70% of the AI agents we tested don't publish their pricing

Hlido is an independent review desk that tests AI agents hands-on and publishes the evidence. If you are choosing one, the first question you ask is usually "what does it cost?" For most of the agents we have tested, the vendor's public surface does not answer it.

The number

Hlido runs a dated, evidence-backed check on every agent it reviews: are pricing or access terms reachable on the vendor's own public surface, without a sales call and without an account? Across 623 tested agents where that check produced a definitive verdict, 189 passed and 434 failed — 69.7%.

That is not a survey and not a vendor self-report. Each verdict is attached to a specific test date and is published on the agent's own record, alongside the rest of the evidence.

Of the agents where we could extract a concrete commercial model from our own review evidence, the shape looks like this:

Model we recorded	Agents
Open source	72
Paid (tier or plan, amount not stated)	22
Free tier / freemium	20
Subscription	15
Usage-based (per call, per credit, per token)	12
Enterprise quote only	1

The long tail is the story: for the large majority, we hold no commercial model at all, because nothing on the public surface stated one.

What "discloses pricing" actually looks like

A handful of the agents we tested publish numbers a reader can act on immediately, and we quote them verbatim from our own review evidence rather than from a vendor's marketing page:

• Omniwork — "Free (100+ agents, 5 deep tasks/mo), Pro $69/mo, Ultimate $1999/yr"
• Lindy — "Plus $49.99/mo, Pro $59.99/mo, Enterprise custom"
• Workflow Machine — self-serve pricing on the page: "Free / $18 / $48, with explicit run counts, AI-credit allowances, active-workflow limits"
• AgentScrape — prices every tool at "$0.001 in on-chain USDC"

Four examples out of hundreds is the point. Concrete, self-serve, machine-readable pricing is the exception in this market, not the norm.

One claim this data cannot support

There is a sharper headline sitting in this dataset: agents in our top tier disclose pricing far more often than agents in our bottom tier. It is true as arithmetic, and it is circular by construction, so we are reporting it as an artefact rather than a finding.

Pricing visibility is one of the inputs to a dimension we score. An agent that keeps its pricing off its public surface loses ground on that dimension, which moves it down the tiers. A correlation between "discloses pricing" and "scores well" is therefore partly a restatement of how the score is built — a definition dressed as a discovery.

We audit vendor claims for exactly that shape. The standard has to survive contact with our own data.

What the data does support is plainer and more useful: most AI agent vendors do not let you find out what they cost without talking to them.

What you can do with this

Every agent record now carries a Pricing & access section: the commercial model we recorded, any price points quoted verbatim from our review evidence, and the dated pass/fail of the pricing check itself. Where we hold nothing, the row is absent — we do not guess, and we never assume "free" by default.

It is machine-readable too. The whole set is published at /data/pricing-facts.json, and each agent's own scorecard at /data/scorecards/{slug}.json carries a pricing_facts block, so an agent doing procurement research can read it directly.

Two caveats we would want stated if this were someone else's data. Everything above is derived from evidence already held in our reviews, so a price we recorded is a price that was true on the test date and may have changed. And a failed pricing check says the terms were not reachable when we looked — it is not a claim that the vendor is hiding something.

Prices move. The check is dated for exactly that reason.