workforce itself.

Three ways customers start

Govern the estate
$500K

Opencontroller + basecode. Every agent on the record, per year.

Govern and build
$850K

Opencontroller + basecode, one block of 150 agent-months, Open Cowork. Per year.

Fully sovereign
$1.6M

Optimus, Opencontroller + basecode, one block of agent-months, Openfinetuner, Open Cowork. Year one, including the hardware.

Questions finance teams ask

What counts as one agent?

Anything with its own identity, its own policy and its own deployment lifecycle. The test: can it be promoted, permissioned and rolled back without touching anything else? If yes, it is an agent. Versions, clones, replicas and specialists inside a workflow are part of their parent.

Do you charge for tokens?

No. You pay your model provider for tokens, or run open-weight models on your own hardware. Lyzr never adds a second line on the same unit of work. On Lyzr-hosted deployments a fair-use envelope sits in the terms, not on the invoice.

How do you tell development from production?

By the runtime path, not a checkbox. An agent bound to your production identity provider, production tools and production data plane is in production. Everything else is an intern and draws nothing.

Who decides an agent's class?

The controller does, from the permissions actually granted. A read-only tool keeps an agent at A. A write into an enterprise system makes it B. An attached sandbox makes it C. Nobody can change a class by relabeling it.

What happens when we outgrow a block?

Additional agent-months bill monthly in arrears at $1,700 each, or you buy the next block of 150 for $250K. Unused agent-months roll over on multi-year terms. Archiving an agent stops its draw the same month.

Do agents we did not build on Lyzr cost agent-months?

No. Agents from Copilot, Bedrock, LangGraph or anywhere else are governed under Opencontroller at the flat price. Agent-months are only for agents built and run on Lyzr.

What if a production agent only runs a few times a month?

It is part-time and draws a quarter of its class rate for that month: 0.25 for a Class B agent, 0.06 for a Class A. Class C stays at full rate because a sandbox with production credentials is standing capacity. If an agent is seasonal, archive it between uses: the draw stops, the history stays, and you re-promote it in one click.

What counts as a run?

One unit of work completed by the agent: a request answered, a ticket closed, a claim decided, a record processed. One execution of a workflow that handles one case is one run, however many calls happen inside it. A batch that processes 10,000 records is 10,000 runs.

Size your first block with us.

Bring your process list. We will classify it, count the interns, and show you the agent-months before you sign anything.

Build your estimate