The CrashMire AI agent barometer
Free, public
Our reference agent, played on the market's models. A general view, for everyone.
The barometer →First step
Torn between several models, or want to keep a cheaper one? From the CrashMire AI agent barometer, the monthly barometer of AI models' resistance to manipulation, we tell you which one holds best under pressure for your use case, and which guardrails to put in the code if you keep a weaker model.
1,900 € excl. VAT 2,400 € excl. VAT
one-off, for one agent · written report within 5 working days
Among the models you consider, the one that best resists pressure and manipulation in situations like yours, according to the CrashMire AI agent barometer. Measured facts only: "in our tests, this model gave in N times out of M".
If you keep a cheaper or weaker model: what to lock in the code, not in the prompt. Amount caps, identity checks, human approval above a threshold, rights limited tool by tool.
Turnkey fixes (4,900 € excl. VAT, as an option of a crash test) and/or a crash test of your agent to prove those guardrails hold (from 4,900 € excl. VAT for a large account, 1,900 € excl. VAT online for an SMB).
An instruction in a prompt is a request, not a lock.
Included in the diagnostic report, at the same price. For each action of your agent (money, data, outgoing messages, deletion, code): what happens to the service if the AI is cut, where to put the stop so that it does not depend on the AI, and what to set or ask your provider if your agent is a setting in a tool you bought.
The rule: the AI writes a request, the code executes it. If you cut the AI, the request is filled in another way and the service keeps running.
The checklist for stopping an AI agent without stopping the service →
For professionals only. Terms of sale
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 method →
The CrashMire AI agent barometer, which it relies on today: edition of 6 Nov 2026, 14 models, 14 tests.
You are considering a model the barometer does not show yet, or you publish it? We measure it on the barometer's protocol: the same reference agent, the same tests, the same number of runs, the same judging rules. You get its result next to the published models.
900 € excl. VAT per model, alone · 500 € excl. VAT more, added to the diagnostic: the report then rests on its own measurement
Requested by a company: the model may enter a later public edition marked "measured at a client's request", never your name. Requested by its publisher or seller: the same price, the result published whatever it is, marked "requested by the publisher". The model must be reachable on OpenRouter or at your own endpoint.
Have a model measuredFree, public
Our reference agent, played on the market's models. A general view, for everyone.
The barometer →1,900 € excl. VAT
Your use case, read through those measurements: the model that holds best for you, and the guardrails to put in the code. No test on your agent.
From 1,900 € excl. VAT online · 4,900 € excl. VAT for large accounts
Your agent itself, with its rules and tools, attacked and pushed to the limit on several models: the proof that your guardrails hold.
Pricing →