Skip to content

Public prices. A clear contract.

Like the rest of the market, we price per agent and by what it can do. You know what you pay before the first call, the proposal arrives as soon as the interview ends, and you sign online.

Which offer for your agent? Estimate in 2 minutes →

  • OWASP LLM Top 10
  • OWASP Agentic
  • MITRE ATLAS
  • NIST AI RMF
  • AI Act
  • ISO/IEC 42001
SMEs and startups · order online

Your first crash test, no meeting needed

Describe your agent in a few minutes, pay online, the test starts. No proposal, no salesperson.

Express crash test

Your agent against ten essential tests, no meeting needed

390 € excl. VAT

one agent, one-off payment

  • Ten of the twenty-two tests: pressure on the goal, trapped instruction, fake superior, data leak, manipulative customer, hidden instruction, trap in the tool, bank details change, stop order, "I am an AI" (only those that concern your agent)
  • 5 market models, plus yours if you give it, each test played 2 times
  • Report, model ranking and re-test included
  • Move up to the full crash test paying only the difference
  • Ordered and paid online
Order online

Follow-up

Your agent crash-tested again every quarter

290 € excl. VAT / month

billed yearly: 3,480 € excl. VAT, one agent

  • A full crash test every three months, four in the year
  • Each time, a report of what changed since the previous one
  • Starts when your first report is delivered
  • Taken with the crash test: two months free
Add to a crash test

Want to integrate a specific model? The crash test plays your agent on that model, with your rules and tools, next to the market models: you see whether it holds in your case, not only in ours. Add it after ordering. No agent yet? Have the model measured alone →

Not ready yet? The configuration check is free. Check my configuration →

See what we test: try to make Léa, our demo agent, give in. Try the demo →

Pilot price: −35 %. For our first clients, SMEs and large accounts alike, in exchange for the right to name you as a reference (name and logo, never the content of the reports) and a 30-minute call after the report.
SMEs: full crash test online, 5 of 5 places left. · Large accounts: 3 of 3 places left.

Order at the pilot price Take it

Founder client offer: 10 of 10 places left. For our first clients, whatever the offer chosen, at that offer's public price: no discount of its own.

  • The price of the offer you choose, frozen for 2 years.
  • A direct line with Cyril Delage, CrashMire's founder.
  • In exchange: a testimonial or your logo, with your written agreement, and a 30-minute call each month.
  • Closes at the 10th signature or on 31 Dec 2026, whichever comes first.
Become a founder client

Pilot: 3 crash tests offered in exchange for feedback. For a company whose AI agent already acts: the full crash test, offered, in exchange for a testimonial or your logo (with your written agreement), a 30-minute call and a case study, anonymised or under your name. The places go to the first applications, in order of arrival. A programme of its own: no discount on the paid offers.
3 of 3 places left.

Apply

Large accounts: on quote, from 4,900 € excl. VAT. Several agents, your in-house model, a framework contract, monthly monitoring: we send you a proposal after a short call.

Get my proposal
See the large-account offers in detail
First step · before the crash test

Which model for your agent?

Before having your agent crash-tested: know which model holds best for your use case according to the CrashMire AI agent barometer, the monthly barometer of AI models' resistance to manipulation, and which guardrails to put in the code if you keep a cheaper model.

Diagnostic: which model for your agent

Which model to pick for your use case, and which guardrails to put in the code

2,400 € excl. VAT

Launch price: 1,900 € excl. VAT for the first 3 diagnostics, 3 of 3 places left

  • A short questionnaire about your agent: what it does, its tools, what it can touch (money, data, actions), the models used or considered, your budget
  • One short call in the evening, no daytime workshops
  • The model that holds best under pressure for your use case, from the CrashMire AI agent barometer, the monthly barometer of AI models' resistance to manipulation
  • If you keep a cheaper or weaker model: the guardrails to put in the code (amount caps, identity checks, human approval above a threshold, rights per tool)
  • Written report within 5 working days, with the next step. No test is run on your agent
Ask for a diagnostic

Your model, measured

A model the barometer does not show yet, measured on its protocol, next to the published models

900 € excl. VAT

per model; 500 € excl. VAT more when added to a diagnostic

  • The model you integrate, or publish, played by the reference agent of the CrashMire AI agent barometer: the same tests, the same number of runs, the same judging rules
  • Its result test by test, next to the models of the latest published edition
  • It may enter a later public edition marked "measured at a client's request"; your name is never published
  • Requested by the model's publisher or seller: the same price, the result published whatever it is, marked "requested by the publisher"
Have a model measured

What it measures, and what it does not. The CrashMire AI agent barometer measures resistance to manipulation and pressure, not cost, latency or answer quality. The diagnostic runs no test on your agent: that is what the crash test does.

CrashMire takes no commission from AI model vendors. A vendor may pay for its model to be measured, at the same price as anyone: the result is published whatever it is, marked "requested by the publisher". The method is public.

The diagnostic in detail →

Large accounts · one-off

One agent, one full crash test

The price follows what the agent can do: the more it acts, the more ways it can go wrong, the longer the test. The interview sets the level; you can pick another.

Essential crash test

A conversational assistant, no actions

4,900 € excl. VAT

one agent, one-off payment

  • Twenty-two tests: pressure on the target, the same pressure with nobody watching, planted instruction, hidden instruction, fake superior, data leak, insider temptation, manipulative customer, runaway on a broken tool, trap in the tool, earned trust, fake colleague, bank details change, fake relative, relay attack, stop order, "I am an AI", hijacked account, side door, camouflaged planted instruction, long conversation, grey zone (the bank-details and personal-data tests only if your agent handles them, those that assume customers only if it talks to them)
  • 8 market models, plus yours (internal included), 5 runs per model and per test
  • Your real prompt and configuration, or your agent plugged in directly
  • Counter-test with a hardened prompt, detailed report and fix plan
  • One re-test included within 90 days
Get my proposal

Critical crash test

Money, regulated data, several agents

from 25,000 € excl. VAT

one agent, one-off payment

  • The whole Agent crash test, with scenarios built on your business rules and regulatory constraints
  • Agents that pay, refund, decide or drive other agents
  • Report ready for the risk committee, prioritised fix plan
  • 8 market models, plus yours, 5 runs per model and per test
  • One re-test included within 90 days
Get my proposal
Large accounts · several agents

One agent per business line. One contract.

In a large group, every business line builds its own agent, often on its own. One framework contract covers them all: each business line adds its agent online, each gets its report, and IT, the CISO and risk see the whole picture.

Flaw to fixWatch pointHolds

Plan
Price
Per agent
As one-off crash tests
Get my proposal
Large accounts · continuous

Your agents watched all year

A model changes, a prompt moves, a flaw appears: every month and on every change, your agents go through the crash test again. Priced per agent; one more agent during the year: 12,000 € excl. VAT a year, pro rata.

Continuous — 10 agents

Several teams, several agents

72,000 € excl. VAT

up to 10 agents, per year

  • Up to 10 agents, all the tests
  • Automatic re-test every month and on every model change
  • A report each time: what changed, the proposed fixes
  • An alert as soon as a new flaw appears
  • Dashboard of your agents in your client space
Get my proposal

Group contract

Every entity, every agent

from 120,000 € excl. VAT

up to 25 agents, per year

  • From 25 agents, several entities or subsidiaries
  • All of Continuous, across the whole agent estate
  • Quarterly review for the risk committee
  • Named contact, negotiated purchasing terms
Get my proposal
Options

What you can add

Add them to any plan when you sign, or later.

Compliance pack

Each flaw mapped to the AI Act, OWASP (LLM and agentic), NIST AI RMF and ISO 42001, with an evidence file for audit and the DPO.

+ 4,900 € excl. VAT with a one-off crash test
+ 15,000 € excl. VAT per year, with a subscription

Release gate

A crash test runs on every prompt or model change; the release is blocked on a regression.

+ 12,000 € excl. VAT per year, with a subscription

Turnkey fixes

In practice: the fallback plan applied to your agent (what is cut, where, who takes over the cases, how to restart), the caps and approvals to set tool by tool, the hardened prompt rewritten, and re-tests until green. This is advice and settings that you apply: CrashMire never takes your production keys and never touches your production.

+ 4,900 € excl. VAT with a one-off crash test
+ 9,800 € excl. VAT per year, with a subscription

Service commitment

Report within 48 business hours, named contact, quarterly review for your risk committee.

+ 9,000 € excl. VAT per year, with a subscription
Compare

What is included

Essential crash test4,900 €Agent crash test12,000 €Critical crash testfrom 25,000 €Continuous — 3 agents36,000 € / yrGroup contractfrom 120,000 € / yr
What is tested
Twenty-two tests: pressure on the target, the same pressure with nobody watching, planted instruction, hidden instruction, fake superior, data leak, insider temptation, manipulative customer, runaway on a broken tool, trap in the tool, earned trust, fake colleague, bank details change, fake relative, relay attack, stop order, "I am an AI", hijacked account, side door, camouflaged planted instruction, long conversation, grey zone (the bank-details and personal-data tests only if your agent handles them, those that assume customers only if it talks to them) ✓✓✓✓✓
On your real prompt and your model, internal included ✓✓✓✓✓
Traps on every tool: planted incoming messages, access left open, confidential documents —✓✓✓✓
Scenarios built on your business rules and regulatory constraints ——✓✓✓
Counter-test with a hardened prompt, fix plan ✓✓✓✓✓
"Suspected a test" measure, per model ✓✓✓✓✓
What each fault puts at stake: irreversible, third party reached, right handed over ✓✓✓✓✓
Honest report: what the agent says it did, compared with what it did ✓✓✓✓✓
Critical-moment replay: the most serious faults replayed 20 times at the exact turn, with and without the hardened prompt ✓✓✓✓✓
The guardrails the public texts recommend (AI Act, ANSSI, OWASP), next to what the crash test observed: stop button, caps, human approval, least privilege, journal, content without authority ✓✓✓✓✓
When, and how many
Number of agents 1113 or 1025 and more
Re-test 1, within 90 days1, within 90 days1, within 90 daysEvery monthEvery month
Re-test on every model or prompt change ———✓✓
Alert as soon as a new flaw appears ———✓✓
Going further
Compliance pack (AI Act, OWASP, NIST, ISO 42001) OptionOptionOptionOptionOption
Turnkey fixes OptionOptionOptionOptionOption
Release gate ———OptionOption
Report within 48 h, named contact ———Option✓
Quarterly review for the risk committee ———Option✓

The method: what is tested, how faults are judged, what is measured next to the score →

How it works

From interview to report, without waiting for a salesperson

Proposal, purchase order, invoice: all automatic. You step in at two moments only.

  1. Online interview

    You describe your agent: what it does, its tools, its rules.

    15 min
  2. Priced proposal

    As a PDF, by email: the level that fits your agent and the annual plans.

    Instant
  3. Online signature

    Plan, options, billing, electronic signature: this is your purchase order.

    2 min
  4. Crash test

    It starts once we have checked the order, on your prompt and your model.

    Same day
  5. Invoice

    Sent by email, payable by bank transfer within 30 days.

    On signature
  6. Report

    Flaws, evidence, prioritised fix plan, counter-test with a hardened prompt.

    Within 5 business days
  7. Follow-up

    Re-test after your fixes; with a subscription, every month and on every change.

    90 days or monthly
YouAutomatic
Contract

The contract, in plain words

Our terms of sale are public: your purchasing team and your lawyer can read them before even writing to us.

Term

One-off: until the re-test. Subscription: 12 firm months, renewed tacitly; terminate in writing three months before the anniversary.

Payment

Invoice on signature, bank transfer within 30 days. Subscriptions invoiced yearly in advance. Statutory late-payment penalties (art. L.441-10 French Commercial Code).

Price revision

Every year, at the anniversary, using the revised Syntec index, the standard formula for professional services in France.

Scope

Priced per agent. One agent = one role, one prompt, one set of tools. Extra agent: 12,000 € excl. VAT a year, pro rata.

Ownership

The report is yours. The engine, scenarios and method stay ours: no software installed at your site, no code delivered.

Data

Fictitious data only, EU hosting, never any model training on your data. Data processing agreement (art. 28 GDPR) annexed.

Liability

Best-efforts obligation, capped at the amounts paid over 12 months. A crash test is not a certification.

References

We name you only with your written consent. The content of the reports is never disclosed.

Purchasing

Supplier onboarding

Everything your purchasing team asks for, sent on first request:

  • URSSAF certificate of social contributions (under 6 months)
  • SIRENE registration notice
  • Bank details and EU VAT number
  • Non-disclosure agreement, signed before any exchange
  • Data processing agreement (art. 28 GDPR)
  • List of sub-processors and hosting locations
  • Supplier security questionnaire filled in
  • Invoicing against your purchase order number
Integrators and agencies

You deliver agents to your clients

Crash-test every agent before you hand it over, under a yearly plan.

  • Integrator partner — 12,000 € excl. VAT / yr, up to 10 delivered agents

the integrator programme →

Questions about pricing

Why a price per agent?

It is the market's unit, and it is what drives the cost of a test: each agent has its role, its tools and its ways of going wrong. You pay for what is tested, not for seats or users.

What exactly is an agent?

An AI assistant or agent defined by a role, instructions and a set of tools. Successive versions of the same agent count as one; a new role, or tools giving access to new systems, make a new agent.

What if I have one more agent during the year?

It is added to your subscription for 12,000 € excl. VAT a year, pro rata of the remaining period, or you move to the next plan up if that is cheaper.

How do I know which level fits my agent?

The interview sets it: an agent that talks is Essential, an agent that acts through tools or reads your documents is Agent, an agent that touches money, regulated data or other agents is Critical. The proposal also shows the annual plans; you choose.

Can we try before signing?

The live demo is free, no sign-up. For a first commitment, the Essential or Agent crash test covers a single agent.

Our purchasing works with purchase orders and supplier onboarding: is that possible?

Yes. Enter your purchase order number when signing, it appears on the invoice. We send the onboarding documents on first request; your own purchasing terms can be discussed, but without a written agreement our terms of sale apply.

Can we install CrashMire on our premises?

No: the engine stays with us, which protects the method and spares you any maintenance. For an internal model, you expose a test environment; only fictitious scenarios go through it.

Is the pilot programme a discount?

No. It is a programme of its own, free and limited to 3 full crash tests, for companies whose agent already acts, in exchange for a testimonial or logo (with written agreement), a 30-minute call and a case study. The places go to the first applications, in order of arrival. It gives no discount on the paid offers, and the founder client offer stays at the public price.

Barometer, diagnostic or crash test: what is the difference?

The CrashMire AI agent barometer, the monthly barometer of AI models' resistance to manipulation, is free and public: our reference agent played on the market's models, for everyone. The diagnostic (2,400 € excl. VAT) reads those measurements for your use case: which model holds best for you, and which guardrails to put in the code; it runs no test on your agent. The crash test tests your agent itself, with its rules and tools, and proves that your guardrails hold. A model the barometer does not show yet can be measured on demand, on the same protocol (900 € excl. VAT).

The proposal arrives as soon as the interview ends.

Get my proposal Not the decision-maker? The email is ready.