Research on AI models, agents and tools. Every review says what we tested, how, and what we did not test.
What we cover
Articles
Explanations and analysis, written to be checked against sources.
In preparationReviews
Hands-on tests of AI products, with a fixed structure and a test date.
In preparationModels
What each model is good at, where it fails and what it costs to run.
In preparationAgents
How agent systems behave in real tasks, including their failure modes.
In preparationPlugins
What a plugin can access and change before you install it.
In preparationConnectors
What connecting a service gives an AI access to, and how to limit it.
In preparationAutomation
Workflows that work in practice, and the ones that look better than they run.
In preparationSecurity
Permissions, data handling and risks, described by what was and was not tested.
In preparation
How we review
Every review follows the same structure, so two products can be compared on the same questions.
- What it is
- The product and who makes it.
- What it actually does
- Observed behaviour, not marketing claims.
- What it can access
- Files, accounts, network and devices.
- Permissions
- What it asks for and what it really needs.
- Privacy and data handling
- Where data goes and how long it stays.
- Best uses
- Tasks where it is a good fit.
- Poor uses and limits
- Where it fails or should not be used.
- Security observations
- Findings, with the scope of what was tested.
- Real Codelab test
- The test we ran, so you can repeat it.
- Result
- What happened, with evidence.
- Alternatives
- Other options worth comparing.
- Pricing and value
- What it costs and whether it is worth it.
- Last tested
- The date, because products change.
- Sources
- Where every claim comes from.
We do not call a product safe in absolute terms. We say what we tested, and what we did not.